From 46f0f3cee90f7cf852bf5bcf89866b57448f1ffa Mon Sep 17 00:00:00 2001 From: zhaoying Date: Mon, 26 Jan 2026 17:43:25 +0800 Subject: [PATCH 1/7] feat(web): update read_all_config select valueKey --- web/src/components/CustomSelect/index.tsx | 19 +++++++++++++------ web/src/views/ApplicationConfig/Agent.tsx | 6 ++++-- web/src/views/Workflow/constant.ts | 4 ++-- web/src/views/Workflow/types.ts | 2 +- 4 files changed, 20 insertions(+), 11 deletions(-) diff --git a/web/src/components/CustomSelect/index.tsx b/web/src/components/CustomSelect/index.tsx index 1887d635..6153a76d 100644 --- a/web/src/components/CustomSelect/index.tsx +++ b/web/src/components/CustomSelect/index.tsx @@ -15,7 +15,7 @@ interface ApiResponse { interface CustomSelectProps extends Omit { url: string; params?: Record; - valueKey?: string; + valueKey?: string | string[]; labelKey?: string; placeholder?: string; hasAll?: boolean; @@ -66,11 +66,18 @@ const CustomSelect: FC = ({ {...props} > {hasAll && {allTitle || t('common.all')}} - {displayOptions.map((option) => ( - - {String(option[labelKey])} - - ))} + {displayOptions.map((option) => { + const getValue = () => { + if (typeof valueKey === 'string') return option[valueKey]; + return valueKey.find(key => option[key] != null) ? option[valueKey.find(key => option[key] != null)!] : undefined; + }; + const value = getValue(); + return ( + + {String(option[labelKey])} + + ); + })} ); }; diff --git a/web/src/views/ApplicationConfig/Agent.tsx b/web/src/views/ApplicationConfig/Agent.tsx index 77e90440..97a622d1 100644 --- a/web/src/views/ApplicationConfig/Agent.tsx +++ b/web/src/views/ApplicationConfig/Agent.tsx @@ -79,7 +79,7 @@ const SelectWrapper: FC<{ title: string, desc: string, name: string | string[], placeholder={t('common.pleaseSelect')} url={url} hasAll={false} - valueKey='config_id' + valueKey={['config_id_old', 'config_id']} labelKey="config_name" /> @@ -126,12 +126,14 @@ const Agent = forwardRef((_props, ref) => { getApplicationConfig(id as string).then(res => { const response = res as Config let allTools = Array.isArray(response.tools) ? response.tools : [] + const memoryContent = response.memory?.memory_content + const convertedMemoryContent = memoryContent && !isNaN(Number(memoryContent)) ? Number(memoryContent) : memoryContent form.setFieldsValue({ ...response, tools: allTools, memory: { ...response.memory, - memory_content: response.memory?.memory_content ? Number(response.memory?.memory_content) : undefined + memory_content: convertedMemoryContent } }) setData({ diff --git a/web/src/views/Workflow/constant.ts b/web/src/views/Workflow/constant.ts index e250e184..aab8be7d 100644 --- a/web/src/views/Workflow/constant.ts +++ b/web/src/views/Workflow/constant.ts @@ -200,7 +200,7 @@ export const nodeLibrary: NodeLibrary[] = [ config_id: { type: 'customSelect', url: memoryConfigListUrl, - valueKey: 'config_id', + valueKey: ['config_id_old', 'config_id'], labelKey: 'config_name' }, search_switch: { @@ -223,7 +223,7 @@ export const nodeLibrary: NodeLibrary[] = [ config_id: { type: 'customSelect', url: memoryConfigListUrl, - valueKey: 'config_id', + valueKey: ['config_id_old', 'config_id'], labelKey: 'config_name' } } diff --git a/web/src/views/Workflow/types.ts b/web/src/views/Workflow/types.ts index 909c30e4..31d1f512 100644 --- a/web/src/views/Workflow/types.ts +++ b/web/src/views/Workflow/types.ts @@ -14,7 +14,7 @@ export interface NodeConfig { url?: string; params?: { [key: string]: unknown; } - valueKey?: string; + valueKey?: string | string[]; labelKey?: string; defaultValue?: any; From f76bffb4823252ed482867c3fe112a1cf09f5a16 Mon Sep 17 00:00:00 2001 From: zhaoying Date: Mon, 26 Jan 2026 18:32:18 +0800 Subject: [PATCH 2/7] fix(web): KnowledgeConfigModal bugfix --- .../components/Knowledge/KnowledgeConfigModal.tsx | 2 +- .../Properties/Knowledge/KnowledgeConfigModal.tsx | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/web/src/views/ApplicationConfig/components/Knowledge/KnowledgeConfigModal.tsx b/web/src/views/ApplicationConfig/components/Knowledge/KnowledgeConfigModal.tsx index abf56b18..70b17a11 100644 --- a/web/src/views/ApplicationConfig/components/Knowledge/KnowledgeConfigModal.tsx +++ b/web/src/views/ApplicationConfig/components/Knowledge/KnowledgeConfigModal.tsx @@ -66,7 +66,7 @@ const KnowledgeConfigModal = forwardRef { if (values?.retrieve_type) { const fieldsToReset = Object.keys(values).filter(key => - key !== 'kb_id' && key !== 'retrieve_type' + key !== 'kb_id' && key !== 'retrieve_type' && key !== 'top_k' ) as (keyof KnowledgeConfigForm)[]; form.resetFields(fieldsToReset); } diff --git a/web/src/views/Workflow/components/Properties/Knowledge/KnowledgeConfigModal.tsx b/web/src/views/Workflow/components/Properties/Knowledge/KnowledgeConfigModal.tsx index 77ca21a2..196ce8e3 100644 --- a/web/src/views/Workflow/components/Properties/Knowledge/KnowledgeConfigModal.tsx +++ b/web/src/views/Workflow/components/Properties/Knowledge/KnowledgeConfigModal.tsx @@ -66,7 +66,7 @@ const KnowledgeConfigModal = forwardRef { if (values?.retrieve_type) { const fieldsToReset = Object.keys(values).filter(key => - key !== 'kb_id' && key !== 'retrieve_type' + key !== 'kb_id' && key !== 'retrieve_type' && key !== 'top_k' ) as (keyof KnowledgeConfigForm)[]; form.resetFields(fieldsToReset); } @@ -108,6 +108,7 @@ const KnowledgeConfigModal = forwardRef {/* Top K */} @@ -116,13 +117,12 @@ const KnowledgeConfigModal = forwardRef form.setFieldValue('top_k', value)} + // onChange={(value) => form.setFieldValue('top_k', value)} /> {/* 语义相似度阈值 similarity_threshold */} From 5267bd60a566893d9269a20c4a073642b479fa33 Mon Sep 17 00:00:00 2001 From: zhaoying Date: Mon, 26 Jan 2026 18:40:28 +0800 Subject: [PATCH 3/7] fix(web): iteration's variable add parameter-extractor node --- web/src/views/Workflow/constant.ts | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/web/src/views/Workflow/constant.ts b/web/src/views/Workflow/constant.ts index e250e184..7b15c049 100644 --- a/web/src/views/Workflow/constant.ts +++ b/web/src/views/Workflow/constant.ts @@ -284,7 +284,7 @@ export const nodeLibrary: NodeLibrary[] = [ config: { input: { type: 'variableList', - filterNodeTypes: ['knowledge-retrieval', 'iteration', 'loop'], + filterNodeTypes: ['knowledge-retrieval', 'iteration', 'loop', 'parameter-extractor'], filterVariableNames: ['message'] }, parallel: { From a5b8d3afa5ef19723ae5cce57f3fbbee70ff51f8 Mon Sep 17 00:00:00 2001 From: lixinyue11 <94037597+lixinyue11@users.noreply.github.com> Date: Mon, 26 Jan 2026 19:05:07 +0800 Subject: [PATCH 4/7] Fix/memory bug fix (#200) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * 图谱数据量限制数量去掉 * 图谱数据量限制数量去掉 * 图谱数据量限制数量去掉 * 用户详情优化 * 用户详情优化 * 用户详情优化 * 用户详情优化 * 用户详情优化 * 用户详情优化 * 读取的接口，去掉全局锁 * 输出数组 * 反思优化1.0（优化隐私输出、时间检索） * 反思优化1.0（优化隐私输出、时间检索） * 反思优化1.0（优化隐私输出、时间检索） * 反思优化测试接口 * 反思优化测试接口 * 读取接口内层嵌套BUG修复 * 读取接口内层嵌套BUG修复 * 读取接口内层嵌套BUG修复 * 读取接口内层嵌套BUG修复 * 读取接口内层嵌套BUG修复 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察) * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 新增中翻英功能（记忆时间线）(用户摘要)(兴趣分布接口)(查询核心档案)(记忆洞察)-接口添加翻译字段 * 把group_id替换end_user_id * 把group_id替换end_user_id_ * 把group_id替换end_user_id_ * config_config替换成memory_config * config_config替换成memory_config * [fix]Fix the memory interface to use end_user_id. * config_config替换成memory_config * config_config替换成memory_config * config_config替换成memory_config * config_id字段改成UUID * config_id字段改成UUID * config_id字段改成UUID * config_id字段改成UUID，与develop校对恢复 * 检查项目，修复group_id的遗留问题 * 检查项目，修复group_id的遗留问题 * 解决冲突 * 解决冲突 * end_user_id清理干净 * end_user_id清理干净 * 修复遗留合并BUG * 修复遗留合并BUG * 修复遗留合并BUG * 修复遗留合并BUG * 感知meta_data字段BUG修复 * user_id->现实为config_id_old * user_id->显示为config_id_old传输 * user_id->显示为config_id_old传输 * user_id->显示为config_id_old传输 --------- Co-authored-by: lanceyq <1982376970@qq.com> --- api/app/services/memory_storage_service.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/api/app/services/memory_storage_service.py b/api/app/services/memory_storage_service.py index 1707f8fa..0ede7bd3 100644 --- a/api/app/services/memory_storage_service.py +++ b/api/app/services/memory_storage_service.py @@ -188,7 +188,7 @@ class DataConfigService: # 数据配置服务类（PostgreSQL） "config_desc": config.config_desc, "workspace_id": str(config.workspace_id) if config.workspace_id else None, "end_user_id": config.end_user_id, - "config_id_old": config.config_id_old, + "config_id_old": int(config.user_id), "apply_id": config.apply_id, "llm_id": config.llm_id, "embedding_id": config.embedding_id, From 80ca247435fe9d79a0c0c71fd4b0113284eaf359 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E4=B9=90=E5=8A=9B=E9=BD=90?= <162269739+lanceyq@users.noreply.github.com> Date: Mon, 26 Jan 2026 19:05:20 +0800 Subject: [PATCH 5/7] Refactor/benchmark test (#196) * [changes]refactor locomo_test * [fix]Fix the circular import of ModelParameters * [changes]The benchmark test can run stably. * [fix]Complete end-to-end LoCoMo repair * [fix]Complete the end-to-end longmemeval and memsciqa fixes * [changes]Complete the benchmark test description document to ensure that the configuration parameters take effect. * [changes]refactor locomo_test * [fix]Fix the circular import of ModelParameters * [changes]The benchmark test can run stably. * [fix]Complete end-to-end LoCoMo repair * [fix]Complete the end-to-end longmemeval and memsciqa fixes * [changes]Complete the benchmark test description document to ensure that the configuration parameters take effect. * [changes]Benchmark test adaptation for end_user_id * [changes]refactor locomo_test * [fix]Fix the circular import of ModelParameters * [changes]The benchmark test can run stably. * [fix]Complete end-to-end LoCoMo repair * [fix]Complete the end-to-end longmemeval and memsciqa fixes * [changes]Complete the benchmark test description document to ensure that the configuration parameters take effect. * [fix]Complete the end-to-end longmemeval and memsciqa fixes * [changes]Complete the benchmark test description document to ensure that the configuration parameters take effect. * [changes]Benchmark test adaptation for end_user_id --- .../memory/evaluation/.env.evaluation.example | 224 +++++ api/app/core/memory/evaluation/.gitignore | 13 + api/app/core/memory/evaluation/benchmark.md | 772 +++++++++++++++- .../memory/evaluation/check_enduser_data.py | 371 ++++++++ .../core/memory/evaluation/common/metrics.py | 2 +- .../memory/evaluation/dialogue_queries.py | 6 +- .../memory/evaluation/extraction_utils.py | 369 +++++--- .../evaluation/locomo/locomo_benchmark.py | 863 +++++++++++------- .../memory/evaluation/locomo/locomo_test.py | 293 +++--- .../memory/evaluation/locomo/locomo_utils.py | 69 +- .../evaluation/locomo/qwen_search_eval.py | 96 +- ...earch_eval.py => longmemeval_benchmark.py} | 181 ++-- .../evaluation/longmemeval/test_eval.py | 132 ++- .../evaluation/memsciqa/memsciqa-test.py | 224 +++-- .../{evaluate_qa.py => memsciqa_benchmark.py} | 123 ++- api/app/core/memory/evaluation/run_eval.py | 21 +- .../extraction_orchestrator.py | 22 +- api/app/models/agent_app_config_model.py | 2 +- api/app/models/multi_agent_model.py | 2 +- api/app/schemas/multi_agent_schema.py | 2 +- api/app/services/master_agent_router.py | 2 +- api/app/utils/app_config_utils.py | 2 +- 22 files changed, 2760 insertions(+), 1031 deletions(-) create mode 100644 api/app/core/memory/evaluation/.env.evaluation.example create mode 100644 api/app/core/memory/evaluation/.gitignore create mode 100644 api/app/core/memory/evaluation/check_enduser_data.py rename api/app/core/memory/evaluation/longmemeval/{qwen_search_eval.py => longmemeval_benchmark.py} (93%) rename api/app/core/memory/evaluation/memsciqa/{evaluate_qa.py => memsciqa_benchmark.py} (76%) diff --git a/api/app/core/memory/evaluation/.env.evaluation.example b/api/app/core/memory/evaluation/.env.evaluation.example new file mode 100644 index 00000000..be089eb4 --- /dev/null +++ b/api/app/core/memory/evaluation/.env.evaluation.example @@ -0,0 +1,224 @@ +# ============================================================================ +# 基准测试统一配置文件示例 +# ============================================================================ +# 复制此文件为 .env.evaluation 并根据需要修改 +# 支持的基准测试：LoCoMo、LongMemEval、MemSciQA +# ============================================================================ + +# ============================================================================ +# 通用配置（所有基准测试共用） +# ============================================================================ + +# ---------------------------------------------------------------------------- +# Neo4j 配置 +# ---------------------------------------------------------------------------- +# 默认 Group ID（建议各基准测试使用独立的 group） +EVAL_GROUP_ID=benchmark_default + +# ---------------------------------------------------------------------------- +# 模型配置（必需） +# ---------------------------------------------------------------------------- +# ⚠️ 必填：从数据库 models 表中选择有效的模型 ID +# +# 如何获取模型 ID： +# 1. 查询数据库：SELECT id, model_name FROM models WHERE is_active = true; +# 2. 或通过系统管理界面查看 +# 3. 确保模型可用且配置正确 + +# LLM 模型 ID（必填） +EVAL_LLM_ID=your_llm_model_id_here + +# Embedding 模型 ID（必填） +EVAL_EMBEDDING_ID=your_embedding_model_id_here + +# ---------------------------------------------------------------------------- +# 检索参数 +# ---------------------------------------------------------------------------- +# 检索类型: "keyword", "embedding", "hybrid" +EVAL_SEARCH_TYPE=hybrid + +# 检索结果数量限制（默认值） +EVAL_SEARCH_LIMIT=12 + +# 上下文最大字符数（默认值） +EVAL_MAX_CONTEXT_CHARS=8000 + +# ---------------------------------------------------------------------------- +# LLM 参数 +# ---------------------------------------------------------------------------- +# LLM 温度参数（0.0 = 确定性输出） +EVAL_LLM_TEMPERATURE=0.0 + +# LLM 最大生成 token 数 +EVAL_LLM_MAX_TOKENS=32 + +# LLM 超时时间（秒） +EVAL_LLM_TIMEOUT=10.0 + +# LLM 最大重试次数 +EVAL_LLM_MAX_RETRIES=1 + +# ---------------------------------------------------------------------------- +# 数据处理参数 +# ---------------------------------------------------------------------------- +# Chunker 策略 +EVAL_CHUNKER_STRATEGY=RecursiveChunker + +# 是否在导入前清空现有数据 +EVAL_RESET_ON_INGEST=true + +# 是否保存详细日志 +EVAL_SAVE_DETAILED_LOGS=true + +# ============================================================================ +# LoCoMo 基准测试专用配置 +# ============================================================================ +# 数据集：locomo10.json +# 运行：python locomo_benchmark.py --sample_size 20 +# ---------------------------------------------------------------------------- + +# Group ID（LoCoMo 专用） +LOCOMO_GROUP_ID=locomo_benchmark + +# 测试样本数量 +# 建议值：20（快速测试）、100（中等测试）、1986（完整测试） +LOCOMO_SAMPLE_SIZE=20 + +# 检索结果数量限制 +LOCOMO_SEARCH_LIMIT=12 + +# 上下文最大字符数 +LOCOMO_CONTEXT_CHAR_BUDGET=8000 + +# 导入的对话数量 +LOCOMO_MAX_DIALOGUES=1 + +# 跳过数据摄入（true=跳过，false=摄入） +# 首次运行设置为 false，后续运行可设置为 true 以节省时间 +LOCOMO_SKIP_INGEST=false + +# 结果保存目录 +LOCOMO_OUTPUT_DIR=locomo/results + +# ============================================================================ +# LongMemEval 基准测试专用配置 +# ============================================================================ +# 数据集：longmemeval_oracle_zh.json +# 运行：python longmemeval_benchmark.py --sample_size 3 +# 特点：支持时间推理问题的增强检索 +# ---------------------------------------------------------------------------- + +# Group ID（LongMemEval 专用） +LONGMEMEVAL_GROUP_ID=longmemeval_zh_bak_3 + +# 测试样本数量（<=0 表示全部样本） +LONGMEMEVAL_SAMPLE_SIZE=3 + +# 起始样本索引 +LONGMEMEVAL_START_INDEX=0 + +# 检索结果数量限制 +LONGMEMEVAL_SEARCH_LIMIT=8 + +# 上下文最大字符数 +LONGMEMEVAL_CONTEXT_CHAR_BUDGET=4000 + +# LLM 最大生成 token 数 +LONGMEMEVAL_LLM_MAX_TOKENS=16 + +# 每条样本最多摄入的上下文段数 +LONGMEMEVAL_MAX_CONTEXTS_PER_ITEM=2 + +# 是否保存分块结果 +LONGMEMEVAL_SAVE_CHUNK_OUTPUT=true + +# 自定义分块输出路径（留空使用默认） +LONGMEMEVAL_SAVE_CHUNK_OUTPUT_PATH= + +# 摄入前是否清空组数据 +LONGMEMEVAL_RESET_GROUP_BEFORE_INGEST=false + +# 是否跳过摄入，仅检索评估 +LONGMEMEVAL_SKIP_INGEST=false + +# 结果保存目录 +LONGMEMEVAL_OUTPUT_DIR=longmemeval/results + +# ============================================================================ +# MemSciQA 基准测试专用配置 +# ============================================================================ +# 数据集：msc_self_instruct.jsonl +# 运行：python memsciqa_benchmark.py --sample_size 1 +# 特点：对话记忆检索评估 +# ---------------------------------------------------------------------------- + +# Group ID（MemSciQA 专用，独立数据集） +MEMSCIQA_GROUP_ID=memsciqa_benchmark + +# 测试样本数量 +MEMSCIQA_SAMPLE_SIZE=1 # 0或者-1标识测试数据集中的所有样本 + +# 检索结果数量限制 +MEMSCIQA_SEARCH_LIMIT=8 + +# 上下文最大字符数 +MEMSCIQA_CONTEXT_CHAR_BUDGET=4000 + +# LLM 最大生成 token 数 +MEMSCIQA_LLM_MAX_TOKENS=64 + +# 跳过数据摄入（true=跳过，false=摄入） +# 首次运行设置为 false，后续运行可设置为 true 以节省时间 +MEMSCIQA_SKIP_INGEST=false + +# 结果保存目录（相对于 memsciqa 脚本所在目录） +# 使用 "results" 会保存到 api/app/core/memory/evaluation/memsciqa/results/ +MEMSCIQA_OUTPUT_DIR=results + +# ============================================================================ +# 高级配置（可选） +# ============================================================================ + +# BM25 权重（用于混合检索，0.0-1.0） +EVAL_RERANK_ALPHA=0.6 + +# 是否使用遗忘重排序 +EVAL_USE_FORGETTING_RERANK=false + +# 是否使用 LLM 重排序 +EVAL_USE_LLM_RERANK=false + +# 连接重置间隔（每 N 个问题重置一次） +EVAL_RESET_INTERVAL=5 + +# 性能阈值（低于此值触发重置） +EVAL_PERFORMANCE_THRESHOLD=0.6 + +# ============================================================================ +# 快速配置指南 +# ============================================================================ +# 1. 复制此文件为 .env.evaluation +# 2. 修改 EVAL_LLM_ID 和 EVAL_EMBEDDING_ID 为你的模型 ID +# 3. 根据需要修改各基准测试的专用配置 +# 4. 运行测试： +# - LoCoMo: python locomo/locomo_benchmark.py --sample_size 20 +# - LongMemEval: python longmemeval/longmemeval_benchmark.py --sample_size 3 --all +# - MemSciQA: python memsciqa/memsciqa_benchmark.py --sample_size 10 +# 配置优先级： +# 命令行参数 > 特定配置（如 LOCOMO_*）> 通用配置（EVAL_*）> 代码默认值 +# ============================================================================ + + +# 执行LoCoMo测试 +# 只摄入前5条消息，评估3个问题（最小测试） +# python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 3 --max_ingest_messages 5 +# +# 如果数据已经摄入，跳过摄入阶段直接测试 +# python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 5 --skip_ingest + + +# 执行longmemeval测试 +# python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --sample-size 10 --max-contexts-per-item 3 --reset-group-before-ingest + +# 执行memsciqa测试 +# python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --sample-size 1 diff --git a/api/app/core/memory/evaluation/.gitignore b/api/app/core/memory/evaluation/.gitignore new file mode 100644 index 00000000..38b1055a --- /dev/null +++ b/api/app/core/memory/evaluation/.gitignore @@ -0,0 +1,13 @@ +# 忽略实际的评估配置文件（包含敏感信息） +.env.evaluation + +# 保留示例文件 +!.env.evaluation.example + +# 忽略测试结果文件 +*/results/*.json +*/results/*.log + +# 忽略数据集文件（文件过大，不应提交到 Git） +dataset/*.json +dataset/*.jsonl diff --git a/api/app/core/memory/evaluation/benchmark.md b/api/app/core/memory/evaluation/benchmark.md index 2853b22b..7c31cccd 100644 --- a/api/app/core/memory/evaluation/benchmark.md +++ b/api/app/core/memory/evaluation/benchmark.md @@ -1,30 +1,748 @@ -⏬数据集下载地址： - Locomo10.json：https://github.com/snap-research/locomo/tree/main/data - LongMemEval_oracle.json：https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned - msc_self_instruct.jsonl:https://huggingface.co/datasets/MemGPT/MSC-Self-Instruct - 上方数据集下载好后全部放入app/core/memory/data文件夹中 +# 1.数据集下载地址 +Locomo10.json : https://github.com/snap-research/locomo/tree/main/data +LongMemEval_oracle.json : https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned +msc_self_instruct.jsonl : https://huggingface.co/datasets/MemGPT/MSC-Self-Instruct -全流程基准测试运行： - locomo： - python -m app.core.memory.evaluation.run_eval --dataset locomo --sample-size 1 --reset-group --group-id yyw1 --search-type hybrid --search-limit 8 --context-char-budget 12000 --llm-max-tokens 32 - LongMemEval： - python -m app.core.memory.evaluation.run_eval --dataset longmemeval --sample-size 10 --start-index 0 --group-id longmemeval_zh_bak_2 --search-limit 8 --context-char-budget 4000 --search-type hybrid --max-contexts-per-item 2 --reset-group - memsciqa： - python -m app.core.memory.evaluation.run_eval --dataset memsciqa --sample-size 10 --reset-group --group-id group_memsci +数据集下载之后保存至api\app\core\memory\evaluation\dataset目录下 +# 2.配置说明 +文件api\app\core\memory\evaluation\.env.evaluation.example对三个基准测试所需配置有着详细的说明 +**实际配置文件**：api\app\core\memory\evaluation\.env.evaluation +```python +# 当使用不带配置参数的命令行执行基准测试，基准测试所需的配置参数根据.env.evaluation中的参数执行 +python -m app.core.memory.evaluation.locomo.locomo_benchmark +``` +**检查neo4j指定的grou_id是否摄入数据** +```python +# 1. 进入交互模式 +python -m app.core.memory.evaluation.check_enduser_data -单独检索评估运行命令： - python -m app.core.memory.evaluation.locomo.locomo_test - python -m app.core.memory.evaluation.longmemeval.test_eval - python -m app.core.memory.evaluation.memsciqa.memsciqa-test - 需要先在项目中修改需要检测评估的group_id。 +# 2. 选择 "1" 检查指定 group +# 3. 输入 group_id，例如: locomo_benchmark +# 4. 选择是否显示详细统计 (y/n) +``` +# 3.locomo -参数及解释： - ● --dataset longmemeval - 指定数据集 - ● --sample-size 10 - 评估10个样本 - ● --start-index 0 - 从第0个样本开始 - ● --group-id longmemeval_zh_bak_2 - 使用指定的组ID - ● --search-limit 8 - 检索限制8条 - ● --context-char-budget 4000 - 上下文字符预算4000 - ● --search-type hybrid - 使用混合检索 - ● --max-contexts-per-item 2 - 每个样本最多摄入2个上下文 - ● --reset-group - 运行前清空组数据 \ No newline at end of file +### (1)locomo执行命令 +```python +# 首先进入api目录 +cd api + +# 只摄入前5条消息，评估3个问题（最小测试） +python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 3 --max_ingest_messages 5 + +# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) +python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 5 --skip_ingest +``` +### (2)locomo结果说明 + +#### 结果示例 +```json +{ + "dataset": "locomo", + "sample_size": 0, + "timestamp": "2026-01-26T11:24:28.239156", + "params": { + "group_id": "locomo_benchmark", + "search_type": "hybrid", + "search_limit": 12, + "context_char_budget": 8000, + "llm_id": "2c9b0782-7a85-4740-ba84-4baf77f256c4", + "embedding_id": "e2a6392d-ca63-4d59-a523-647420b59cb2" + }, + "overall_metrics": { + "f1": 0.0, + "bleu1": 0.0, + "jaccard": 0.0, + "locomo_f1": 0.0 + }, + "by_category": {}, + "latency": { + "search": { + "mean": 0.0, + "p50": 0.0, + "p95": 0.0, + "iqr": 0.0 + }, + "llm": { + "mean": 0.0, + "p50": 0.0, + "p95": 0.0, + "iqr": 0.0 + } + }, + "context_stats": { + "avg_retrieved_docs": 0.0, + "avg_context_chars": 0.0, + "avg_context_tokens": 0.0 + }, + "samples": [] +} +``` + +#### 参数详解 + +##### 1. 核心评估指标 (overall_metrics) + +**🎯 关键进步指标：** + +- **`f1`** (F1 Score): 精确率和召回率的调和平均值 + - 范围：0.0 - 1.0 + - **越高越好**，衡量检索和生成答案的准确性 + - 这是最重要的综合性能指标 + - 优秀标准：> 0.85 + +- **`bleu1`** (BLEU-1): 单词级别的匹配度 + - 范围：0.0 - 1.0 + - **越高越好**，衡量生成答案与标准答案的词汇重叠度 + - 关注词汇层面的准确性 + +- **`jaccard`** (Jaccard 相似度): 集合相似度 + - 范围：0.0 - 1.0 + - **越高越好**，衡量答案集合的相似性 + - 计算公式：交集大小 / 并集大小 + +- **`locomo_f1`**: Locomo 特定的 F1 分数 + - 范围：0.0 - 1.0 + - **越高越好**，针对 Locomo 数据集优化的评估指标 + - 考虑了长对话记忆的特殊性 + +##### 2. 性能指标 (latency) + +**⚡ 关键效率指标：** + +- **`search`**: 检索延迟统计（单位：毫秒） + - `mean`: 平均延迟 + - `p50`: 中位数延迟（50%的请求在此时间内完成） + - `p95`: 95分位数延迟（95%的请求在此时间内完成） + - `iqr`: 四分位距（Q3-Q1，衡量稳定性） + - **越低越好**，衡量记忆检索速度 + - 优秀标准：p95 < 2000ms + +- **`llm`**: LLM 推理延迟统计（单位：毫秒） + - `mean`: 平均推理时间 + - `p50`: 中位数推理时间 + - `p95`: 95分位数推理时间 + - `iqr`: 四分位距（越小越稳定） + - **越低越好**，衡量答案生成速度 + - 优秀标准：p95 < 3000ms + +##### 3. 上下文统计 (context_stats) + +**📊 资源效率指标：** + +- **`avg_retrieved_docs`**: 平均检索文档数 + - 反映检索策略的广度 + - 需要平衡：太少可能信息不足，太多增加噪音和延迟 + - 建议范围：8-15 个文档 + +- **`avg_context_chars`**: 平均上下文字符数 + - 反映检索内容的总量 + - 应在满足准确性前提下尽量精简 + - 受 `context_char_budget` 参数限制 + +- **`avg_context_tokens`**: 平均上下文 token 数 + - **越低越好**（在保持准确性前提下） + - 直接影响 API 调用成本和推理速度 + - 成本效益比 = f1 / avg_context_tokens + +##### 4. 分类统计 (by_category) + +- 按问题类型分类的性能指标 +- 帮助识别系统在不同场景下的强弱项 +- 可针对性优化特定类型的问题 + +#### 系统进步衡量标准 + +**一级指标（最重要）：** +- `f1` 和 `locomo_f1` 提升 → 核心能力提升 +- 目标：f1 > 0.85 + +**二级指标（重要）：** +- `latency.p95` 降低 → 用户体验提升 +- 目标：search.p95 < 2000ms, llm.p95 < 3000ms + +**三级指标（辅助）：** +- `avg_context_tokens` 降低（在保持 f1 前提下）→ 成本优化 +- `iqr` 降低 → 性能稳定性提升 +# 4.longmemeval +支持时间推理问题的增强检索 +### (1)执行命令 +```python +# 首先进入api目录 +cd api + +# 不带参数运行 - 使用环境变量 +python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark + +# 命令行参数覆盖环境变量 +python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --sample-size 2 + +# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) +python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --skip_ingest +``` +### (2)结果说明 + +#### 结果示例 +```json +{ + "dataset": "longmemeval", + "items": 1, + "accuracy_by_type": { + "single-session-user": 1.0 + }, + "f1_by_type": { + "single-session-user": 1.0 + }, + "jaccard_by_type": { + "single-session-user": 1.0 + }, + "samples": [ + { + "question": "What degree did I graduate with?", + "prediction": "Business Administration", + "answer": "Business Administration", + "question_type": "single-session-user", + "is_temporal": false, + "question_id": "e47becba", + "options": [], + "context_count": 13, + "context_chars": 1268, + "retrieved_dialogue_count": 0, + "retrieved_statement_count": 12, + "metrics": { + "exact_match": true, + "f1": 1.0, + "jaccard": 1.0 + }, + "timing": { + "search_ms": 1483.100175857544, + "llm_ms": 995.8682060241699 + } + } + ], + "latency": { + "search": { + "mean": 1483.100175857544, + "p50": 1483.100175857544, + "p95": 1483.100175857544, + "iqr": 0.0 + }, + "llm": { + "mean": 995.8682060241699, + "p50": 995.8682060241699, + "p95": 995.8682060241699, + "iqr": 0.0 + } + }, + "context": { + "avg_tokens": 204.0, + "avg_chars": 1268, + "count_avg": 13 + }, + "params": { + "group_id": "longmemeval_zh_bak_3", + "search_limit": 8, + "context_char_budget": 4000, + "search_type": "hybrid", + "llm_id": "6dc52e1b-9cec-4194-af66-a74c6307fc3f", + "embedding_id": "e2a6392d-ca63-4d59-a523-647420b59cb2", + "sample_size": 1, + "start_index": 0 + }, + "timestamp": "2026-01-24T21:36:10.818308", + "metric_summary": { + "score_accuracy": 100.0, + "latency_median_s": 2.478968381881714, + "latency_iqr_s": 0.0, + "avg_context_tokens_k": 0.204 + }, + "diagnostics": { + "duplicate_previews_top": [], + "unique_preview_count": 1 + } +} +``` + +#### 参数详解 + +##### 1. 核心评估指标 + +**🎯 关键进步指标：** + +- **`accuracy_by_type`**: 按问题类型分类的准确率 + - 范围：0.0 - 1.0 + - **越高越好**，1.0 表示 100% 准确 + - 问题类型包括： + - `single-session-user`: 单会话用户信息 + - `single-session-event`: 单会话事件信息 + - `multi-session-user`: 多会话用户信息 + - `multi-session-event`: 多会话事件信息 + - 可以识别系统在不同场景下的强弱项 + +- **`f1_by_type`**: 按问题类型的 F1 分数 + - 范围：0.0 - 1.0 + - **越高越好**，综合评估精确率和召回率 + - 比单纯的准确率更全面 + +- **`jaccard_by_type`**: 按问题类型的 Jaccard 相似度 + - 范围：0.0 - 1.0 + - **越高越好**，衡量答案集合匹配度 + - 对于集合类答案特别有用 + +##### 2. 样本级指标 (samples) + +**详细诊断指标：** + +- **`metrics.exact_match`**: 精确匹配（布尔值） + - **true 越多越好**，最严格的评估标准 + - 要求预测答案与标准答案完全一致 + +- **`metrics.f1`**: 单个样本的 F1 分数 + - 范围：0.0 - 1.0 + - **越高越好**，衡量单个问题的回答质量 + +- **`is_temporal`**: 是否为时间推理问题 + - 布尔值，标识问题是否涉及时间推理 + - 时间推理问题通常更具挑战性 + +- **`context_count`**: 检索到的上下文数量 + - 反映检索策略的有效性 + - 建议范围：8-15 个上下文片段 + +- **`retrieved_dialogue_count`**: 检索到的对话数 +- **`retrieved_statement_count`**: 检索到的陈述数 + - 这两个指标帮助理解检索的内容类型分布 + - 可用于优化检索策略 + +- **`timing.search_ms`**: 单个问题的检索延迟（毫秒） +- **`timing.llm_ms`**: 单个问题的 LLM 推理延迟（毫秒） + - **越低越好**，反映单次查询的响应速度 + +##### 3. 汇总指标 (metric_summary) + +**📊 关键 KPI：** + +- **`score_accuracy`**: 总体准确率百分比 + - 范围：0.0 - 100.0 + - **越高越好**，最直观的性能指标 + - 优秀标准：> 90.0 + +- **`latency_median_s`**: 中位延迟（秒） + - **越低越好**，反映真实响应速度 + - 优秀标准：< 3.0 秒 + +- **`latency_iqr_s`**: 延迟四分位距（秒） + - **越低越好**，反映性能稳定性 + - 越小说明响应时间越稳定 + +- **`avg_context_tokens_k`**: 平均上下文 token 数（千） + - **越低越好**（在保持准确性前提下） + - 直接影响 API 调用成本 + - 成本效益比 = score_accuracy / (avg_context_tokens_k * 1000) + +##### 4. 上下文统计 (context) + +- **`avg_tokens`**: 平均 token 数 +- **`avg_chars`**: 平均字符数 +- **`count_avg`**: 平均上下文片段数 + - 这些指标反映检索内容的规模 + - 需要在准确性和效率之间平衡 + +##### 5. 性能指标 (latency) + +**⚡ 效率指标：** + +- **`search`**: 检索延迟统计（单位：毫秒） + - `mean`: 平均延迟 + - `p50`: 中位数延迟 + - `p95`: 95分位数延迟 + - `iqr`: 四分位距 + - **越低越好**，衡量记忆检索速度 + +- **`llm`**: LLM 推理延迟统计（单位：毫秒） + - `mean`: 平均推理时间 + - `p50`: 中位数推理时间 + - `p95`: 95分位数推理时间 + - `iqr`: 四分位距 + - **越低越好**，衡量答案生成速度 + +##### 6. 诊断信息 (diagnostics) + +- **`duplicate_previews_top`**: 重复预览统计 + - 列出出现频率最高的重复内容 + - 帮助发现检索冗余问题 + - 应该尽量减少重复 + +- **`unique_preview_count`**: 唯一预览数量 + - 反映检索多样性 + - **越高越好**，说明检索到的内容更丰富 + +#### 系统进步衡量标准 + +**一级指标（最重要）：** +- `score_accuracy` 提升 → 核心能力提升 +- 目标：> 90.0% +- 各类型的 `accuracy_by_type` 均衡提升 → 全面能力提升 + +**二级指标（重要）：** +- `latency_median_s` 降低 → 用户体验提升 +- 目标：< 3.0 秒 +- `exact_match` 比例提升 → 精确度提升 + +**三级指标（辅助）：** +- `avg_context_tokens_k` 降低（在保持准确性前提下）→ 成本优化 +- `unique_preview_count` 提升 → 检索多样性提升 +- `latency_iqr_s` 降低 → 性能稳定性提升 + +**特殊关注：** +- 时间推理问题（`is_temporal: true`）的准确率 +- 多会话问题的准确率（通常更具挑战性） +# 5.memsciqa +对话记忆检索评估 +### (1)执行命令 +```python +# 首先进入api目录 +cd api + +# 不带参数运行 - 使用环境变量 +python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark + +# 命令行参数覆盖环境变量 +python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --sample-size 100 + +# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) +python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --skip_ingest +``` +### (2)结果说明 + +#### 结果示例 +```json +{ + "dataset": "memsciqa", + "items": 1, + "metrics": { + "accuracy": 0.0, + "f1": 0.0, + "bleu1": 0.0, + "jaccard": 0.0 + }, + "latency": { + "search": { + "mean": 0.0, + "p50": 0.0, + "p95": 0.0, + "iqr": 0.0 + }, + "llm": { + "mean": 3067.7285194396973, + "p50": 3067.7285194396973, + "p95": 3067.7285194396973, + "iqr": 0.0 + } + }, + "avg_context_tokens": 4.0 +} +``` + +#### 参数详解 + +##### 1. 核心评估指标 (metrics) + +**🎯 关键进步指标：** + +- **`accuracy`**: 准确率 + - 范围：0.0 - 1.0 + - **越高越好**，最直接的性能指标 + - 衡量系统回答正确的问题比例 + - 优秀标准：> 0.85 + +- **`f1`**: F1 分数 + - 范围：0.0 - 1.0 + - **越高越好**，平衡精确率和召回率 + - 计算公式：2 * (precision * recall) / (precision + recall) + - 比单纯的准确率更全面，特别适合不平衡数据集 + +- **`bleu1`**: BLEU-1 分数 + - 范围：0.0 - 1.0 + - **越高越好**，衡量词汇级别的匹配度 + - 关注生成答案与标准答案的单词重叠 + - 源自机器翻译评估，适用于自然语言生成 + +- **`jaccard`**: Jaccard 相似度 + - 范围：0.0 - 1.0 + - **越高越好**，衡量集合相似性 + - 计算公式：|A ∩ B| / |A ∪ B| + - 对于多答案或集合类问题特别有用 + +##### 2. 性能指标 (latency) + +**⚡ 效率指标：** + +- **`search`**: 检索延迟统计（单位：毫秒） + - `mean`: 平均检索延迟 + - `p50`: 中位数延迟（50%的请求在此时间内完成） + - `p95`: 95分位数延迟（95%的请求在此时间内完成） + - `iqr`: 四分位距（Q3-Q1，衡量稳定性） + - **越低越好**，衡量记忆检索效率 + - 优秀标准：p95 < 2000ms + +- **`llm`**: LLM 推理延迟统计（单位：毫秒） + - `mean`: 平均推理时间 + - `p50`: 中位数推理时间 + - `p95`: 95分位数推理时间 + - `iqr`: 四分位距（越小越稳定） + - **越低越好**，衡量答案生成速度 + - 优秀标准：p95 < 3000ms + - 注意：LLM 延迟通常占总延迟的大部分 + +##### 3. 资源指标 + +- **`avg_context_tokens`**: 平均上下文 token 数 + - **越低越好**（在保持准确性前提下） + - 直接影响： + - API 调用成本（按 token 计费） + - 推理速度（token 越多越慢） + - 上下文窗口占用 + - 成本效益比 = accuracy / avg_context_tokens + - 建议范围：根据模型上下文窗口和成本预算调整 + +##### 4. 数据集特点 + +- **`items`**: 评估的问题数量 + - 样本量越大，评估结果越可靠 + - 建议至少 100 个样本以获得稳定的评估结果 + +- **对话记忆特性**： + - MemSciQA 专注于对话历史中的记忆检索 + - 评估系统从多轮对话中提取和回忆信息的能力 + - 模拟真实的对话场景 + +#### 系统进步衡量标准 + +**一级指标（最重要）：** +- `accuracy` 提升 → 核心能力提升 +- 目标：> 0.85 +- `f1` 提升 → 综合性能提升 +- 目标：> 0.80 + +**二级指标（重要）：** +- `latency.p95` 降低 → 用户体验提升 + - search.p95 目标：< 2000ms + - llm.p95 目标：< 3000ms +- `iqr` 降低 → 性能稳定性提升 + +**三级指标（辅助）：** +- `avg_context_tokens` 降低（在保持准确性前提下）→ 成本优化 +- `bleu1` 和 `jaccard` 提升 → 答案质量提升 + +**综合评估：** +- 成本效益比 = accuracy / avg_context_tokens + - 该比值越高，说明系统在相同成本下性能越好 +- 总延迟 = search.p95 + llm.p95 + - 应控制在 5 秒以内以保证良好的用户体验 + +#### 优化建议 + +**提升准确性：** +- 优化检索算法（调整 hybrid search 参数） +- 改进 embedding 模型质量 +- 增加检索上下文数量（`search_limit`） +- 优化 prompt 工程 + +**提升效率：** +- 减少不必要的检索文档 +- 使用更快的 LLM 模型或量化版本 +- 实施缓存策略（相似问题复用结果） +- 优化数据库索引 + +**平衡性能：** +- 监控 accuracy vs latency 的权衡 +- 监控 accuracy vs cost (tokens) 的权衡 +- 根据业务需求调整优先级 + + +--- + +# 6. 三个基准测试对比总结 + +## 6.1 测试特点对比 + +| 基准测试 | 主要评估目标 | 数据集特点 | 适用场景 | +|---------|------------|-----------|---------| +| **Locomo** | 长对话记忆检索 | 长对话历史，多轮交互 | 评估长期记忆保持和检索能力 | +| **LongMemEval** | 时间推理和多会话记忆 | 支持时间推理，多会话场景 | 评估时间感知和跨会话记忆能力 | +| **MemSciQA** | 对话记忆问答 | 对话历史问答 | 评估对话上下文理解和记忆提取 | + +## 6.2 核心指标对比 + +### 准确性指标 + +| 指标 | Locomo | LongMemEval | MemSciQA | 说明 | +|-----|--------|-------------|----------|------| +| **F1 Score** | ✅ | ✅ | ✅ | 所有测试都使用，最重要的综合指标 | +| **Accuracy** | ❌ | ✅ | ✅ | 直观的准确率指标 | +| **BLEU-1** | ✅ | ❌ | ✅ | 词汇级别匹配度 | +| **Jaccard** | ✅ | ✅ | ✅ | 集合相似度 | +| **Exact Match** | ❌ | ✅ | ❌ | 最严格的评估标准 | + +### 性能指标 + +所有三个测试都包含： +- **检索延迟** (search latency): mean, p50, p95, iqr +- **LLM 延迟** (llm latency): mean, p50, p95, iqr +- **上下文统计**: token 数、字符数、文档数 + +## 6.3 关键进步指标优先级 + +### 🥇 一级指标（必须关注） + +1. **准确性指标** + - Locomo: `f1`, `locomo_f1` + - LongMemEval: `score_accuracy`, `accuracy_by_type` + - MemSciQA: `accuracy`, `f1` + - **目标**: > 85% 或 > 0.85 + +2. **综合性能** + - 所有测试的 F1 分数应保持一致性 + - 不同类型问题的准确率应均衡 + +### 🥈 二级指标（重要） + +3. **响应延迟** + - `latency.p95` (95分位数延迟) + - **目标**: + - search.p95 < 2000ms + - llm.p95 < 3000ms + - 总延迟 < 5000ms + +4. **性能稳定性** + - `iqr` (四分位距) + - **目标**: 越小越好，说明性能稳定 + +### 🥉 三级指标（优化） + +5. **成本效率** + - `avg_context_tokens` + - **目标**: 在保持准确性前提下最小化 + - 成本效益比 = accuracy / avg_context_tokens + +6. **检索质量** + - `avg_retrieved_docs` 的合理性 + - `unique_preview_count` (LongMemEval) + - 检索内容的多样性和相关性 + +## 6.4 系统优化路径 + +### 阶段一：提升准确性（优先级最高） + +**目标**: 所有测试的准确率 > 85% + +**优化方向**: +1. 改进 embedding 模型质量 +2. 优化检索算法（hybrid search 参数） +3. 增加检索上下文数量（`search_limit`） +4. 优化 prompt 工程 +5. 改进记忆存储结构 + +**监控指标**: +- Locomo: `f1`, `locomo_f1` +- LongMemEval: `score_accuracy`, `exact_match` 比例 +- MemSciQA: `accuracy`, `f1` + +### 阶段二：优化性能（准确性达标后） + +**目标**: p95 延迟 < 5 秒，性能稳定 + +**优化方向**: +1. 优化数据库索引和查询 +2. 实施缓存策略 +3. 使用更快的 LLM 模型 +4. 并行化检索和推理 +5. 减少不必要的检索 + +**监控指标**: +- `latency.p50`, `latency.p95` +- `iqr` (稳定性) +- 各阶段耗时分布 + +### 阶段三：降低成本（性能达标后） + +**目标**: 在保持准确性和性能前提下，最小化成本 + +**优化方向**: +1. 精简检索上下文 +2. 优化 context 选择策略 +3. 使用更小的 LLM 模型 +4. 实施智能缓存 +5. 批处理优化 + +**监控指标**: +- `avg_context_tokens` +- 成本效益比 = accuracy / avg_context_tokens +- API 调用成本 + +## 6.5 评估最佳实践 + +### 测试执行建议 + +1. **初始测试**: 使用小样本快速验证 + ```bash + --sample_size 10 + ``` + +2. **完整评估**: 使用足够大的样本量 + ```bash + --sample_size 100 # 或更多 + ``` + +3. **增量测试**: 数据已摄入时跳过摄入阶段 + ```bash + --skip_ingest + ``` + +4. **参数调优**: 系统性地调整参数并记录结果 + - 调整 `search_limit`: 4, 8, 12, 16 + - 调整 `context_char_budget`: 2000, 4000, 8000 + - 尝试不同的 `search_type`: vector, keyword, hybrid + +### 结果分析建议 + +1. **横向对比**: 比较三个测试的结果，识别系统的强弱项 +2. **纵向对比**: 跟踪同一测试在不同版本的表现 +3. **分类分析**: 关注不同问题类型的性能差异 +4. **异常诊断**: 分析失败案例，找出根本原因 + +### 持续监控 + +建议建立监控仪表板，跟踪： +- 核心指标趋势（准确率、延迟） +- 成本效益比趋势 +- 不同问题类型的性能分布 +- 异常样本和失败模式 + +## 6.6 性能基准参考 + +### 优秀水平（Production Ready） + +- **准确性**: accuracy/f1 > 0.90 +- **延迟**: p95 < 3 秒 +- **稳定性**: iqr < 500ms +- **成本效益**: accuracy/tokens > 0.0001 + +### 良好水平（Acceptable） + +- **准确性**: accuracy/f1 > 0.85 +- **延迟**: p95 < 5 秒 +- **稳定性**: iqr < 1000ms +- **成本效益**: accuracy/tokens > 0.00005 + +### 需要改进（Below Target） + +- **准确性**: accuracy/f1 < 0.85 +- **延迟**: p95 > 5 秒 +- **稳定性**: iqr > 1000ms +- **成本效益**: accuracy/tokens < 0.00005 + +--- + +**注**: 以上标准仅供参考，实际目标应根据具体业务需求和资源约束调整。 diff --git a/api/app/core/memory/evaluation/check_enduser_data.py b/api/app/core/memory/evaluation/check_enduser_data.py new file mode 100644 index 00000000..18ecbb34 --- /dev/null +++ b/api/app/core/memory/evaluation/check_enduser_data.py @@ -0,0 +1,371 @@ +""" +交互式 Neo4j End User 数据检查工具 + +用于查询指定 end_user_id 在 Neo4j 中是否存在数据，以及数据的详细统计信息。 + +使用方法: + python check_group_data.py + python check_group_data.py --group-id locomo_benchmark + python check_group_data.py --group-id memsciqa_benchmark --detailed +""" + +import asyncio +import argparse +import os +from pathlib import Path +from typing import Dict, Any +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}\n") + +from app.repositories.neo4j.neo4j_connector import Neo4jConnector + + +async def check_group_exists(end_user_id: str) -> Dict[str, Any]: + """ + 检查指定 end_user_id 是否存在数据 + + Args: + end_user_id: 要检查的 end_user ID + + Returns: + 包含统计信息的字典 + """ + connector = Neo4jConnector() + + try: + # 查询该 end_user 的节点总数 + query_total = """ + MATCH (n {end_user_id: $end_user_id}) + RETURN count(n) as total_nodes + """ + result_total = await connector.execute_query(query_total, end_user_id=end_user_id) + total_nodes = result_total[0]["total_nodes"] if result_total else 0 + + # 查询各类型节点的数量 + query_by_type = """ + MATCH (n {end_user_id: $end_user_id}) + RETURN labels(n) as labels, count(n) as count + ORDER BY count DESC + """ + result_by_type = await connector.execute_query(query_by_type, end_user_id=end_user_id) + + # 查询关系数量 + query_relationships = """ + MATCH (n {end_user_id: $end_user_id})-[r]-() + RETURN count(DISTINCT r) as total_relationships + """ + result_rel = await connector.execute_query(query_relationships, end_user_id=end_user_id) + total_relationships = result_rel[0]["total_relationships"] if result_rel else 0 + + return { + "exists": total_nodes > 0, + "total_nodes": total_nodes, + "total_relationships": total_relationships, + "nodes_by_type": result_by_type + } + + finally: + await connector.close() + + +async def get_detailed_stats(end_user_id: str) -> Dict[str, Any]: + """ + 获取详细的统计信息 + + Args: + end_user_id: 要检查的 end_user ID + + Returns: + 详细统计信息字典 + """ + connector = Neo4jConnector() + + try: + stats = {} + + # Chunk 节点统计 + query_chunks = """ + MATCH (c:Chunk {end_user_id: $end_user_id}) + RETURN count(c) as count, + avg(size(c.content)) as avg_content_length + """ + result_chunks = await connector.execute_query(query_chunks, end_user_id=end_user_id) + if result_chunks and result_chunks[0]["count"] > 0: + stats["chunks"] = { + "count": result_chunks[0]["count"], + "avg_content_length": int(result_chunks[0]["avg_content_length"]) if result_chunks[0]["avg_content_length"] else 0 + } + + # Statement 节点统计 + query_statements = """ + MATCH (s:Statement {end_user_id: $end_user_id}) + RETURN count(s) as count + """ + result_statements = await connector.execute_query(query_statements, end_user_id=end_user_id) + if result_statements and result_statements[0]["count"] > 0: + stats["statements"] = { + "count": result_statements[0]["count"] + } + + # Entity 节点统计 + query_entities = """ + MATCH (e:Entity {end_user_id: $end_user_id}) + RETURN count(e) as count, + count(DISTINCT e.entity_type) as unique_types + """ + result_entities = await connector.execute_query(query_entities, end_user_id=end_user_id) + if result_entities and result_entities[0]["count"] > 0: + stats["entities"] = { + "count": result_entities[0]["count"], + "unique_types": result_entities[0]["unique_types"] + } + + # Dialogue 节点统计 + query_dialogues = """ + MATCH (d:Dialogue {end_user_id: $end_user_id}) + RETURN count(d) as count + """ + result_dialogues = await connector.execute_query(query_dialogues, end_user_id=end_user_id) + if result_dialogues and result_dialogues[0]["count"] > 0: + stats["dialogues"] = { + "count": result_dialogues[0]["count"] + } + + # Summary 节点统计 + query_summaries = """ + MATCH (s:Summary {end_user_id: $end_user_id}) + RETURN count(s) as count + """ + result_summaries = await connector.execute_query(query_summaries, end_user_id=end_user_id) + if result_summaries and result_summaries[0]["count"] > 0: + stats["summaries"] = { + "count": result_summaries[0]["count"] + } + + return stats + + finally: + await connector.close() + + +async def list_all_end_users() -> list: + """ + 列出数据库中所有的 end_user_id + + Returns: + end_user_id 列表及其节点数量 + """ + connector = Neo4jConnector() + + try: + query = """ + MATCH (n) + WHERE n.end_user_id IS NOT NULL + RETURN DISTINCT n.end_user_id as end_user_id, count(n) as node_count + ORDER BY node_count DESC + """ + results = await connector.execute_query(query) + return results + + finally: + await connector.close() + + +def print_results(end_user_id: str, stats: Dict[str, Any], detailed_stats: Dict[str, Any] = None): + """ + 打印查询结果 + + Args: + end_user_id: End User ID + stats: 基本统计信息 + detailed_stats: 详细统计信息（可选） + """ + print(f"\n{'='*60}") + print(f"📊 End User ID: {end_user_id}") + print(f"{'='*60}\n") + + if not stats["exists"]: + print("❌ 该 end_user_id 不存在数据") + print("\n💡 提示: 请先运行基准测试以摄入数据") + return + + print(f"✅ 该 end_user_id 存在数据\n") + print(f"📈 基本统计:") + print(f" 总节点数: {stats['total_nodes']}") + print(f" 总关系数: {stats['total_relationships']}") + + if stats["nodes_by_type"]: + print(f"\n📋 节点类型分布:") + for item in stats["nodes_by_type"]: + labels = ", ".join(item["labels"]) + count = item["count"] + print(f" {labels}: {count}") + + if detailed_stats: + print(f"\n🔍 详细统计:") + + if "chunks" in detailed_stats: + print(f" Chunks: {detailed_stats['chunks']['count']} 个") + print(f" 平均内容长度: {detailed_stats['chunks']['avg_content_length']} 字符") + + if "statements" in detailed_stats: + print(f" Statements: {detailed_stats['statements']['count']} 个") + + if "entities" in detailed_stats: + print(f" Entities: {detailed_stats['entities']['count']} 个") + print(f" 唯一类型数: {detailed_stats['entities']['unique_types']}") + + if "dialogues" in detailed_stats: + print(f" Dialogues: {detailed_stats['dialogues']['count']} 个") + + if "summaries" in detailed_stats: + print(f" Summaries: {detailed_stats['summaries']['count']} 个") + + print(f"\n{'='*60}\n") + + +async def interactive_mode(): + """ + 交互式模式 + """ + print("\n" + "="*60) + print("🔍 Neo4j End User 数据检查工具 - 交互模式") + print("="*60 + "\n") + + while True: + print("\n请选择操作:") + print(" 1. 检查指定 end_user_id") + print(" 2. 列出所有 end_user_id") + print(" 3. 退出") + + choice = input("\n请输入选项 (1-3): ").strip() + + if choice == "1": + end_user_id = input("\n请输入 end_user_id: ").strip() + if not end_user_id: + print("❌ end_user_id 不能为空") + continue + + detailed = input("是否显示详细统计? (y/n, 默认 n): ").strip().lower() == 'y' + + print("\n🔄 正在查询...") + stats = await check_group_exists(end_user_id) + + detailed_stats = None + if detailed and stats["exists"]: + detailed_stats = await get_detailed_stats(end_user_id) + + print_results(end_user_id, stats, detailed_stats) + + elif choice == "2": + print("\n🔄 正在查询所有 end_user_id...") + end_users = await list_all_end_users() + + if not end_users: + print("\n❌ 数据库中没有任何 end_user 数据") + else: + print(f"\n{'='*60}") + print(f"📋 数据库中的所有 End User ID") + print(f"{'='*60}\n") + + for idx, end_user in enumerate(end_users, 1): + print(f" {idx}. {end_user['end_user_id']}") + print(f" 节点数: {end_user['node_count']}") + + print(f"\n{'='*60}\n") + + elif choice == "3": + print("\n👋 再见！") + break + + else: + print("\n❌ 无效的选项，请重新选择") + + +async def main(): + """ + 主函数 + """ + parser = argparse.ArgumentParser( + description="检查 Neo4j 中指定 end_user_id 的数据情况", + formatter_class=argparse.RawDescriptionHelpFormatter, + epilog=""" +示例: + # 交互模式 + python check_group_data.py + + # 检查指定 end_user + python check_group_data.py --end-user-id locomo_benchmark + + # 检查并显示详细统计 + python check_group_data.py --end-user-id memsciqa_benchmark --detailed + + # 列出所有 end_user + python check_group_data.py --list-all + """ + ) + + parser.add_argument( + "--end-user-id", + type=str, + help="要检查的 end_user ID" + ) + + parser.add_argument( + "--detailed", + action="store_true", + help="显示详细统计信息" + ) + + parser.add_argument( + "--list-all", + action="store_true", + help="列出所有 end_user_id" + ) + + args = parser.parse_args() + + # 如果没有提供任何参数，进入交互模式 + if not args.end_user_id and not args.list_all: + await interactive_mode() + return + + # 列出所有 end_user + if args.list_all: + print("\n🔄 正在查询所有 end_user_id...") + end_users = await list_all_end_users() + + if not end_users: + print("\n❌ 数据库中没有任何 end_user 数据") + else: + print(f"\n{'='*60}") + print(f"📋 数据库中的所有 End User ID") + print(f"{'='*60}\n") + + for idx, end_user in enumerate(end_users, 1): + print(f" {idx}. {end_user['end_user_id']}") + print(f" 节点数: {end_user['node_count']}") + + print(f"\n{'='*60}\n") + return + + # 检查指定 end_user + if args.end_user_id: + print(f"\n🔄 正在查询 end_user_id: {args.end_user_id}...") + stats = await check_group_exists(args.end_user_id) + + detailed_stats = None + if args.detailed and stats["exists"]: + print("🔄 正在获取详细统计...") + detailed_stats = await get_detailed_stats(args.end_user_id) + + print_results(args.end_user_id, stats, detailed_stats) + + +if __name__ == "__main__": + asyncio.run(main()) diff --git a/api/app/core/memory/evaluation/common/metrics.py b/api/app/core/memory/evaluation/common/metrics.py index acc27fb9..961ce7f0 100644 --- a/api/app/core/memory/evaluation/common/metrics.py +++ b/api/app/core/memory/evaluation/common/metrics.py @@ -2,7 +2,7 @@ import math import re from typing import List, Dict - +# 评估指标的实现 def _normalize(text: str) -> List[str]: """Lowercase, strip punctuation, and split into tokens.""" text = text.lower().strip() diff --git a/api/app/core/memory/evaluation/dialogue_queries.py b/api/app/core/memory/evaluation/dialogue_queries.py index 25abe64e..0aace0ec 100644 --- a/api/app/core/memory/evaluation/dialogue_queries.py +++ b/api/app/core/memory/evaluation/dialogue_queries.py @@ -4,15 +4,17 @@ This file contains Cypher queries for searching dialogues, entities, and chunks. Placed in evaluation directory to avoid circular imports with src modules. """ +# 应该是neo4j browser的cypher语句，需要修改文件名 + # Entity search queries SEARCH_ENTITIES_BY_NAME = """ -MATCH (e:Entity) +MATCH (e:ExtractedEntity) WHERE e.name = $name RETURN e """ SEARCH_ENTITIES_BY_NAME_FALLBACK = """ -MATCH (e:Entity) +MATCH (e:ExtractedEntity) WHERE e.name CONTAINS $name RETURN e """ diff --git a/api/app/core/memory/evaluation/extraction_utils.py b/api/app/core/memory/evaluation/extraction_utils.py index 9e70bc28..43ef6fe0 100644 --- a/api/app/core/memory/evaluation/extraction_utils.py +++ b/api/app/core/memory/evaluation/extraction_utils.py @@ -1,34 +1,33 @@ +import os import asyncio import json -import os -import re +from typing import List, Dict, Any, Optional from datetime import datetime -from typing import Any, Dict, List, Optional +from uuid import UUID +import re from app.core.memory.llm_tools.openai_client import LLMClient -from app.core.memory.models.message_models import ( - ConversationContext, - ConversationMessage, - DialogData, -) +from app.core.memory.storage_services.extraction_engine.knowledge_extraction.chunk_extraction import DialogueChunker +from app.core.memory.models.message_models import DialogData, ConversationContext, ConversationMessage +import os +import sys +from pathlib import Path +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent / "app" / "core" / "memory" / "evaluation" / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") + +from app.repositories.neo4j.neo4j_connector import Neo4jConnector +from app.core.memory.utils.llm.llm_utils import get_llm_client # 使用新的模块化架构 -from app.core.memory.storage_services.extraction_engine.extraction_orchestrator import ( - ExtractionOrchestrator, -) -from app.core.memory.storage_services.extraction_engine.knowledge_extraction.chunk_extraction import ( - DialogueChunker, -) -from app.core.memory.utils.config.definitions import ( - SELECTED_CHUNKER_STRATEGY, - SELECTED_EMBEDDING_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.db import get_db_context +from app.core.memory.storage_services.extraction_engine.extraction_orchestrator import ExtractionOrchestrator # Import from database module from app.repositories.neo4j.graph_saver import save_dialog_and_statements_to_neo4j -from app.repositories.neo4j.neo4j_connector import Neo4jConnector # Cypher queries for evaluation # Note: Entity, chunk, and dialogue search queries have been moved to evaluation/dialogue_queries.py @@ -41,11 +40,14 @@ async def ingest_contexts_via_full_pipeline( embedding_name: str | None = None, save_chunk_output: bool = False, save_chunk_output_path: str | None = None, + reset_group: bool = False, ) -> bool: - """DEPRECATED: 此函数使用旧的流水线架构，建议使用新的 ExtractionOrchestrator + """ + 使用新的 ExtractionOrchestrator 运行完整的提取流水线 Run the full extraction pipeline on provided dialogue contexts and save to Neo4j. - This function mirrors the steps in main(), but starts from raw text contexts. + This function uses the new ExtractionOrchestrator architecture for better maintainability. + Args: contexts: List of dialogue texts, each containing lines like "role: message". end_user_id: Group ID to assign to generated DialogData and graph nodes. @@ -53,25 +55,59 @@ async def ingest_contexts_via_full_pipeline( embedding_name: Optional embedding model ID; defaults to SELECTED_EMBEDDING_ID. save_chunk_output: If True, write chunked DialogData list to a JSON file for debugging. save_chunk_output_path: Optional output path; defaults to src/chunker_test_output.txt. + reset_group: If True, clear existing data for this group before ingestion. Returns: True if data saved successfully, False otherwise. """ - chunker_strategy = chunker_strategy or SELECTED_CHUNKER_STRATEGY - embedding_name = embedding_name or SELECTED_EMBEDDING_ID + chunker_strategy = chunker_strategy or os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker") + embedding_name = embedding_name or os.getenv("EVAL_EMBEDDING_ID") + + # Check if we should reset from environment variable if not explicitly set + if not reset_group: + reset_group = os.getenv("EVAL_RESET_ON_INGEST", "false").lower() in ("true", "1", "yes") + + # Step 0: Reset group if requested + if reset_group: + print(f"[Ingestion] 🗑️ 清空 end_user '{end_user_id}' 的现有数据...") + try: + from app.repositories.neo4j.neo4j_connector import Neo4jConnector + connector = Neo4jConnector() + try: + # 删除该 end_user 的所有节点和关系 + query = """ + MATCH (n {end_user_id: $end_user_id}) + DETACH DELETE n + """ + await connector.execute_query(query, end_user_id=end_user_id) + print(f"[Ingestion] ✅ End User '{end_user_id}' 已清空") + finally: + await connector.close() + except Exception as e: + print(f"[Ingestion] ⚠️ 清空 end_user 失败: {e}") + # 继续执行，不中断摄入流程 - # Initialize llm client with graceful fallback + # Step 1: Initialize LLM client llm_client = None - llm_available = True try: - from app.core.memory.utils.config import definitions as config_defs - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(config_defs.SELECTED_LLM_ID) + # 使用评估配置中的 LLM ID + llm_id = os.getenv("EVAL_LLM_ID") + if not llm_id: + print("[Ingestion] ❌ EVAL_LLM_ID not set in .env.evaluation") + return False + + from app.db import get_db + + db = next(get_db()) + try: + llm_client = get_llm_client(llm_id, db) + finally: + db.close() except Exception as e: - print(f"[Ingestion] LLM client unavailable, will skip LLM-dependent steps: {e}") - llm_available = False + print(f"[Ingestion] LLM client unavailable: {e}") + return False - # Step A: Build DialogData list from contexts with robust parsing + # Step 2: Parse contexts and create DialogData with chunks + print(f"[Ingestion] Parsing {len(contexts)} contexts...") chunker = DialogueChunker(chunker_strategy) dialog_data_list: List[DialogData] = [] @@ -94,7 +130,7 @@ async def ingest_contexts_via_full_pipeline( line = raw.strip() if not line: continue - m = re.match(r'^\s*([^:：]+)\s*[：:]\s*(.+)$', line) + m = re.match(r'^\s*([^:：]+)\s*[：:]\s*(.+)', line) if m: role = m.group(1).strip() msg = m.group(2).strip() @@ -118,10 +154,12 @@ async def ingest_contexts_via_full_pipeline( dialog_data_list.append(dialog) if not dialog_data_list: - print("No dialogs to process for ingestion.") + print("[Ingestion] No dialogs to process.") return False - # Optionally save chunking outputs for debugging + print(f"[Ingestion] Parsed {len(dialog_data_list)} dialogs with chunks") + + # Step 3: Optionally save chunking outputs for debugging if save_chunk_output: try: def _serialize_datetime(obj): @@ -137,124 +175,185 @@ async def ingest_contexts_via_full_pipeline( combined_output = [dd.model_dump() for dd in dialog_data_list] with open(out_path, "w", encoding="utf-8") as f: json.dump(combined_output, f, ensure_ascii=False, indent=4, default=_serialize_datetime) - print(f"Saved chunking results to: {out_path}") + print(f"[Ingestion] Saved chunking results to: {out_path}") except Exception as e: - print(f"Failed to save chunking results: {e}") + print(f"[Ingestion] Failed to save chunking results: {e}") - # Step B-G: 使用新的 ExtractionOrchestrator 执行完整的提取流水线 - if not llm_available: - print("[Ingestion] Skipping extraction pipeline (no LLM).") - return False - - # 初始化 embedder 客户端 - from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient + # Step 4: Initialize embedder client from app.core.models.base import RedBearModelConfig - from app.services.memory_config_service import MemoryConfigService + from app.core.memory.utils.config.config_utils import get_embedder_config + from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient + from app.db import get_db try: - with get_db_context() as db: - embedder_config_dict = MemoryConfigService(db).get_embedder_config(embedding_name or SELECTED_EMBEDDING_ID) - embedder_config = RedBearModelConfig(**embedder_config_dict) - embedder_client = OpenAIEmbedderClient(embedder_config) + db = next(get_db()) + try: + embedder_config_dict = get_embedder_config(embedding_name, db) + embedder_config = RedBearModelConfig(**embedder_config_dict) + embedder_client = OpenAIEmbedderClient(embedder_config) + finally: + db.close() except Exception as e: print(f"[Ingestion] Failed to initialize embedder client: {e}") - print("[Ingestion] Skipping extraction pipeline (embedder initialization failed).") return False + # Step 5: Initialize Neo4j connector connector = Neo4jConnector() - # 初始化并运行 ExtractionOrchestrator - from app.core.memory.utils.config.config_utils import get_pipeline_config - config = get_pipeline_config() + # Step 6: 构建 MemoryConfig（从环境变量直接构建，不依赖数据库） + print("[Ingestion] 构建 MemoryConfig from environment variables...") + from app.schemas.memory_config_schema import MemoryConfig + + try: + # 从环境变量获取配置参数 + llm_id = os.getenv("EVAL_LLM_ID") + embedding_id = os.getenv("EVAL_EMBEDDING_ID") + chunker_strategy_env = os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker") + + if not llm_id or not embedding_id: + print("[Ingestion] ❌ EVAL_LLM_ID or EVAL_EMBEDDING_ID is not set in .env.evaluation") + print("[Ingestion] Please set both EVAL_LLM_ID and EVAL_EMBEDDING_ID") + await connector.close() + return False + + # 从数据库获取模型信息（仅用于显示名称） + from app.db import get_db + db = next(get_db()) + try: + from sqlalchemy import text + # 获取 LLM 模型信息（从 model_configs 表） + llm_result = db.execute( + text("SELECT name FROM model_configs WHERE id = :id"), + {"id": llm_id} + ).fetchone() + llm_model_name = llm_result[0] if llm_result else "Unknown LLM" + + # 获取 Embedding 模型信息（从 model_configs 表） + emb_result = db.execute( + text("SELECT name FROM model_configs WHERE id = :id"), + {"id": embedding_id} + ).fetchone() + embedding_model_name = emb_result[0] if emb_result else "Unknown Embedding" + except Exception as e: + # 如果查询失败，使用默认名称 + print(f"[Ingestion] Warning: Failed to query model names from database: {e}") + llm_model_name = f"LLM ({llm_id[:8]}...)" + embedding_model_name = f"Embedding ({embedding_id[:8]}...)" + finally: + db.close() + + # 构建 MemoryConfig 对象（使用最小必需配置） + from uuid import uuid4 + memory_config = MemoryConfig( + config_id=0, # 评估环境不需要真实的 config_id + config_name="evaluation_config", + workspace_id=uuid4(), # 临时 workspace_id + workspace_name="evaluation_workspace", + tenant_id=uuid4(), # 临时 tenant_id + llm_model_id=UUID(llm_id), + llm_model_name=llm_model_name, + embedding_model_id=UUID(embedding_id), + embedding_model_name=embedding_model_name, + storage_type="neo4j", + chunker_strategy=chunker_strategy_env, + reflexion_enabled=False, + reflexion_iteration_period=3, + reflexion_range="partial", + reflexion_baseline="TIME", + loaded_at=datetime.now(), + # 可选字段使用默认值 + rerank_model_id=None, + rerank_model_name=None, + llm_params={}, + embedding_params={}, + config_version="2.0", + ) + + print(f"[Ingestion] ✅ 构建 MemoryConfig 成功") + print(f"[Ingestion] LLM: {llm_model_name}") + print(f"[Ingestion] Embedding: {embedding_model_name}") + print(f"[Ingestion] Chunker: {chunker_strategy_env}") + + except Exception as e: + print(f"[Ingestion] ❌ Failed to build MemoryConfig: {e}") + print(f"[Ingestion] Please check:") + print(f"[Ingestion] 1. EVAL_LLM_ID and EVAL_EMBEDDING_ID are set in .env.evaluation") + print(f"[Ingestion] 2. Model IDs exist in the models table") + print(f"[Ingestion] 3. Database connection is working") + await connector.close() + return False + + # Step 7: Initialize and run ExtractionOrchestrator + print("[Ingestion] Running extraction pipeline with ExtractionOrchestrator...") + from app.services.memory_config_service import MemoryConfigService + config = MemoryConfigService.get_pipeline_config(memory_config) orchestrator = ExtractionOrchestrator( llm_client=llm_client, embedder_client=embedder_client, connector=connector, config=config, + embedding_id=str(memory_config.embedding_model_id), # 传递 embedding_id ) - # 创建一个包装的 orchestrator 来修复时间提取器的输出 - # 保存原始的 _assign_extracted_data 方法 - original_assign = orchestrator._assign_extracted_data - - def clean_temporal_value(value): - """清理 temporal_validity 字段的值，将无效值转换为 None""" - if value is None: - return None - if isinstance(value, str): - # 处理字符串形式的 'null', 'None', 空字符串等 - if value.lower() in ('null', 'none', '') or value.strip() == '': - return None - return value - - async def patched_assign_extracted_data(*args, **kwargs): - """包装方法：在赋值后清理 temporal_validity 中的无效字符串""" - result = await original_assign(*args, **kwargs) + try: + # Run the complete extraction pipeline + result = await orchestrator.run(dialog_data_list, is_pilot_run=False) - # 清理返回的 dialog_data_list 中的 temporal_validity - for dialog in result: - if hasattr(dialog, 'chunks') and dialog.chunks: - for chunk in dialog.chunks: - if hasattr(chunk, 'statements') and chunk.statements: - for statement in chunk.statements: - if hasattr(statement, 'temporal_validity') and statement.temporal_validity: - tv = statement.temporal_validity - # 清理 valid_at 和 invalid_at - if hasattr(tv, 'valid_at'): - tv.valid_at = clean_temporal_value(tv.valid_at) - if hasattr(tv, 'invalid_at'): - tv.invalid_at = clean_temporal_value(tv.invalid_at) - return result - - # 替换方法 - orchestrator._assign_extracted_data = patched_assign_extracted_data - - # 同时包装 _create_nodes_and_edges 方法，在创建节点前再次清理 - original_create = orchestrator._create_nodes_and_edges - - async def patched_create_nodes_and_edges(dialog_data_list_arg): - """包装方法：在创建节点前再次清理 temporal_validity""" - # 最后一次清理，确保万无一失 - for dialog in dialog_data_list_arg: - if hasattr(dialog, 'chunks') and dialog.chunks: - for chunk in dialog.chunks: - if hasattr(chunk, 'statements') and chunk.statements: - for statement in chunk.statements: - if hasattr(statement, 'temporal_validity') and statement.temporal_validity: - tv = statement.temporal_validity - if hasattr(tv, 'valid_at'): - tv.valid_at = clean_temporal_value(tv.valid_at) - if hasattr(tv, 'invalid_at'): - tv.invalid_at = clean_temporal_value(tv.invalid_at) + # Handle different return formats: + # - Pilot mode: 7 values (without dedup_details) + # - Normal mode: 8 values (with dedup_details at the end) + if len(result) == 8: + # Normal mode: includes dedup_details + ( + dialogue_nodes, + chunk_nodes, + statement_nodes, + entity_nodes, + statement_chunk_edges, + statement_entity_edges, + entity_entity_edges, + _, # dedup_details - not needed here + ) = result + elif len(result) == 7: + # Pilot mode or older version: no dedup_details + ( + dialogue_nodes, + chunk_nodes, + statement_nodes, + entity_nodes, + statement_chunk_edges, + statement_entity_edges, + entity_entity_edges, + ) = result + else: + raise ValueError(f"Unexpected number of return values: {len(result)}") - return await original_create(dialog_data_list_arg) - - orchestrator._create_nodes_and_edges = patched_create_nodes_and_edges - - # 运行完整的提取流水线 - # orchestrator.run 返回 7 个元素的元组 - result = await orchestrator.run(dialog_data_list, is_pilot_run=False) - ( - dialogue_nodes, - chunk_nodes, - statement_nodes, - entity_nodes, - statement_chunk_edges, - statement_entity_edges, - entity_entity_edges, - ) = result - - # statement_chunk_edges 已经由 orchestrator 创建，无需重复创建 + print(f"[Ingestion] Extraction completed: {len(statement_nodes)} statements, {len(entity_nodes)} entities") + + except ValueError as e: + # If unpacking fails, provide helpful error message + print(f"[Ingestion] Extraction pipeline result unpacking failed: {e}") + print(f"[Ingestion] Result type: {type(result)}, length: {len(result) if hasattr(result, '__len__') else 'N/A'}") + if hasattr(result, '__len__') and len(result) > 0: + print(f"[Ingestion] First element type: {type(result[0])}") + await connector.close() + return False + except Exception as e: + print(f"[Ingestion] Extraction pipeline failed: {e}") + import traceback + traceback.print_exc() + await connector.close() + return False - # Step G: 生成记忆摘要 + # Step 7: Generate memory summaries print("[Ingestion] Generating memory summaries...") try: from app.core.memory.storage_services.extraction_engine.knowledge_extraction.memory_summary import ( memory_summary_generation, ) - from app.repositories.neo4j.add_edges import add_memory_summary_statement_edges from app.repositories.neo4j.add_nodes import add_memory_summary_nodes + from app.repositories.neo4j.add_edges import add_memory_summary_statement_edges summaries = await memory_summary_generation( chunked_dialogs=dialog_data_list, @@ -266,7 +365,8 @@ async def ingest_contexts_via_full_pipeline( print(f"[Ingestion] Warning: Failed to generate memory summaries: {e}") summaries = [] - # Step H: Save to Neo4j + # Step 8: Save to Neo4j + print("[Ingestion] Saving to Neo4j...") try: success = await save_dialog_and_statements_to_neo4j( dialogue_nodes=dialogue_nodes, @@ -284,18 +384,21 @@ async def ingest_contexts_via_full_pipeline( try: await add_memory_summary_nodes(summaries, connector) await add_memory_summary_statement_edges(summaries, connector) - print(f"Successfully saved {len(summaries)} memory summary nodes to Neo4j") + print(f"[Ingestion] Saved {len(summaries)} memory summary nodes to Neo4j") except Exception as e: - print(f"Warning: Failed to save summary nodes: {e}") + print(f"[Ingestion] Warning: Failed to save summary nodes: {e}") await connector.close() + if success: - print("Successfully saved extracted data to Neo4j!") + print("[Ingestion] Successfully saved all data to Neo4j!") else: - print("Failed to save data to Neo4j") + print("[Ingestion] Failed to save data to Neo4j") return success + except Exception as e: - print(f"Failed to save data to Neo4j: {e}") + print(f"[Ingestion] Failed to save data to Neo4j: {e}") + await connector.close() return False diff --git a/api/app/core/memory/evaluation/locomo/locomo_benchmark.py b/api/app/core/memory/evaluation/locomo/locomo_benchmark.py index 1c70c28e..eed75016 100644 --- a/api/app/core/memory/evaluation/locomo/locomo_benchmark.py +++ b/api/app/core/memory/evaluation/locomo/locomo_benchmark.py @@ -15,134 +15,145 @@ import json import os import time from datetime import datetime -from typing import Any, Dict, List, Optional +from typing import List, Dict, Any, Optional +from pathlib import Path +from dotenv import load_dotenv -try: - from dotenv import load_dotenv -except ImportError: - def load_dotenv(): - pass +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") +from app.repositories.neo4j.neo4j_connector import Neo4jConnector +from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient +from app.core.models.base import RedBearModelConfig +from app.core.memory.utils.config.config_utils import get_embedder_config +from app.core.memory.utils.llm.llm_utils import get_llm_client from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - bleu1, f1_score, + bleu1, jaccard, latency_stats, + avg_context_tokens ) from app.core.memory.evaluation.locomo.locomo_metrics import ( - get_category_name, locomo_f1_score, locomo_multi_f1, + get_category_name ) from app.core.memory.evaluation.locomo.locomo_utils import ( - extract_conversations, - ingest_conversations_if_needed, load_locomo_data, + extract_conversations, resolve_temporal_references, - retrieve_relevant_information, select_and_format_information, -) -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.memory.utils.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_end_user_id, - SELECTED_LLM_ID, + retrieve_relevant_information, ) from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.core.models.base import RedBearModelConfig from app.db import get_db_context -from app.repositories.neo4j.neo4j_connector import Neo4jConnector from app.services.memory_config_service import MemoryConfigService +# Get configuration from environment variables +PROJECT_ROOT = str(Path(__file__).resolve().parents[5]) # api directory +SELECTED_EMBEDDING_ID = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") +SELECTED_end_user_id = os.getenv("LOCOMO_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "locomo_benchmark") +SELECTED_LLM_ID = os.getenv("EVAL_LLM_ID", "2c9b0782-7a85-4740-ba84-4baf77f256c4") -async def run_locomo_benchmark( - sample_size: int = 20, - end_user_id: Optional[str] = None, - search_type: str = "hybrid", - search_limit: int = 12, - context_char_budget: int = 8000, - reset_group: bool = False, - skip_ingest: bool = False, - output_dir: Optional[str] = None -) -> Dict[str, Any]: + +# ============================================================================ +# Step 1: Data Loading +# ============================================================================ + +def step_load_data(data_path: str, sample_size: int) -> List[Dict[str, Any]]: """ - Run LoCoMo benchmark evaluation. - - This function orchestrates the complete evaluation pipeline: - 1. Load LoCoMo dataset (only QA pairs from first conversation) - 2. Check/ingest conversations into database (only first conversation, unless skip_ingest=True) - 3. For each question: - - Retrieve relevant information - - Generate answer using LLM - - Calculate metrics - 4. Aggregate results and save to file - - Note: By default, only the first conversation is ingested into the database, - and only QA pairs from that conversation are evaluated. This ensures that - all questions have corresponding memory in the database for retrieval. + Load QA pairs from LoCoMo dataset. Args: - sample_size: Number of QA pairs to evaluate (from first conversation) - end_user_id: Database group ID for retrieval (uses default if None) - search_type: "keyword", "embedding", or "hybrid" - search_limit: Max documents to retrieve per query - context_char_budget: Max characters for context - reset_group: Whether to clear and re-ingest data (not implemented) - skip_ingest: If True, skip data ingestion and use existing data in Neo4j - output_dir: Directory to save results (uses default if None) + data_path: Path to locomo10.json file + sample_size: Number of QA pairs to load (0 for all) Returns: - Dictionary with evaluation results including metrics, timing, and samples + List of QA items from the first conversation """ - # Use default end_user_id if not provided - end_user_id = end_user_id or SELECTED_end_user_id + print("📂 Loading LoCoMo data...") - # Determine data path - data_path = os.path.join(PROJECT_ROOT, "data", "locomo10.json") - if not os.path.exists(data_path): - # Fallback to current directory - data_path = os.path.join(os.getcwd(), "data", "locomo10.json") + # Load the dataset + qa_items = load_locomo_data(data_path, sample_size) - print(f"\n{'='*60}") - print("🚀 Starting LoCoMo Benchmark Evaluation") - print(f"{'='*60}") - print("📊 Configuration:") - print(f" Sample size: {sample_size}") - print(f" Group ID: {end_user_id}") - print(f" Search type: {search_type}") - print(f" Search limit: {search_limit}") - print(f" Context budget: {context_char_budget} chars") - print(f" Data path: {data_path}") - print(f"{'='*60}\n") + print(f"✅ Loaded {len(qa_items)} QA pairs from first conversation\n") + return qa_items + + +# ============================================================================ +# Step 2: Data Ingestion +# ============================================================================ + +async def ingest_conversations_if_needed( + conversations: List[str], + end_user_id: str, + reset: bool = False +) -> bool: + """ + Ingest conversations into Neo4j database. - # Step 1: Load LoCoMo data - print("📂 Loading LoCoMo dataset...") + Args: + conversations: List of conversation strings (already formatted) + end_user_id: Database end_user ID + reset: Whether to reset the group before ingestion + + Returns: + True if successful, False otherwise + """ try: - # Only load QA pairs from the first conversation (index 0) - # since we only ingest the first conversation into the database - qa_items = load_locomo_data(data_path, sample_size, conversation_index=0) - print(f"✅ Loaded {len(qa_items)} QA pairs from conversation 0\n") + from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline + + # Conversations are already formatted as strings, use them directly + await ingest_contexts_via_full_pipeline(conversations, end_user_id) + return True + except Exception as e: - print(f"❌ Failed to load data: {e}") - return { - "error": f"Data loading failed: {e}", - "timestamp": datetime.now().isoformat() - } + print(f"⚠️ Ingestion error: {e}") + import traceback + traceback.print_exc() + return False + + +async def step_ingest_data( + data_path: str, + end_user_id: str, + skip_ingest: bool, + reset_group: bool, + max_messages: Optional[int] = None +) -> bool: + """ + Ingest conversations into Neo4j database if needed. - # Step 2: Extract conversations and ingest if needed + Args: + data_path: Path to locomo10.json file + end_user_id: Database end_user ID + skip_ingest: Whether to skip ingestion + reset_group: Whether to reset the group before ingestion + max_messages: Maximum messages per dialogue to ingest (for testing) + + Returns: + True if ingestion succeeded or was skipped, False otherwise + """ if skip_ingest: print("⏭️ Skipping data ingestion (using existing data in Neo4j)") - print(f" Group ID: {end_user_id}\n") + print(f" End User ID: {end_user_id}\n") else: print("💾 Checking database ingestion...") try: - conversations = extract_conversations(data_path, max_dialogues=1) + # Extract conversations with optional message limit + conversations = extract_conversations( + data_path, + max_dialogues=1, + max_messages_per_dialogue=max_messages + ) print(f"📝 Extracted {len(conversations)} conversations") # Always ingest for now (ingestion check not implemented) - print(f"🔄 Ingesting conversations into group '{end_user_id}'...") + print(f"🔄 Ingesting conversations into end_user '{end_user_id}'...") success = await ingest_conversations_if_needed( conversations=conversations, end_user_id=end_user_id, @@ -156,238 +167,249 @@ async def run_locomo_benchmark( except Exception as e: print(f"❌ Ingestion failed: {e}") + import traceback + traceback.print_exc() print("⚠️ Continuing with evaluation (database may be empty)\n") - # Step 3: Initialize clients + return True + + +# ============================================================================ +# Step 3: Initialize Clients +# ============================================================================ + +def step_initialize_clients(llm_id: str, embedding_id: str): + """ + Initialize Neo4j connector, LLM client, and embedder. + + Args: + llm_id: LLM model ID + embedding_id: Embedding model ID + + Returns: + Tuple of (connector, llm_client, embedder) + """ print("🔧 Initializing clients...") + connector = Neo4jConnector() - # Initialize LLM client with database context - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(SELECTED_LLM_ID) + # Get database session + from app.db import get_db + db = next(get_db()) + try: + llm_client = get_llm_client(llm_id, db) + cfg_dict = get_embedder_config(embedding_id, db) + embedder = OpenAIEmbedderClient( + model_config=RedBearModelConfig.model_validate(cfg_dict) + ) + finally: + db.close() - # Initialize embedder - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) print("✅ Clients initialized\n") - - # Step 4: Process questions + return connector, llm_client, embedder + + +# ============================================================================ +# Step 4: Process Questions +# ============================================================================ + +async def step_process_all_questions( + qa_items: List[Dict[str, Any]], + end_user_id: str, + search_type: str, + search_limit: int, + context_char_budget: int, + connector: Neo4jConnector, + embedder: OpenAIEmbedderClient, + llm_client: Any +) -> List[Dict[str, Any]]: + """Process all QA items: retrieve, generate, and calculate metrics.""" print(f"🔍 Processing {len(qa_items)} questions...") print(f"{'='*60}\n") - # Tracking variables - latencies_search: List[float] = [] - latencies_llm: List[float] = [] - context_counts: List[int] = [] - context_chars: List[int] = [] - context_tokens: List[int] = [] - - # Metric lists - f1_scores: List[float] = [] - bleu1_scores: List[float] = [] - jaccard_scores: List[float] = [] - locomo_f1_scores: List[float] = [] - - # Per-category tracking - category_counts: Dict[str, int] = {} - category_f1: Dict[str, List[float]] = {} - category_bleu1: Dict[str, List[float]] = {} - category_jaccard: Dict[str, List[float]] = {} - category_locomo_f1: Dict[str, List[float]] = {} - - # Detailed samples samples: List[Dict[str, Any]] = [] - - # Fixed anchor date for temporal resolution anchor_date = datetime(2023, 5, 8) - try: - for idx, item in enumerate(qa_items, 1): - question = item.get("question", "") - ground_truth = item.get("answer", "") - category = get_category_name(item) - - # Ensure ground truth is a string - ground_truth_str = str(ground_truth) if ground_truth is not None else "" - - print(f"[{idx}/{len(qa_items)}] Category: {category}") - print(f"❓ Question: {question}") - print(f"✅ Ground Truth: {ground_truth_str}") - - # Step 4a: Retrieve relevant information - t_search_start = time.time() - try: - retrieved_info = await retrieve_relevant_information( - question=question, - end_user_id=end_user_id, - search_type=search_type, - search_limit=search_limit, - connector=connector, - embedder=embedder - ) - t_search_end = time.time() - search_latency = (t_search_end - t_search_start) * 1000 - latencies_search.append(search_latency) - - print(f"🔍 Retrieved {len(retrieved_info)} documents ({search_latency:.1f}ms)") - - except Exception as e: - print(f"❌ Retrieval failed: {e}") - retrieved_info = [] - search_latency = 0.0 - latencies_search.append(search_latency) - - # Step 4b: Select and format context - context_text = select_and_format_information( - retrieved_info=retrieved_info, + for idx, item in enumerate(qa_items, 1): + question = item.get("question", "") + ground_truth = item.get("answer", "") + category = get_category_name(item) + ground_truth_str = str(ground_truth) if ground_truth is not None else "" + + print(f"[{idx}/{len(qa_items)}] Category: {category}") + print(f"❓ Question: {question}") + print(f"✅ Ground Truth: {ground_truth_str}") + + # Retrieve + t_search_start = time.time() + try: + retrieved_info = await retrieve_relevant_information( question=question, - max_chars=context_char_budget + end_user_id=end_user_id, + search_type=search_type, + search_limit=search_limit, + connector=connector, + embedder=embedder ) - - # Resolve temporal references - context_text = resolve_temporal_references(context_text, anchor_date) - - # Add reference date to context - if context_text: - context_text = f"Reference date: {anchor_date.date().isoformat()}\n\n{context_text}" + search_latency = (time.time() - t_search_start) * 1000 + print(f"🔍 Retrieved {len(retrieved_info)} documents ({search_latency:.1f}ms)") + except Exception as e: + print(f"❌ Retrieval failed: {e}") + retrieved_info = [] + search_latency = 0.0 + + # Format context + context_text = select_and_format_information( + retrieved_info=retrieved_info, + question=question, + max_chars=context_char_budget + ) + context_text = resolve_temporal_references(context_text, anchor_date) + if context_text: + context_text = f"Reference date: {anchor_date.date().isoformat()}\n\n{context_text}" + else: + context_text = "No relevant context found." + + print(f"📝 Context: {len(context_text)} chars, {len(retrieved_info)} docs") + + # Generate answer + messages = [ + { + "role": "system", + "content": ( + "You are a precise QA assistant. Answer following these rules:\n" + "1) Extract the EXACT information mentioned in the context\n" + "2) For time questions: calculate actual dates from relative times\n" + "3) Return ONLY the answer text in simplest form\n" + "4) For dates, use format 'DD Month YYYY' (e.g., '7 May 2023')\n" + "5) If no clear answer found, respond with 'Unknown'" + ) + }, + { + "role": "user", + "content": f"Question: {question}\n\nContext:\n{context_text}" + } + ] + + t_llm_start = time.time() + try: + response = await llm_client.chat(messages=messages) + llm_latency = (time.time() - t_llm_start) * 1000 + if hasattr(response, 'content'): + prediction = response.content.strip() + elif isinstance(response, dict): + prediction = response["choices"][0]["message"]["content"].strip() else: - context_text = "No relevant context found." - - # Track context statistics - context_counts.append(len(retrieved_info)) - context_chars.append(len(context_text)) - context_tokens.append(len(context_text.split())) - - print(f"📝 Context: {len(context_text)} chars, {len(retrieved_info)} docs") - - # Step 4c: Generate answer with LLM - messages = [ - { - "role": "system", - "content": ( - "You are a precise QA assistant. Answer following these rules:\n" - "1) Extract the EXACT information mentioned in the context\n" - "2) For time questions: calculate actual dates from relative times\n" - "3) Return ONLY the answer text in simplest form\n" - "4) For dates, use format 'DD Month YYYY' (e.g., '7 May 2023')\n" - "5) If no clear answer found, respond with 'Unknown'" - ) - }, - { - "role": "user", - "content": f"Question: {question}\n\nContext:\n{context_text}" - } - ] - - t_llm_start = time.time() - try: - response = await llm_client.chat(messages=messages) - t_llm_end = time.time() - llm_latency = (t_llm_end - t_llm_start) * 1000 - latencies_llm.append(llm_latency) - - # Extract prediction from response - if hasattr(response, 'content'): - prediction = response.content.strip() - elif isinstance(response, dict): - prediction = response["choices"][0]["message"]["content"].strip() - else: - prediction = "Unknown" - - print(f"🤖 Prediction: {prediction} ({llm_latency:.1f}ms)") - - except Exception as e: - print(f"❌ LLM failed: {e}") prediction = "Unknown" - llm_latency = 0.0 - latencies_llm.append(llm_latency) - - # Step 4d: Calculate metrics - f1_val = f1_score(prediction, ground_truth_str) - bleu1_val = bleu1(prediction, ground_truth_str) - jaccard_val = jaccard(prediction, ground_truth_str) - - # LoCoMo-specific F1: use multi-answer for category 1 (Multi-Hop) - if item.get("category") == 1: - locomo_f1_val = locomo_multi_f1(prediction, ground_truth_str) - else: - locomo_f1_val = locomo_f1_score(prediction, ground_truth_str) - - # Accumulate metrics - f1_scores.append(f1_val) - bleu1_scores.append(bleu1_val) - jaccard_scores.append(jaccard_val) - locomo_f1_scores.append(locomo_f1_val) - - # Track by category - category_counts[category] = category_counts.get(category, 0) + 1 - category_f1.setdefault(category, []).append(f1_val) - category_bleu1.setdefault(category, []).append(bleu1_val) - category_jaccard.setdefault(category, []).append(jaccard_val) - category_locomo_f1.setdefault(category, []).append(locomo_f1_val) - - print(f"📊 Metrics - F1: {f1_val:.3f}, BLEU-1: {bleu1_val:.3f}, " - f"Jaccard: {jaccard_val:.3f}, LoCoMo F1: {locomo_f1_val:.3f}") - print() - - # Save sample details - samples.append({ - "question": question, - "ground_truth": ground_truth_str, - "prediction": prediction, - "category": category, - "metrics": { - "f1": f1_val, - "bleu1": bleu1_val, - "jaccard": jaccard_val, - "locomo_f1": locomo_f1_val - }, - "retrieval": { - "num_docs": len(retrieved_info), - "context_length": len(context_text) - }, - "timing": { - "search_ms": search_latency, - "llm_ms": llm_latency - } - }) + print(f"🤖 Prediction: {prediction} ({llm_latency:.1f}ms)") + except Exception as e: + print(f"❌ LLM failed: {e}") + prediction = "Unknown" + llm_latency = 0.0 + + # Calculate metrics + f1_val = f1_score(prediction, ground_truth_str) + bleu1_val = bleu1(prediction, ground_truth_str) + jaccard_val = jaccard(prediction, ground_truth_str) + if item.get("category") == 1: + locomo_f1_val = locomo_multi_f1(prediction, ground_truth_str) + else: + locomo_f1_val = locomo_f1_score(prediction, ground_truth_str) + + print(f"📊 Metrics - F1: {f1_val:.3f}, BLEU-1: {bleu1_val:.3f}, " + f"Jaccard: {jaccard_val:.3f}, LoCoMo F1: {locomo_f1_val:.3f}") + print() + + samples.append({ + "question": question, + "ground_truth": ground_truth_str, + "prediction": prediction, + "category": category, + "metrics": { + "f1": f1_val, + "bleu1": bleu1_val, + "jaccard": jaccard_val, + "locomo_f1": locomo_f1_val + }, + "retrieval": { + "num_docs": len(retrieved_info), + "context_length": len(context_text) + }, + "context_tokens": len(context_text.split()), + "timing": { + "search_ms": search_latency, + "llm_ms": llm_latency + } + }) - finally: - # Close connector - await connector.close() - - # Step 5: Aggregate results + return samples + + +# ============================================================================ +# Step 5: Aggregate Results +# ============================================================================ + +def step_aggregate_results(samples: List[Dict[str, Any]]) -> Dict[str, Any]: + """Aggregate metrics from all samples.""" print(f"\n{'='*60}") print("📊 Aggregating Results") print(f"{'='*60}\n") + if not samples: + return { + "overall_metrics": {}, + "by_category": {}, + "latency": {}, + "context_stats": {} + } + + # Extract metrics + f1_scores = [s["metrics"]["f1"] for s in samples] + bleu1_scores = [s["metrics"]["bleu1"] for s in samples] + jaccard_scores = [s["metrics"]["jaccard"] for s in samples] + locomo_f1_scores = [s["metrics"]["locomo_f1"] for s in samples] + + # Extract timing + latencies_search = [s["timing"]["search_ms"] for s in samples] + latencies_llm = [s["timing"]["llm_ms"] for s in samples] + + # Extract context stats + context_counts = [s["retrieval"]["num_docs"] for s in samples] + context_chars = [s["retrieval"]["context_length"] for s in samples] + context_tokens = [s["context_tokens"] for s in samples] + # Overall metrics overall_metrics = { - "f1": sum(f1_scores) / max(len(f1_scores), 1) if f1_scores else 0.0, - "bleu1": sum(bleu1_scores) / max(len(bleu1_scores), 1) if bleu1_scores else 0.0, - "jaccard": sum(jaccard_scores) / max(len(jaccard_scores), 1) if jaccard_scores else 0.0, - "locomo_f1": sum(locomo_f1_scores) / max(len(locomo_f1_scores), 1) if locomo_f1_scores else 0.0 + "f1": sum(f1_scores) / len(f1_scores) if f1_scores else 0.0, + "bleu1": sum(bleu1_scores) / len(bleu1_scores) if bleu1_scores else 0.0, + "jaccard": sum(jaccard_scores) / len(jaccard_scores) if jaccard_scores else 0.0, + "locomo_f1": sum(locomo_f1_scores) / len(locomo_f1_scores) if locomo_f1_scores else 0.0 } # Per-category metrics + category_data: Dict[str, Dict[str, List[float]]] = {} + for sample in samples: + cat = sample["category"] + if cat not in category_data: + category_data[cat] = { + "f1": [], + "bleu1": [], + "jaccard": [], + "locomo_f1": [] + } + category_data[cat]["f1"].append(sample["metrics"]["f1"]) + category_data[cat]["bleu1"].append(sample["metrics"]["bleu1"]) + category_data[cat]["jaccard"].append(sample["metrics"]["jaccard"]) + category_data[cat]["locomo_f1"].append(sample["metrics"]["locomo_f1"]) + by_category: Dict[str, Dict[str, Any]] = {} - for cat in category_counts: - f1_list = category_f1.get(cat, []) - b1_list = category_bleu1.get(cat, []) - j_list = category_jaccard.get(cat, []) - lf_list = category_locomo_f1.get(cat, []) - + for cat, metrics_lists in category_data.items(): by_category[cat] = { - "count": category_counts[cat], - "f1": sum(f1_list) / max(len(f1_list), 1) if f1_list else 0.0, - "bleu1": sum(b1_list) / max(len(b1_list), 1) if b1_list else 0.0, - "jaccard": sum(j_list) / max(len(j_list), 1) if j_list else 0.0, - "locomo_f1": sum(lf_list) / max(len(lf_list), 1) if lf_list else 0.0 + "count": len(metrics_lists["f1"]), + "f1": sum(metrics_lists["f1"]) / len(metrics_lists["f1"]), + "bleu1": sum(metrics_lists["bleu1"]) / len(metrics_lists["bleu1"]), + "jaccard": sum(metrics_lists["jaccard"]) / len(metrics_lists["jaccard"]), + "locomo_f1": sum(metrics_lists["locomo_f1"]) / len(metrics_lists["locomo_f1"]) } # Latency statistics @@ -398,12 +420,181 @@ async def run_locomo_benchmark( # Context statistics context_stats = { - "avg_retrieved_docs": sum(context_counts) / max(len(context_counts), 1) if context_counts else 0.0, - "avg_context_chars": sum(context_chars) / max(len(context_chars), 1) if context_chars else 0.0, - "avg_context_tokens": sum(context_tokens) / max(len(context_tokens), 1) if context_tokens else 0.0 + "avg_retrieved_docs": sum(context_counts) / len(context_counts) if context_counts else 0.0, + "avg_context_chars": sum(context_chars) / len(context_chars) if context_chars else 0.0, + "avg_context_tokens": sum(context_tokens) / len(context_tokens) if context_tokens else 0.0 } - # Build result dictionary + return { + "overall_metrics": overall_metrics, + "by_category": by_category, + "latency": latency, + "context_stats": context_stats + } + + +# ============================================================================ +# Step 6: Result Saving +# ============================================================================ + +def step_save_results( + result: Dict[str, Any], + output_dir: Optional[str] +) -> str: + """ + Save evaluation results to JSON file. + + Args: + result: Complete result dictionary + output_dir: Directory to save results (uses default if None) + + Returns: + Path to saved file + """ + if output_dir is None: + # Use absolute path to ensure results are saved in the correct location + script_dir = Path(__file__).resolve().parent + output_dir = script_dir / "results" + else: + # Convert to Path object + output_dir = Path(output_dir) + # If relative path, make it relative to script directory + if not output_dir.is_absolute(): + script_dir = Path(__file__).resolve().parent + output_dir = script_dir / output_dir + + # Create directory if it doesn't exist + output_dir.mkdir(parents=True, exist_ok=True) + + timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S") + output_path = output_dir / f"locomo_{timestamp_str}.json" + + try: + with open(output_path, "w", encoding="utf-8") as f: + json.dump(result, f, ensure_ascii=False, indent=2) + print(f"✅ Results saved to: {output_path}\n") + return str(output_path) + except Exception as e: + print(f"❌ Failed to save results: {e}") + print("📊 Printing results to console instead:\n") + print(json.dumps(result, ensure_ascii=False, indent=2)) + return "" + + +# ============================================================================ +# Main Orchestration Function +# ============================================================================ + + +async def run_locomo_benchmark( + sample_size: int = 20, + end_user_id: Optional[str] = None, + search_type: str = "hybrid", + search_limit: int = 12, + context_char_budget: int = 8000, + reset_group: bool = False, + skip_ingest: bool = False, + output_dir: Optional[str] = None, + max_ingest_messages: Optional[int] = None +) -> Dict[str, Any]: + """ + Run LoCoMo benchmark evaluation. + + This function orchestrates the complete evaluation pipeline by calling + well-defined step functions: + 1. Load LoCoMo dataset (only QA pairs from first conversation) + 2. Ingest conversations into database (unless skip_ingest=True) + 3. Initialize clients (Neo4j, LLM, Embedder) + 4. Process all questions (retrieve, generate, calculate metrics) + 5. Aggregate results + 6. Save results to file + + Note: By default, only the first conversation is ingested into the database, + and only QA pairs from that conversation are evaluated. This ensures that + all questions have corresponding memory in the database for retrieval. + + Args: + sample_size: Number of QA pairs to evaluate (from first conversation) + end_user_id: Database end_user ID for retrieval (uses default if None) + search_type: "keyword", "embedding", or "hybrid" + search_limit: Max documents to retrieve per query + context_char_budget: Max characters for context + reset_group: Whether to clear and re-ingest data + skip_ingest: If True, skip data ingestion and use existing data in Neo4j + output_dir: Directory to save results (uses default if None) + max_ingest_messages: Max messages per dialogue to ingest (for testing, None = all) + + Returns: + Dictionary with evaluation results including metrics, timing, and samples + """ + # Use default end_user_id if not provided + # 优先级：命令行参数 > LOCOMO_END_USER_ID > EVAL_END_USER_ID > 默认值 + if end_user_id is None: + end_user_id = os.getenv("LOCOMO_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "locomo_benchmark") + + # Get model IDs from config + llm_id = os.getenv("EVAL_LLM_ID", "6dc52e1b-9cec-4194-af66-a74c6307fc3f") + embedding_id = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") + + # Determine data path + dataset_dir = Path(__file__).resolve().parent.parent / "dataset" + data_path = dataset_dir / "locomo10.json" + if not os.path.exists(data_path): + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 locomo10.json 放置在: {dataset_dir}" + ) + + # Print configuration + print(f"\n{'='*60}") + print("🚀 Starting LoCoMo Benchmark Evaluation") + print(f"{'='*60}") + print("📊 Configuration:") + print(f" Sample size: {sample_size}") + print(f" End User ID: {end_user_id}") + print(f" Search type: {search_type}") + print(f" Search limit: {search_limit}") + print(f" Context budget: {context_char_budget} chars") + print(f" Data path: {data_path}") + if max_ingest_messages: + print(f" Max ingest messages: {max_ingest_messages} (testing mode)") + print(f"{'='*60}\n") + + # Step 1: Load LoCoMo data （加载数据） + try: + qa_items = step_load_data(data_path, sample_size) + except Exception as e: + print(f"❌ Failed to load data: {e}") + return { + "error": f"Data loading failed: {e}", + "timestamp": datetime.now().isoformat() + } + + # Step 2: Ingest data if needed（数据摄入） + await step_ingest_data(data_path, end_user_id, skip_ingest, reset_group, max_ingest_messages) + + # Step 3: Initialize clients （初始化客户端） + connector, llm_client, embedder = step_initialize_clients(llm_id, embedding_id) + + # Step 4: Process all questions （处理所有问题） + try: + samples = await step_process_all_questions( + qa_items=qa_items, + end_user_id=end_user_id, + search_type=search_type, + search_limit=search_limit, + context_char_budget=context_char_budget, + connector=connector, + embedder=embedder, + llm_client=llm_client + ) + finally: + await connector.close() + + # Step 5: Aggregate results （聚合答案） + aggregated = step_aggregate_results(samples) + + # Build final result dictionary result = { "dataset": "locomo", "sample_size": len(qa_items), @@ -413,37 +604,18 @@ async def run_locomo_benchmark( "search_type": search_type, "search_limit": search_limit, "context_char_budget": context_char_budget, - "llm_id": SELECTED_LLM_ID, - "embedding_id": SELECTED_EMBEDDING_ID + "llm_id": llm_id, + "embedding_id": embedding_id }, - "overall_metrics": overall_metrics, - "by_category": by_category, - "latency": latency, - "context_stats": context_stats, + "overall_metrics": aggregated["overall_metrics"], + "by_category": aggregated["by_category"], + "latency": aggregated["latency"], + "context_stats": aggregated["context_stats"], "samples": samples } - # Step 6: Save results - if output_dir is None: - output_dir = os.path.join( - os.path.dirname(__file__), - "results" - ) - - os.makedirs(output_dir, exist_ok=True) - - # Generate timestamped filename - timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S") - output_path = os.path.join(output_dir, f"locomo_{timestamp_str}.json") - - try: - with open(output_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"✅ Results saved to: {output_path}\n") - except Exception as e: - print(f"❌ Failed to save results: {e}") - print("📊 Printing results to console instead:\n") - print(json.dumps(result, ensure_ascii=False, indent=2)) + # Step 6: Save results （保存结果） + step_save_results(result, output_dir) return result @@ -454,7 +626,25 @@ def main(): This function provides a CLI interface for running LoCoMo benchmarks with configurable parameters. + + Configuration priority: Command-line args > Environment variables > Code defaults """ + # Load environment variables first + load_dotenv() + + # Get defaults from environment variables + env_sample_size = os.getenv("LOCOMO_SAMPLE_SIZE") + env_search_limit = os.getenv("LOCOMO_SEARCH_LIMIT") + env_context_budget = os.getenv("LOCOMO_CONTEXT_CHAR_BUDGET") + env_output_dir = os.getenv("LOCOMO_OUTPUT_DIR") + env_skip_ingest = os.getenv("LOCOMO_SKIP_INGEST", "false").lower() in ("true", "1", "yes") + + # Convert to appropriate types with fallback to code defaults + default_sample_size = int(env_sample_size) if env_sample_size else 20 + default_search_limit = int(env_search_limit) if env_search_limit else 12 + default_context_budget = int(env_context_budget) if env_context_budget else 8000 + default_output_dir = env_output_dir if env_output_dir else None + parser = argparse.ArgumentParser( description="Run LoCoMo benchmark evaluation", formatter_class=argparse.ArgumentDefaultsHelpFormatter @@ -463,14 +653,14 @@ def main(): parser.add_argument( "--sample_size", type=int, - default=20, - help="Number of QA pairs to evaluate" + default=default_sample_size, + help=f"Number of QA pairs to evaluate (env: LOCOMO_SAMPLE_SIZE={env_sample_size or 'not set'}, 0 for all)" ) parser.add_argument( "--end_user_id", type=str, default=None, - help="Database group ID for retrieval (uses default if not specified)" + help="Database end user ID for retrieval (uses LOCOMO_END_USER_ID or EVAL_END_USER_ID if not specified)" ) parser.add_argument( "--search_type", @@ -482,14 +672,14 @@ def main(): parser.add_argument( "--search_limit", type=int, - default=12, - help="Maximum number of documents to retrieve per query" + default=default_search_limit, + help=f"Maximum number of documents to retrieve per query (env: LOCOMO_SEARCH_LIMIT={env_search_limit or 'not set'})" ) parser.add_argument( "--context_char_budget", type=int, - default=8000, - help="Maximum characters for context" + default=default_context_budget, + help=f"Maximum characters for context (env: LOCOMO_CONTEXT_CHAR_BUDGET={env_context_budget or 'not set'})" ) parser.add_argument( "--reset_group", @@ -499,20 +689,24 @@ def main(): parser.add_argument( "--skip_ingest", action="store_true", - help="Skip data ingestion and use existing data in Neo4j" + default=env_skip_ingest, + help=f"Skip data ingestion and use existing data in Neo4j (env: LOCOMO_SKIP_INGEST={os.getenv('LOCOMO_SKIP_INGEST', 'false')})" ) parser.add_argument( "--output_dir", type=str, + default=default_output_dir, + help=f"Directory to save results (env: LOCOMO_OUTPUT_DIR={env_output_dir or 'not set'})" + ) + parser.add_argument( + "--max_ingest_messages", + type=int, default=None, - help="Directory to save results (uses default if not specified)" + help="Maximum messages per dialogue to ingest (for testing, default: all messages)" ) args = parser.parse_args() - # Load environment variables - load_dotenv() - # Run benchmark result = asyncio.run(run_locomo_benchmark( sample_size=args.sample_size, @@ -522,7 +716,8 @@ def main(): context_char_budget=args.context_char_budget, reset_group=args.reset_group, skip_ingest=args.skip_ingest, - output_dir=args.output_dir + output_dir=args.output_dir, + max_ingest_messages=args.max_ingest_messages )) # Print summary diff --git a/api/app/core/memory/evaluation/locomo/locomo_test.py b/api/app/core/memory/evaluation/locomo/locomo_test.py index 01c45123..2cb0664c 100644 --- a/api/app/core/memory/evaluation/locomo/locomo_test.py +++ b/api/app/core/memory/evaluation/locomo/locomo_test.py @@ -1,30 +1,29 @@ # file name: check_neo4j_connection_fixed.py import asyncio -import json -import math import os -import re import sys +import json import time +import math +import re from datetime import datetime, timedelta -from typing import Any, Dict, List +from typing import List, Dict, Any from pathlib import Path - from dotenv import load_dotenv -# 1 -# 添加项目根目录到路径 -current_dir = Path(__file__).resolve().parent -project_root = str(current_dir.parent) -if project_root not in sys.path: - sys.path.insert(0, project_root) -# 关键：将 src 目录置于最前，确保从当前仓库加载模块 -src_dir = os.path.join(project_root, "src") -if src_dir not in sys.path: - sys.path.insert(0, src_dir) - +# Load main .env load_dotenv() +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") + +# Get group_id from config +group_id = os.getenv("EVAL_GROUP_ID", "locomo_test") +print(f"✅ 使用配置的 group_id: {group_id}") + # 首先定义 _loc_normalize 函数，因为其他函数依赖它 def _loc_normalize(text: str) -> str: text = str(text) if text is not None else "" @@ -37,7 +36,7 @@ def _loc_normalize(text: str) -> str: # 尝试从 metrics.py 导入基础指标 try: - from common.metrics import bleu1, f1_score, jaccard + from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard print("✅ 从 metrics.py 导入基础指标成功") except ImportError as e: print(f"❌ 从 metrics.py 导入失败: {e}") @@ -107,23 +106,8 @@ except ImportError as e: # 尝试从 qwen_search_eval.py 导入 LoCoMo 特定指标 try: - # 添加 evaluation 目录路径 - evaluation_dir = os.path.join(project_root, "evaluation") - if evaluation_dir not in sys.path: - sys.path.insert(0, evaluation_dir) - - # 尝试从不同位置导入 - try: - from locomo.qwen_search_eval import ( - _resolve_relative_times, - loc_f1_score, - loc_multi_f1, - ) - print("✅ 从 locomo.qwen_search_eval 导入 LoCoMo 特定指标成功") - except ImportError: - from qwen_search_eval import _resolve_relative_times, loc_f1_score, loc_multi_f1 - print("✅ 从 qwen_search_eval 导入 LoCoMo 特定指标成功") - + from app.core.memory.evaluation.locomo.qwen_search_eval import loc_f1_score, loc_multi_f1, _resolve_relative_times + print("✅ 从 qwen_search_eval 导入 LoCoMo 特定指标成功") except ImportError as e: print(f"❌ 从 qwen_search_eval.py 导入失败: {e}") # 回退到本地实现 LoCoMo 特定函数 @@ -429,31 +413,36 @@ def enhanced_context_selection(contexts: List[str], question: str, question_inde async def run_enhanced_evaluation(): """使用增强方法进行完整评估 - 解决中间性能衰减问题""" - try: - from dotenv import load_dotenv - except Exception: - def load_dotenv(): - return None - + from dotenv import load_dotenv + from uuid import UUID + from datetime import datetime + from dataclasses import dataclass + # 修正导入路径：使用 app.core.memory.src 前缀 - from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient - from app.core.memory.utils.config.definitions import ( - SELECTED_EMBEDDING_ID, - SELECTED_LLM_ID, - ) - from app.core.memory.utils.llm.llm_utils import MemoryClientFactory - from app.core.models.base import RedBearModelConfig - from app.db import get_db_context - from app.repositories.neo4j.graph_search import search_graph_by_embedding from app.repositories.neo4j.neo4j_connector import Neo4jConnector + from app.repositories.neo4j.graph_search import search_graph_by_embedding + from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient + from app.core.models.base import RedBearModelConfig + from app.core.memory.utils.llm.llm_utils import get_llm_client + from app.core.memory.utils.config.config_utils import get_embedder_config + from app.schemas.memory_config_schema import MemoryConfig from app.services.memory_config_service import MemoryConfigService + + # Get model IDs from config + llm_id = os.getenv("EVAL_LLM_ID", "6dc52e1b-9cec-4194-af66-a74c6307fc3f") + embedding_id = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") - # 加载数据 - # 获取项目根目录 - current_file = os.path.abspath(__file__) - evaluation_dir = os.path.dirname(os.path.dirname(current_file)) # evaluation目录 - memory_dir = os.path.dirname(evaluation_dir) # memory目录 - data_path = os.path.join(memory_dir, "data", "locomo10.json") + # 加载数据 - 使用统一的 dataset 目录 + data_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "dataset", "locomo10.json") + + if not os.path.exists(data_path): + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 locomo10.json 放置在: api/app/core/memory/evaluation/dataset/" + ) + + print(f"✅ 找到数据文件: {data_path}") + with open(data_path, "r", encoding="utf-8") as f: raw = json.load(f) @@ -463,64 +452,109 @@ async def run_enhanced_evaluation(): qa_items.extend(entry.get("qa", [])) else: qa_items.extend(raw.get("qa", [])) - - items = qa_items[:20] # 测试多少个问题 + + # 测试多少个问题 - 可通过环境变量设置 + sample_size = int(os.getenv("LOCOMO_SAMPLE_SIZE", "20")) + items = qa_items[:sample_size] + print(f"📊 将测试 {len(items)} 个问题（总共 {len(qa_items)} 个可用）") # 初始化增强监控器 monitor = EnhancedEvaluationMonitor(reset_interval=5, performance_threshold=0.6) - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm = factory.get_llm_client(SELECTED_LLM_ID) + # 获取数据库会话并初始化 LLM 客户端 + from app.db import get_db + db = next(get_db()) - # 初始化embedder - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - - # 初始化连接器 - connector = Neo4jConnector() - - # 初始化结果字典 - results = { - "questions": [], - "overall_metrics": {"f1": 0.0, "b1": 0.0, "j": 0.0, "loc_f1": 0.0}, - "category_metrics": {}, - "retrieval_stats": {"total_questions": len(items), "avg_context_length": 0, "avg_retrieved_docs": 0}, - "performance_trend": "stable", - "timestamp": datetime.now().isoformat(), - "enhanced_strategy": True - } - - total_f1 = 0.0 - total_bleu1 = 0.0 - total_jaccard = 0.0 - total_loc_f1 = 0.0 - total_context_length = 0 - total_retrieved_docs = 0 - category_stats = {} - try: - for i, item in enumerate(items): - monitor.question_count += 1 + llm = get_llm_client(llm_id, db) + + # 初始化embedder + cfg_dict = get_embedder_config(embedding_id, db) + embedder = OpenAIEmbedderClient( + model_config=RedBearModelConfig.model_validate(cfg_dict) + ) + + # 🔧 创建 MemoryConfig 对象用于搜索 + # 方案1：如果有配置ID，从数据库加载 + config_id = os.getenv("EVAL_CONFIG_ID") + if config_id: + print(f"📋 从数据库加载配置 ID: {config_id}") + memory_config_service = MemoryConfigService(db) + memory_config = memory_config_service.load_memory_config(config_id, service_name="locomo_test") + else: + # 方案2：创建临时配置对象用于测试 + print(f"📋 创建临时测试配置") + from uuid import UUID + from datetime import datetime + + # 将字符串 ID 转换为 UUID + try: + embedding_uuid = UUID(embedding_id) + llm_uuid = UUID(llm_id) + except ValueError as e: + raise ValueError(f"无效的 UUID 格式: {e}") + + memory_config = MemoryConfig( + config_id=1, # 临时 ID + config_name="locomo_test_config", + workspace_id=UUID("00000000-0000-0000-0000-000000000000"), # 临时 workspace + workspace_name="test_workspace", + tenant_id=UUID("00000000-0000-0000-0000-000000000000"), # 临时 tenant + embedding_model_id=embedding_uuid, + embedding_model_name="test_embedding", + llm_model_id=llm_uuid, + llm_model_name="test_llm", + storage_type="neo4j", + chunker_strategy="RecursiveChunker", + reflexion_enabled=False, + reflexion_iteration_period=3, + reflexion_range="partial", + reflexion_baseline="Time", + loaded_at=datetime.now() + ) + + print(f"✅ MemoryConfig 已准备: embedding_id={memory_config.embedding_model_id}, llm_id={memory_config.llm_model_id}") + + # 初始化连接器 + connector = Neo4jConnector() - # 获取近期性能用于重置判断 - recent_performance = monitor.get_recent_performance() + # 初始化结果字典 + results = { + "questions": [], + "overall_metrics": {"f1": 0.0, "b1": 0.0, "j": 0.0, "loc_f1": 0.0}, + "category_metrics": {}, + "retrieval_stats": {"total_questions": len(items), "avg_context_length": 0, "avg_retrieved_docs": 0}, + "performance_trend": "stable", + "timestamp": datetime.now().isoformat(), + "enhanced_strategy": True + } - # 增强的重置判断 - should_reset = monitor.should_reset_connections(current_f1=recent_performance) - if should_reset and i > 0: - print(f"🔄 重置Neo4j连接 (问题 {i+1}/{len(items)}, 近期性能: {recent_performance:.3f})...") - await connector.close() - connector = Neo4jConnector() # 创建新连接 - print("✅ 连接重置完成") + total_f1 = 0.0 + total_bleu1 = 0.0 + total_jaccard = 0.0 + total_loc_f1 = 0.0 + total_context_length = 0 + total_retrieved_docs = 0 + category_stats = {} - q = item.get("question", "") - ref = item.get("answer", "") - ref_str = str(ref) if ref is not None else "" + try: + for i, item in enumerate(items): + monitor.question_count += 1 + + # 获取近期性能用于重置判断 + recent_performance = monitor.get_recent_performance() + + # 增强的重置判断 + should_reset = monitor.should_reset_connections(current_f1=recent_performance) + if should_reset and i > 0: + print(f"🔄 重置Neo4j连接 (问题 {i+1}/{len(items)}, 近期性能: {recent_performance:.3f})...") + await connector.close() + connector = Neo4jConnector() # 创建新连接 + print("✅ 连接重置完成") + + q = item.get("question", "") + ref = item.get("answer", "") + ref_str = str(ref) if ref is not None else "" print(f"\n🔍 [{i+1}/{len(items)}] 问题: {q}") print(f"✅ 真实答案: {ref_str}") @@ -548,10 +582,12 @@ async def run_enhanced_evaluation(): contexts_all = [] try: - # 使用统一的搜索服务 - from app.core.memory.storage_services.search import run_hybrid_search + # 使用旧版本的搜索服务（重构前的版本） + from app.core.memory.src.search import run_hybrid_search - print("🔀 使用混合搜索服务...") + print(f"🔀 使用混合搜索服务（旧版本）...") + print(f"📍 检索参数: group_id={group_id}, limit=20, search_type=hybrid") + print(f"📍 查询文本: {q}") search_results = await run_hybrid_search( query_text=q, @@ -559,15 +595,27 @@ async def run_enhanced_evaluation(): end_user_id="locomo_sk", limit=20, include=["statements", "chunks", "entities", "summaries"], - alpha=0.6, # BM25权重 - embedding_id=SELECTED_EMBEDDING_ID + output_path=None, + memory_config=memory_config, # 🔧 添加必需的 memory_config 参数 + rerank_alpha=0.6, # BM25权重 + use_forgetting_rerank=False, + use_llm_rerank=False ) - # 处理搜索结果 - 新的搜索服务返回统一的结构 - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) + # 处理搜索结果 - 旧版本返回包含 reranked_results 的结构 + # 对于 hybrid 搜索，使用 reranked_results + if "reranked_results" in search_results: + reranked = search_results["reranked_results"] + chunks = reranked.get("chunks", []) + statements = reranked.get("statements", []) + entities = reranked.get("entities", []) + summaries = reranked.get("summaries", []) + else: + # 单一搜索类型的结果 + chunks = search_results.get("chunks", []) + statements = search_results.get("statements", []) + entities = search_results.get("entities", []) + summaries = search_results.get("summaries", []) print(f"✅ 混合检索成功: {len(chunks)} chunks, {len(statements)} 条陈述, {len(entities)} 个实体, {len(summaries)} 个摘要") @@ -609,6 +657,8 @@ async def run_enhanced_evaluation(): print(f"📊 有效上下文数量: {len(contexts_all)}") except Exception as e: print(f"❌ 检索失败: {e}") + import traceback + print(f"详细错误信息:\n{traceback.format_exc()}") contexts_all = [] t1 = time.time() @@ -728,14 +778,17 @@ async def run_enhanced_evaluation(): print("="*60) - except Exception as e: - print(f"❌ 评估过程中发生错误: {e}") - # 即使出错，也返回已有的结果 - import traceback - traceback.print_exc() + except Exception as e: + print(f"❌ 评估过程中发生错误: {e}") + # 即使出错，也返回已有的结果 + import traceback + traceback.print_exc() + finally: + await connector.close() + finally: - await connector.close() + db.close() # 关闭数据库会话 # 计算总体指标 n = len(items) diff --git a/api/app/core/memory/evaluation/locomo/locomo_utils.py b/api/app/core/memory/evaluation/locomo/locomo_utils.py index d3b74947..6ad68470 100644 --- a/api/app/core/memory/evaluation/locomo/locomo_utils.py +++ b/api/app/core/memory/evaluation/locomo/locomo_utils.py @@ -15,8 +15,14 @@ import json import re from datetime import datetime, timedelta from typing import List, Dict, Any, Optional +from pathlib import Path +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) -from app.core.memory.utils.definitions import PROJECT_ROOT from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline @@ -82,7 +88,7 @@ def load_locomo_data( return qa_items[:sample_size] -def extract_conversations(data_path: str, max_dialogues: int = 1) -> List[str]: +def extract_conversations(data_path: str, max_dialogues: int = 1, max_messages_per_dialogue: Optional[int] = None) -> List[str]: """ Extract conversation texts from LoCoMo data for ingestion. @@ -93,6 +99,7 @@ def extract_conversations(data_path: str, max_dialogues: int = 1) -> List[str]: Args: data_path: Path to locomo10.json file max_dialogues: Maximum number of dialogues to extract (default: 1) + max_messages_per_dialogue: Maximum messages per dialogue (default: None = all messages) Returns: List of conversation strings formatted for ingestion. @@ -141,13 +148,21 @@ def extract_conversations(data_path: str, max_dialogues: int = 1) -> List[str]: continue lines.append(f"{role}: {text}") + + # Limit messages if specified + if max_messages_per_dialogue and len(lines) >= max_messages_per_dialogue: + break + + # Break outer loop if we've reached the message limit + if max_messages_per_dialogue and len(lines) >= max_messages_per_dialogue: + break if lines: contents.append("\n".join(lines)) return contents - +# 时间解析：将相对时间表达转换为绝对日期 def resolve_temporal_references(text: str, anchor_date: datetime) -> str: """ Resolve relative temporal references to absolute dates. @@ -225,6 +240,8 @@ def resolve_temporal_references(text: str, anchor_date: datetime) -> str: t, flags=re.IGNORECASE ) + + # 中文支持 t = re.sub( r"\bnext\s+week\b", (anchor_date + timedelta(days=7)).date().isoformat(), @@ -345,6 +362,50 @@ def select_and_format_information( return "\n\n".join(selected) +# 记忆系统核心能力：写入与读取 +async def ingest_conversations_if_needed( + conversations: List[str], + end_user_id: str, + reset: bool = False +) -> bool: + """ + Wrapper for conversation ingestion using external extraction pipeline. + + This function populates the Neo4j database with processed conversation data + (chunks, statements, entities) so that the retrieval system has memory to search. + + The ingestion process: + 1. Parses conversation text into dialogue messages + 2. Chunks the dialogues into semantic units + 3. Extracts statements and entities using LLM + 4. Generates embeddings for all content + 5. Stores everything in Neo4j graph database + + Args: + conversations: List of raw conversation texts from LoCoMo dataset + Example: ["User: I went to Paris. AI: When was that?", ...] + end_user_id: Target end_user ID for database storage + reset: Whether to clear existing data first (not implemented in wrapper) + + Returns: + True if successful, False otherwise + + Note: + The external function uses "contexts" to mean "conversation texts". + This runs the full extraction pipeline: chunking → entity extraction → + statement extraction → embedding → Neo4j storage. + """ + try: + success = await ingest_contexts_via_full_pipeline( + contexts=conversations, + end_user_id=end_user_id, + save_chunk_output=True, + reset_group=reset + ) + return success + except Exception as e: + print(f"[Ingestion] Failed to ingest conversations: {e}") + return False async def retrieve_relevant_information( question: str, @@ -385,7 +446,7 @@ async def retrieve_relevant_information( search_graph, search_graph_by_embedding ) - from app.core.memory.storage_services.search import run_hybrid_search + from app.core.memory.src.search import run_hybrid_search contexts_all: List[str] = [] diff --git a/api/app/core/memory/evaluation/locomo/qwen_search_eval.py b/api/app/core/memory/evaluation/locomo/qwen_search_eval.py index 6a5caa0c..889c5065 100644 --- a/api/app/core/memory/evaluation/locomo/qwen_search_eval.py +++ b/api/app/core/memory/evaluation/locomo/qwen_search_eval.py @@ -2,43 +2,29 @@ import argparse import asyncio import json import os -import statistics import time from datetime import datetime, timedelta -from typing import Any, Dict, List - -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None - +from typing import List, Dict, Any +import statistics import re +from pathlib import Path +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") -from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - bleu1, - jaccard, - latency_stats, -) -from app.core.memory.evaluation.common.metrics import f1_score as common_f1 -from app.core.memory.evaluation.extraction_utils import ( - ingest_contexts_via_full_pipeline, -) -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.memory.storage_services.search import run_hybrid_search -from app.core.memory.utils.config.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_GROUP_ID, - SELECTED_LLM_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.core.models.base import RedBearModelConfig -from app.db import get_db_context -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.services.memory_config_service import MemoryConfigService +from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding +from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient +from app.core.models.base import RedBearModelConfig +from app.core.memory.utils.config.config_utils import get_embedder_config +from app.core.memory.src.search import run_hybrid_search # 使用旧版本（重构前） +from app.core.memory.utils.llm.llm_utils import get_llm_client +from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline +from app.core.memory.evaluation.common.metrics import f1_score as common_f1, bleu1, jaccard, latency_stats, avg_context_tokens # 参考 evaluation/locomo/evaluation.py 的 F1 计算逻辑（移除外部依赖，内联实现） @@ -265,7 +251,10 @@ async def run_locomo_eval( end_user_id = end_user_id or SELECTED_end_user_id data_path = os.path.join(PROJECT_ROOT, "data", "locomo10.json") if not os.path.exists(data_path): - data_path = os.path.join(os.getcwd(), "data", "locomo10.json") + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 locomo10.json 放置在: {dataset_dir}" + ) with open(data_path, "r", encoding="utf-8") as f: raw = json.load(f) # LoCoMo 数据结构：顶层为若干对象，每个对象下有 qa 列表 @@ -343,13 +332,9 @@ async def run_locomo_eval( await ingest_contexts_via_full_pipeline(contents, end_user_id, save_chunk_output=True) # 使用异步LLM客户端 - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(SELECTED_LLM_ID) + llm_client = get_llm_client(llm_id) # 初始化embedder用于直接调用 - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) + cfg_dict = get_embedder_config(embedding_id) embedder = OpenAIEmbedderClient( model_config=RedBearModelConfig.model_validate(cfg_dict) ) @@ -480,8 +465,8 @@ async def run_locomo_eval( contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") else: # hybrid - # 🎯 关键修复：混合检索使用更严格的回退机制 - print("🔀 使用混合检索（带回退机制）...") + # 使用旧版本的混合检索（重构前） + print("🔀 使用混合检索（旧版本）...") try: search_results = await run_hybrid_search( query_text=q, @@ -490,16 +475,26 @@ async def run_locomo_eval( limit=adjusted_limit, include=["chunks", "statements", "entities", "summaries"], output_path=None, + rerank_alpha=0.6, + use_forgetting_rerank=False, + use_llm_rerank=False ) - # 🎯 关键修复：正确处理混合检索的扁平结构 - # 新的API返回扁平结构，直接从顶层获取结果 + # 处理旧版本的返回结构（包含 reranked_results） if search_results and isinstance(search_results, dict): - # 新API返回扁平结构：直接从顶层获取 - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) + # 对于 hybrid 搜索，使用 reranked_results + if "reranked_results" in search_results: + reranked = search_results["reranked_results"] + chunks = reranked.get("chunks", []) + statements = reranked.get("statements", []) + entities = reranked.get("entities", []) + summaries = reranked.get("summaries", []) + else: + # 单一搜索类型的结果 + chunks = search_results.get("chunks", []) + statements = search_results.get("statements", []) + entities = search_results.get("entities", []) + summaries = search_results.get("summaries", []) # 检查是否有有效结果 if chunks or statements or entities or summaries: @@ -799,8 +794,9 @@ async def run_locomo_eval( "search_limit": search_limit, "context_char_budget": context_char_budget, "search_type": search_type, - "llm_id": SELECTED_LLM_ID, - "retrieval_embedding_id": SELECTED_EMBEDDING_ID, + "llm_id": llm_id, + "retrieval_embedding_id": embedding_id, + "chunker_strategy": os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker"), "skip_ingest_if_exists": skip_ingest_if_exists, "llm_timeout": llm_timeout, "llm_max_retries": llm_max_retries, diff --git a/api/app/core/memory/evaluation/longmemeval/qwen_search_eval.py b/api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py similarity index 93% rename from api/app/core/memory/evaluation/longmemeval/qwen_search_eval.py rename to api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py index 8710a504..aaf46e35 100644 --- a/api/app/core/memory/evaluation/longmemeval/qwen_search_eval.py +++ b/api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py @@ -2,100 +2,67 @@ import argparse import asyncio import json import os +import time import re import statistics -import time from datetime import datetime, timedelta -from typing import Any, Dict, List - -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None - -# 确保可以找到 src 及项目根路径 -import sys +from typing import List, Dict, Any from pathlib import Path -_THIS_DIR = Path(__file__).resolve().parent -_PROJECT_ROOT = str(_THIS_DIR.parents[2]) -_SRC_DIR = os.path.join(_PROJECT_ROOT, "src") -for _p in (_SRC_DIR, _PROJECT_ROOT): - if _p not in sys.path: - sys.path.insert(0, _p) +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") -# 与现有评估脚本保持一致的导入方式 from app.repositories.neo4j.neo4j_connector import Neo4jConnector - -try: - # 优先从 extraction_utils1 导入 - from app.core.memory.evaluation.extraction_utils import ( - ingest_contexts_via_full_pipeline, # type: ignore - ) -except Exception: - ingest_contexts_via_full_pipeline = None # 在运行时做兜底检查 -from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - jaccard, - latency_stats, -) -from app.core.memory.evaluation.common.metrics import f1_score as common_f1 -from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.memory.utils.config.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_LLM_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.core.models.base import RedBearModelConfig -from app.db import get_db_context +from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding -from app.services.memory_config_service import MemoryConfigService - -try: - from app.core.memory.evaluation.common.metrics import exact_match -except Exception: - # 兜底：简单的大小写不敏感比较 - def exact_match(pred: str, ref: str) -> bool: - return str(pred).strip().lower() == str(ref).strip().lower() +from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient +from app.core.models.base import RedBearModelConfig +from app.core.memory.utils.config.config_utils import get_embedder_config +from app.core.memory.utils.llm.llm_utils import get_llm_client +from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME +from app.core.memory.evaluation.common.metrics import f1_score as common_f1, jaccard, latency_stats, avg_context_tokens +from app.core.memory.evaluation.common.metrics import exact_match def load_dataset_any(path: str) -> List[Dict[str, Any]]: - """健壮地加载数据集（兼容 list 或多段 JSON）。""" + """健壮地加载数据集，支持三种格式： + 1. 标准 JSON 数组: [{...}, {...}] + 2. 单个 JSON 对象: {...} + 3. JSONL 格式（每行一个 JSON）: {...}\n{...}\n{...} + """ with open(path, "r", encoding="utf-8") as f: - s = f.read().strip() + content = f.read().strip() + + # 尝试标准 JSON 解析 try: - obj = json.loads(s) - if isinstance(obj, list): - return obj - elif isinstance(obj, dict): - return [obj] + data = json.loads(content) + if isinstance(data, list): + return [item for item in data if isinstance(item, dict)] + elif isinstance(data, dict): + return [data] except json.JSONDecodeError: pass - dec = json.JSONDecoder() - idx = 0 - items: List[Dict[str, Any]] = [] - while idx < len(s): - while idx < len(s) and s[idx].isspace(): - idx += 1 - if idx >= len(s): - break + + # 尝试 JSONL 格式（每行一个 JSON 对象） + items = [] + for line in content.splitlines(): + line = line.strip() + if not line: + continue try: - obj, end = dec.raw_decode(s, idx) - if isinstance(obj, list): - for it in obj: - if isinstance(it, dict): - items.append(it) - elif isinstance(obj, dict): + obj = json.loads(line) + if isinstance(obj, dict): items.append(obj) - idx = end + elif isinstance(obj, list): + items.extend(item for item in obj if isinstance(item, dict)) except json.JSONDecodeError: - nl = s.find("\n", idx) - if nl == -1: - break - idx = nl + 1 + continue + return items @@ -624,7 +591,7 @@ def _resolve_relative_times_cn_en(text: str, anchor: datetime) -> str: async def run_longmemeval_test( sample_size: int = 3, - end_user_id: str = "longmemeval_zh_bak_3", + end_user_id: str | None = None, search_limit: int = 8, context_char_budget: int = 4000, llm_temperature: float = 0.0, @@ -639,18 +606,22 @@ async def run_longmemeval_test( skip_ingest: bool = False, ) -> Dict[str, Any]: """LongMemEval 评估测试：增强时间推理能力""" + + # Use environment variable with fallback chain + if end_user_id is None: + end_user_id = os.getenv("LONGMEMEVAL_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "longmemeval_zh_bak_3") # 数据路径 if not data_path: - # 固定使用中文数据集：data/longmemeval_oracle_zh.json - zh_proj = os.path.join(PROJECT_ROOT, "data", "longmemeval_oracle_zh.json") - zh_cwd = os.path.join(os.getcwd(), "data", "longmemeval_oracle_zh.json") - if os.path.exists(zh_proj): - data_path = zh_proj - elif os.path.exists(zh_cwd): - data_path = zh_cwd - else: - raise FileNotFoundError("未找到数据集: data/longmemeval_oracle_zh.json，请确保其存在于项目根目录或当前工作目录的 data 目录下。") + # 固定使用中文数据集：dataset/longmemeval_oracle_zh.json + dataset_dir = Path(__file__).resolve().parent.parent / "dataset" + data_path = str(dataset_dir / "longmemeval_oracle_zh.json") + + if not os.path.exists(data_path): + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 longmemeval_oracle_zh.json 放置在: {dataset_dir}" + ) qa_list: List[Dict[str, Any]] = load_dataset_any(data_path) # 支持评估全部样本：当 sample_size <= 0 时，取从 start_index 到末尾 @@ -702,16 +673,19 @@ async def run_longmemeval_test( ) # 初始化组件（摄入后再初始化连接器）- 使用异步LLM客户端 - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(SELECTED_LLM_ID) + from app.db import get_db + + db = next(get_db()) + try: + llm_client = get_llm_client(os.getenv("EVAL_LLM_ID"), db) + cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID"), db) + embedder = OpenAIEmbedderClient( + model_config=RedBearModelConfig.model_validate(cfg_dict) + ) + finally: + db.close() + connector = Neo4jConnector() - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) # 指标收集 latencies_llm: List[float] = [] @@ -768,10 +742,10 @@ async def run_longmemeval_test( if stmt_text: contexts_all.append(stmt_text) - # for sm in summaries: - # summary_text = str(sm.get("summary", "")).strip() - # if summary_text: - # contexts_all.append(summary_text) + for sm in summaries: + summary_text = str(sm.get("summary", "")).strip() + if summary_text: + contexts_all.append(summary_text) # 实体摘要（最多3个） scored = [e for e in entities if e.get("score") is not None] @@ -1228,8 +1202,8 @@ async def run_longmemeval_test( "search_limit": search_limit, "context_char_budget": context_char_budget, "search_type": search_type, - "llm_id": SELECTED_LLM_ID, - "embedding_id": SELECTED_EMBEDDING_ID, + "llm_id": os.getenv("EVAL_LLM_ID"), + "embedding_id": os.getenv("EVAL_EMBEDDING_ID"), "sample_size": sample_size, "start_index": start_index, }, @@ -1288,7 +1262,7 @@ def main(): parser.add_argument("--sample-size", type=int, default=3, help="样本数量（<=0 表示全部）") parser.add_argument("--all", action="store_true", help="评估全部样本（覆盖 --sample-size）") parser.add_argument("--start-index", type=int, default=0, help="起始样本索引") - parser.add_argument("--group-id", type=str, default="longmemeval_zh_bak_3", help="图数据库 Group ID") + parser.add_argument("--end-user-id", type=str, default=None, help="图数据库 End User ID，默认使用环境变量") parser.add_argument("--search-limit", type=int, default=8, help="检索条数上限") parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") @@ -1349,7 +1323,8 @@ def main(): # 保存结果到文件 try: - out_dir = os.path.join(PROJECT_ROOT, "evaluation", "longmemeval", "results") + # 使用相对路径而不是 PROJECT_ROOT + out_dir = Path(__file__).resolve().parent / "results" os.makedirs(out_dir, exist_ok=True) ts = datetime.now().strftime("%Y%m%d_%H%M%S") out_path = os.path.join(out_dir, f"longmemeval_{result['params']['search_type']}_{ts}.json") diff --git a/api/app/core/memory/evaluation/longmemeval/test_eval.py b/api/app/core/memory/evaluation/longmemeval/test_eval.py index 67bd6ec2..08daa890 100644 --- a/api/app/core/memory/evaluation/longmemeval/test_eval.py +++ b/api/app/core/memory/evaluation/longmemeval/test_eval.py @@ -2,81 +2,67 @@ import argparse import asyncio import json import os +import time import re import statistics -import time from datetime import datetime, timedelta -from typing import Any, Dict, List +from typing import List, Dict, Any +from pathlib import Path -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") # 与现有评估脚本保持一致的导入方式 -from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - jaccard, - latency_stats, -) -from app.core.memory.evaluation.common.metrics import f1_score as common_f1 -from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.memory.utils.config.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_LLM_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.core.models.base import RedBearModelConfig -from app.db import get_db_context -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.services.memory_config_service import MemoryConfigService - -try: - from app.core.memory.evaluation.common.metrics import exact_match -except Exception: - # 兜底：简单的大小写不敏感比较 - def exact_match(pred: str, ref: str) -> bool: - return str(pred).strip().lower() == str(ref).strip().lower() +from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding +from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient +from app.core.models.base import RedBearModelConfig +from app.core.memory.utils.config.config_utils import get_embedder_config +from app.core.memory.utils.llm.llm_utils import get_llm_client +from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME +from app.core.memory.evaluation.common.metrics import f1_score as common_f1, jaccard, latency_stats, avg_context_tokens +from app.core.memory.evaluation.common.metrics import exact_match def load_dataset_any(path: str) -> List[Dict[str, Any]]: - """健壮地加载数据集（兼容 list 或多段 JSON）。""" + """健壮地加载数据集，支持三种格式： + 1. 标准 JSON 数组: [{...}, {...}] + 2. 单个 JSON 对象: {...} + 3. JSONL 格式（每行一个 JSON）: {...}\n{...}\n{...} + """ with open(path, "r", encoding="utf-8") as f: - s = f.read().strip() + content = f.read().strip() + + # 尝试标准 JSON 解析 try: - obj = json.loads(s) - if isinstance(obj, list): - return obj - elif isinstance(obj, dict): - return [obj] + data = json.loads(content) + if isinstance(data, list): + return [item for item in data if isinstance(item, dict)] + elif isinstance(data, dict): + return [data] except json.JSONDecodeError: pass - dec = json.JSONDecoder() - idx = 0 - items: List[Dict[str, Any]] = [] - while idx < len(s): - while idx < len(s) and s[idx].isspace(): - idx += 1 - if idx >= len(s): - break + + # 尝试 JSONL 格式（每行一个 JSON 对象） + items = [] + for line in content.splitlines(): + line = line.strip() + if not line: + continue try: - obj, end = dec.raw_decode(s, idx) - if isinstance(obj, list): - for it in obj: - if isinstance(it, dict): - items.append(it) - elif isinstance(obj, dict): + obj = json.loads(line) + if isinstance(obj, dict): items.append(obj) - idx = end + elif isinstance(obj, list): + items.extend(item for item in obj if isinstance(item, dict)) except json.JSONDecodeError: - nl = s.find("\n", idx) - if nl == -1: - break - idx = nl + 1 + continue + return items @@ -640,15 +626,15 @@ async def run_longmemeval_test( # 数据路径 if not data_path: - # 固定使用中文数据集：data/longmemeval_oracle_zh.json - zh_proj = os.path.join(PROJECT_ROOT, "data", "longmemeval_oracle_zh.json") - zh_cwd = os.path.join(os.getcwd(), "data", "longmemeval_oracle_zh.json") - if os.path.exists(zh_proj): - data_path = zh_proj - elif os.path.exists(zh_cwd): - data_path = zh_cwd - else: - raise FileNotFoundError("未找到数据集: data/longmemeval_oracle_zh.json，请确保其存在于项目根目录或当前工作目录的 data 目录下。") + # 固定使用中文数据集：dataset/longmemeval_oracle_zh.json + dataset_dir = Path(__file__).resolve().parent.parent / "dataset" + data_path = str(dataset_dir / "longmemeval_oracle_zh.json") + + if not os.path.exists(data_path): + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 longmemeval_oracle_zh.json 放置在: {dataset_dir}" + ) qa_list: List[Dict[str, Any]] = load_dataset_any(data_path) # 支持评估全部样本：当 sample_size <= 0 时，取从 start_index 到末尾 @@ -658,13 +644,9 @@ async def run_longmemeval_test( items = qa_list[start_index:start_index + sample_size] # 初始化组件 - 使用异步LLM客户端 - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(SELECTED_LLM_ID) + llm_client = get_llm_client(os.getenv("EVAL_LLM_ID")) connector = Neo4jConnector() - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) + cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID")) embedder = OpenAIEmbedderClient( model_config=RedBearModelConfig.model_validate(cfg_dict) ) @@ -1203,8 +1185,8 @@ async def run_longmemeval_test( "search_limit": search_limit, "context_char_budget": context_char_budget, "search_type": search_type, - "llm_id": SELECTED_LLM_ID, - "embedding_id": SELECTED_EMBEDDING_ID, + "llm_id": os.getenv("EVAL_LLM_ID"), + "embedding_id": os.getenv("EVAL_EMBEDDING_ID"), "sample_size": sample_size, "start_index": start_index, }, diff --git a/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py b/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py index 3023020a..e07b0cab 100644 --- a/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py +++ b/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py @@ -2,81 +2,30 @@ import argparse import asyncio import json import os -import re import time from datetime import datetime -from typing import Any, Dict, List - -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None - -# 路径与模块导入保持与现有评估脚本一致 -import sys +from typing import List, Dict, Any +import re from pathlib import Path -_THIS_DIR = Path(__file__).resolve().parent -_PROJECT_ROOT = str(_THIS_DIR.parents[1]) -_SRC_DIR = os.path.join(_PROJECT_ROOT, "src") -for _p in (_SRC_DIR, _PROJECT_ROOT): - if _p not in sys.path: - sys.path.insert(0, _p) +from dotenv import load_dotenv + +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) + print(f"✅ 加载评估配置: {eval_config_path}") -# 对齐 locomo_test 的检索逻辑：直接使用 graph_search 与 Neo4jConnector/Embedder1 -from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - exact_match, - latency_stats, -) -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.memory.utils.config.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_GROUP_ID, - SELECTED_LLM_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.core.models.base import RedBearModelConfig -from app.db import get_db_context -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.services.memory_config_service import MemoryConfigService +from app.core.memory.src.search import run_hybrid_search # 使用与 evaluate_qa.py 相同的检索函数 +from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient +from app.core.models.base import RedBearModelConfig +from app.core.memory.utils.config.config_utils import get_embedder_config -try: - from app.core.memory.evaluation.common.metrics import bleu1, f1_score, jaccard -except Exception: - # 兜底：简单实现（必要时） - def f1_score(pred: str, ref: str) -> float: - ps = pred.lower().split() - rs = ref.lower().split() - if not ps or not rs: - return 0.0 - tp = len(set(ps) & set(rs)) - if tp == 0: - return 0.0 - precision = tp / len(ps) - recall = tp / len(rs) - if precision + recall == 0: - return 0.0 - return 2 * precision * recall / (precision + recall) +from app.core.memory.utils.llm.llm_utils import get_llm_client +from app.core.memory.evaluation.common.metrics import exact_match, latency_stats, avg_context_tokens - def bleu1(pred: str, ref: str) -> float: - ps = pred.lower().split() - rs = ref.lower().split() - if not ps or not rs: - return 0.0 - overlap = len([w for w in ps if w in rs]) - return overlap / max(len(ps), 1) - - def jaccard(pred: str, ref: str) -> float: - ps = set(pred.lower().split()) - rs = set(ref.lower().split()) - union = len(ps | rs) - if union == 0: - return 0.0 - return len(ps & rs) / union +from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: @@ -219,16 +168,16 @@ async def run_memsciqa_test( # 默认使用指定的 memsci 组 ID end_user_id = end_user_id or "group_memsci" - # 数据路径解析（项目根与当前工作目录兜底） + # 数据路径解析 if not data_path: - proj_path = os.path.join(PROJECT_ROOT, "data", "msc_self_instruct.jsonl") - cwd_path = os.path.join(os.getcwd(), "data", "msc_self_instruct.jsonl") - if os.path.exists(proj_path): - data_path = proj_path - elif os.path.exists(cwd_path): - data_path = cwd_path - else: - raise FileNotFoundError("未找到数据集: data/msc_self_instruct.jsonl，请确保其存在于项目根目录或当前工作目录的 data 目录下。") + dataset_dir = Path(__file__).resolve().parent.parent / "dataset" + data_path = str(dataset_dir / "msc_self_instruct.jsonl") + + if not os.path.exists(data_path): + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 msc_self_instruct.jsonl 放置在: {dataset_dir}" + ) # 加载数据 all_items = load_dataset_memsciqa(data_path) @@ -238,17 +187,13 @@ async def run_memsciqa_test( items = all_items[start_index:start_index + sample_size] # 初始化 LLM（纯测试：不进行摄入） - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm = factory.get_llm_client(SELECTED_LLM_ID) + llm = get_llm_client(os.getenv("EVAL_LLM_ID")) # 初始化 Neo4j 连接与向量检索 Embedder（对齐 locomo_test） connector = Neo4jConnector() embedder = None if search_type in ("embedding", "hybrid"): - with get_db_context() as db: - config_service = MemoryConfigService(db) - cfg_dict = config_service.get_embedder_config(SELECTED_EMBEDDING_ID) + cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID")) embedder = OpenAIEmbedderClient( model_config=RedBearModelConfig.model_validate(cfg_dict) ) @@ -273,7 +218,7 @@ async def run_memsciqa_test( question = item.get("self_instruct", {}).get("B", "") or item.get("question", "") reference = item.get("self_instruct", {}).get("A", "") or item.get("answer", "") - # 三路检索：chunks/statements/entities/summaries（对齐 qwen_search_eval.py） + # 检索：使用与 evaluate_qa.py 相同的 run_hybrid_search t0 = time.time() results = None try: @@ -302,57 +247,94 @@ async def run_memsciqa_test( search_ms = (t1 - t0) * 1000 latencies_search.append(search_ms) - # 构建上下文：包含 chunks、陈述、摘要和实体（对齐 qwen_search_eval.py） + # 构建上下文：与 evaluate_qa.py 完全一致的逻辑 contexts_all: List[str] = [] retrieved_counts: Dict[str, int] = {} if results: - chunks = results.get("chunks", []) - statements = results.get("statements", []) - entities = results.get("entities", []) - summaries = results.get("summaries", []) + # 处理 hybrid 搜索结果 + if search_type == "hybrid": + emb = results.get("embedding_search", {}) if isinstance(results.get("embedding_search"), dict) else {} + kw = results.get("keyword_search", {}) if isinstance(results.get("keyword_search"), dict) else {} + emb_dialogs = emb.get("dialogues", []) + emb_statements = emb.get("statements", []) + emb_entities = emb.get("entities", []) + kw_dialogs = kw.get("dialogues", []) + kw_statements = kw.get("statements", []) + kw_entities = kw.get("entities", []) + all_dialogs = emb_dialogs + kw_dialogs + all_statements = emb_statements + kw_statements + all_entities = emb_entities + kw_entities + + # 简单去重 + seen_dialog = set() + dialogues = [] + for d in all_dialogs: + key = (str(d.get("uuid", "")), str(d.get("content", ""))) + if key not in seen_dialog: + dialogues.append(d) + seen_dialog.add(key) + + seen_stmt = set() + statements = [] + for s in all_statements: + key = str(s.get("statement", "")) + if key not in seen_stmt: + statements.append(s) + seen_stmt.add(key) + + seen_ent = set() + entities = [] + for e in all_entities: + key = str(e.get("name", "")) + if key not in seen_ent: + entities.append(e) + seen_ent.add(key) + else: + # embedding 或 keyword 单独搜索 + dialogues = results.get("dialogues", []) + statements = results.get("statements", []) + entities = results.get("entities", []) + retrieved_counts = { - "chunks": len(chunks), + "dialogues": len(dialogues), "statements": len(statements), "entities": len(entities), - "summaries": len(summaries), } - # 优先使用 chunks - for c in chunks: - text = str(c.get("content", "")).strip() + + # 构建上下文文本 + for d in dialogues: + text = str(d.get("content", "")).strip() if text: contexts_all.append(text) - # 然后是 statements + for s in statements: text = str(s.get("statement", "")).strip() if text: contexts_all.append(text) - # 然后是 summaries - for sm in summaries: - text = str(sm.get("summary", "")).strip() - if text: - contexts_all.append(text) - # 实体摘要：最多加入前3个高分实体（对齐 qwen_search_eval.py） - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) + + # 实体摘要 + if entities: + scored = [e for e in entities if e.get("score") is not None] + top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] + if top_entities: + summary_lines = [] + for e in top_entities: + name = str(e.get("name", "")).strip() + etype = str(e.get("entity_type", "")).strip() + score = e.get("score") + if name: + meta = [] + if etype: + meta.append(f"type={etype}") + if isinstance(score, (int, float)): + meta.append(f"score={score:.3f}") + summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") + if summary_lines: + contexts_all.append("\n".join(summary_lines)) if verbose: if retrieved_counts: - print(f"✅ 检索成功: {retrieved_counts.get('chunks',0)} chunks, {retrieved_counts.get('statements',0)} 条陈述, {retrieved_counts.get('entities',0)} 个实体, {retrieved_counts.get('summaries',0)} 个摘要") + print(f"✅ 检索成功: {retrieved_counts.get('dialogues',0)} dialogues, {retrieved_counts.get('statements',0)} 条陈述, {retrieved_counts.get('entities',0)} 个实体, {retrieved_counts.get('summaries',0)} 个摘要") print(f"📊 有效上下文数量: {len(contexts_all)}") q_keywords = extract_question_keywords(question, max_keywords=8) if q_keywords: @@ -507,8 +489,8 @@ async def run_memsciqa_test( "llm_max_tokens": llm_max_tokens, "search_type": search_type, "start_index": start_index, - "llm_id": SELECTED_LLM_ID, - "retrieval_embedding_id": SELECTED_EMBEDDING_ID + "llm_id": os.getenv("EVAL_LLM_ID"), + "retrieval_embedding_id": os.getenv("EVAL_EMBEDDING_ID") }, "timestamp": datetime.now().isoformat(), } @@ -522,7 +504,7 @@ async def run_memsciqa_test( def main(): load_dotenv() parser = argparse.ArgumentParser(description="memsciqa 测试脚本（三路检索 + 智能上下文选择）") - parser.add_argument("--sample-size", type=int, default=30, help="样本数量（<=0 表示全部）") + parser.add_argument("--sample-size", type=int, default=10, help="样本数量（<=0 表示全部）") parser.add_argument("--all", action="store_true", help="评估全部样本（覆盖 --sample-size）") parser.add_argument("--start-index", type=int, default=0, help="起始样本索引") parser.add_argument("--group-id", type=str, default="group_memsci", help="图数据库 Group ID（默认 group_memsci）") diff --git a/api/app/core/memory/evaluation/memsciqa/evaluate_qa.py b/api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py similarity index 76% rename from api/app/core/memory/evaluation/memsciqa/evaluate_qa.py rename to api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py index 869fdb60..40684f4c 100644 --- a/api/app/core/memory/evaluation/memsciqa/evaluate_qa.py +++ b/api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py @@ -4,35 +4,20 @@ import json import os import time from datetime import datetime -from typing import TYPE_CHECKING, Any, Dict, List +from typing import List, Dict, Any +from pathlib import Path +from dotenv import load_dotenv -if TYPE_CHECKING: - from app.schemas.memory_config_schema import MemoryConfig +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None - -from app.core.memory.evaluation.common.metrics import ( - avg_context_tokens, - exact_match, - latency_stats, -) -from app.core.memory.evaluation.extraction_utils import ( - ingest_contexts_via_full_pipeline, -) -from app.core.memory.storage_services.search import run_hybrid_search -from app.core.memory.utils.config.definitions import ( - PROJECT_ROOT, - SELECTED_EMBEDDING_ID, - SELECTED_GROUP_ID, - SELECTED_LLM_ID, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.db import get_db_context from app.repositories.neo4j.neo4j_connector import Neo4jConnector +from app.core.memory.src.search import run_hybrid_search # 使用旧版本（重构前） +from app.core.memory.utils.llm.llm_utils import get_llm_client +from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline +from app.core.memory.evaluation.common.metrics import exact_match, latency_stats, avg_context_tokens def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: @@ -135,24 +120,37 @@ def _combine_dialogues_for_hybrid(results: Dict[str, Any]) -> List[Dict[str, Any return merged + async def run_memsciqa_eval(sample_size: int = 1, end_user_id: str | None = None, search_limit: int = 8, context_char_budget: int = 4000, llm_temperature: float = 0.0, llm_max_tokens: int = 64, search_type: str = "hybrid", memory_config: "MemoryConfig" = None) -> Dict[str, Any]: end_user_id = end_user_id or SELECTED_GROUP_ID + # Load data - data_path = os.path.join(PROJECT_ROOT, "data", "msc_self_instruct.jsonl") + dataset_dir = Path(__file__).resolve().parent.parent / "dataset" + data_path = dataset_dir / "msc_self_instruct.jsonl" + if not os.path.exists(data_path): - data_path = os.path.join(os.getcwd(), "data", "msc_self_instruct.jsonl") + raise FileNotFoundError( + f"数据集文件不存在: {data_path}\n" + f"请将 msc_self_instruct.jsonl 放置在: {dataset_dir}" + ) with open(data_path, "r", encoding="utf-8") as f: lines = f.readlines() items: List[Dict[str, Any]] = [json.loads(l) for l in lines[:sample_size]] + + # 改为：每条样本仅摄入一个上下文（完整对话转录），避免多上下文摄入 # 说明：memsciqa 数据集的每个样本天然只有一个对话，保持按样本一上下文的策略 contexts: List[str] = [build_context_from_dialog(item) for item in items] await ingest_contexts_via_full_pipeline(contexts, end_user_id) # LLM client (使用异步调用) - with get_db_context() as db: - factory = MemoryClientFactory(db) - llm_client = factory.get_llm_client(SELECTED_LLM_ID) + from app.db import get_db + + db = next(get_db()) + try: + llm_client = get_llm_client(os.getenv("EVAL_LLM_ID"), db) + finally: + db.close() # Evaluate each item connector = Neo4jConnector() @@ -177,7 +175,6 @@ async def run_memsciqa_eval(sample_size: int = 1, end_user_id: str | None = None limit=search_limit, include=["dialogues", "statements", "entities"], output_path=None, - memory_config=memory_config, ) except Exception: results = None @@ -261,11 +258,7 @@ async def run_memsciqa_eval(sample_size: int = 1, end_user_id: str | None = None pred = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else str(resp).strip()) # Metrics: F1, BLEU-1, Jaccard; keep exact match for reference correct_flags.append(exact_match(pred, reference)) - from app.core.memory.evaluation.common.metrics import ( - bleu1, - f1_score, - jaccard, - ) + from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard f1s.append(f1_score(str(pred), str(reference))) b1s.append(bleu1(str(pred), str(reference))) jss.append(jaccard(str(pred), str(reference))) @@ -295,15 +288,39 @@ async def run_memsciqa_eval(sample_size: int = 1, end_user_id: str | None = None def main(): + # Load environment variables first load_dotenv() + + # Get defaults from environment variables + env_sample_size = os.getenv("MEMSCIQA_SAMPLE_SIZE") + env_search_limit = os.getenv("MEMSCIQA_SEARCH_LIMIT") + env_context_budget = os.getenv("MEMSCIQA_CONTEXT_CHAR_BUDGET") + env_llm_max_tokens = os.getenv("MEMSCIQA_LLM_MAX_TOKENS") + env_skip_ingest = os.getenv("MEMSCIQA_SKIP_INGEST", "false").lower() in ("true", "1", "yes") + env_output_dir = os.getenv("MEMSCIQA_OUTPUT_DIR") + + # Convert to appropriate types with fallback to code defaults + default_sample_size = int(env_sample_size) if env_sample_size else 1 + default_search_limit = int(env_search_limit) if env_search_limit else 8 + default_context_budget = int(env_context_budget) if env_context_budget else 4000 + default_llm_max_tokens = int(env_llm_max_tokens) if env_llm_max_tokens else 64 + default_output_dir = env_output_dir if env_output_dir else None + parser = argparse.ArgumentParser(description="Evaluate DMR (memsciqa) with graph search and Qwen") + parser.add_argument("--sample-size", type=int, default=1, help="评测样本数量") - parser.add_argument("--group-id", type=str, default=None, help="可选 end_user_id，默认取 runtime.json") + parser.add_argument("--end-user-id", type=str, default=None, help="可选 end_user_id，默认使用环境变量") parser.add_argument("--search-limit", type=int, default=8, help="每类检索最大返回数") parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") + parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") - parser.add_argument("--llm-max-tokens", type=int, default=64, help="LLM 最大生成长度") + parser.add_argument("--llm-max-tokens", type=int, default=default_llm_max_tokens, + help=f"LLM 最大生成长度 (env: MEMSCIQA_LLM_MAX_TOKENS={env_llm_max_tokens or 'not set'})") parser.add_argument("--search-type", type=str, choices=["keyword","embedding","hybrid"], default="hybrid", help="检索类型") + parser.add_argument("--skip-ingest", action="store_true", default=env_skip_ingest, + help=f"跳过数据摄入，使用 Neo4j 中的现有数据 (env: MEMSCIQA_SKIP_INGEST={os.getenv('MEMSCIQA_SKIP_INGEST', 'false')})") + parser.add_argument("--output-dir", type=str, default=default_output_dir, + help=f"结果保存目录 (env: MEMSCIQA_OUTPUT_DIR={env_output_dir or 'not set'})") args = parser.parse_args() result = asyncio.run( @@ -315,9 +332,37 @@ def main(): llm_temperature=args.llm_temperature, llm_max_tokens=args.llm_max_tokens, search_type=args.search_type, + skip_ingest=args.skip_ingest, ) ) + + # Print results to console print(json.dumps(result, ensure_ascii=False, indent=2)) + + # Save results to file + output_dir = args.output_dir + if output_dir is None: + # Use absolute path to ensure results are saved in the correct location + script_dir = Path(__file__).resolve().parent + output_dir = script_dir / "results" + elif not Path(output_dir).is_absolute(): + # If relative path, make it relative to this script's directory + script_dir = Path(__file__).resolve().parent + output_dir = script_dir / output_dir + else: + output_dir = Path(output_dir) + + output_dir.mkdir(parents=True, exist_ok=True) + + timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S") + output_path = output_dir / f"memsciqa_{timestamp_str}.json" + + try: + with open(output_path, "w", encoding="utf-8") as f: + json.dump(result, f, ensure_ascii=False, indent=2) + print(f"\n✅ 结果已保存到: {output_path}") + except Exception as e: + print(f"\n❌ 保存结果失败: {e}") if __name__ == "__main__": diff --git a/api/app/core/memory/evaluation/run_eval.py b/api/app/core/memory/evaluation/run_eval.py index c5aacb2f..56b2e790 100644 --- a/api/app/core/memory/evaluation/run_eval.py +++ b/api/app/core/memory/evaluation/run_eval.py @@ -2,20 +2,16 @@ import argparse import asyncio import json import os -import sys from typing import Any, Dict +from pathlib import Path +from dotenv import load_dotenv -# Add src directory to Python path for proper imports when running from evaluation directory -sys.path.insert(0, os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), 'src')) - -try: - from dotenv import load_dotenv -except Exception: - def load_dotenv(): - return None +# Load evaluation config +eval_config_path = Path(__file__).resolve().parent / ".env.evaluation" +if eval_config_path.exists(): + load_dotenv(eval_config_path, override=True) from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.utils.config.definitions import SELECTED_GROUP_ID, PROJECT_ROOT from app.core.memory.evaluation.memsciqa.evaluate_qa import run_memsciqa_eval from app.core.memory.evaluation.longmemeval.qwen_search_eval import run_longmemeval_test @@ -36,8 +32,9 @@ async def run( start_index: int | None = None, max_contexts_per_item: int | None = None, ) -> Dict[str, Any]: - # 恢复原始风格：统一入口做路由，并沿用各数据集既有默认 - end_user_id = end_user_id or SELECTED_GROUP_ID + # Use environment variable with fallback chain if not provided + if end_user_id is None: + end_user_id = os.getenv("EVAL_END_USER_ID", "benchmark_default") if reset_group: connector = Neo4jConnector() diff --git a/api/app/core/memory/storage_services/extraction_engine/extraction_orchestrator.py b/api/app/core/memory/storage_services/extraction_engine/extraction_orchestrator.py index 8c69c7cf..7b7e854b 100644 --- a/api/app/core/memory/storage_services/extraction_engine/extraction_orchestrator.py +++ b/api/app/core/memory/storage_services/extraction_engine/extraction_orchestrator.py @@ -1064,13 +1064,16 @@ class ExtractionOrchestrator: if statement.triplet_extraction_info: triplet_info = statement.triplet_extraction_info - # 创建实体索引到ID的映射 + # 创建实体索引到ID的映射（支持多种索引方式） entity_idx_to_id = {} # 创建实体节点 for entity_idx, entity in enumerate(triplet_info.entities): - # 映射实体索引到实体ID + # 映射实体索引到实体ID（使用多个键以提高容错性） + # 1. 使用实体自己的 entity_idx entity_idx_to_id[entity.entity_idx] = entity.id + # 2. 使用枚举索引（从0开始） + entity_idx_to_id[entity_idx] = entity.id if entity.id not in entity_id_set: entity_connect_strength = getattr(entity, 'connect_strength', 'Strong') @@ -1149,9 +1152,18 @@ class ExtractionOrchestrator: relationship_result ) else: - logger.warning( - f"跳过三元组 - 无法找到实体ID: subject_id={triplet.subject_id}, " - f"object_id={triplet.object_id}, statement_id={statement.id}" + # 改进的警告信息，包含更多调试信息 + missing_subject = "subject" if not subject_entity_id else "" + missing_object = "object" if not object_entity_id else "" + missing_both = " and " if (not subject_entity_id and not object_entity_id) else "" + + logger.debug( + f"跳过三元组 - 无法找到{missing_subject}{missing_both}{missing_object}实体ID: " + f"subject_id={triplet.subject_id} ({triplet.subject_name}), " + f"object_id={triplet.object_id} ({triplet.object_name}), " + f"predicate={triplet.predicate}, " + f"statement_id={statement.id}, " + f"available_indices={sorted(entity_idx_to_id.keys())}" ) logger.info( diff --git a/api/app/models/agent_app_config_model.py b/api/app/models/agent_app_config_model.py index 0a7a5935..96752c8e 100644 --- a/api/app/models/agent_app_config_model.py +++ b/api/app/models/agent_app_config_model.py @@ -6,7 +6,7 @@ from sqlalchemy.orm import relationship from app.base.type import PydanticType from app.db import Base -from app.schemas import ModelParameters +from app.schemas.app_schema import ModelParameters class AgentConfig(Base): diff --git a/api/app/models/multi_agent_model.py b/api/app/models/multi_agent_model.py index 544ddb27..400c05ad 100644 --- a/api/app/models/multi_agent_model.py +++ b/api/app/models/multi_agent_model.py @@ -10,7 +10,7 @@ from sqlalchemy.orm import relationship from app.base.type import PydanticType from app.db import Base -from app.schemas import ModelParameters +from app.schemas.app_schema import ModelParameters class OrchestrationMode(StrEnum): diff --git a/api/app/schemas/multi_agent_schema.py b/api/app/schemas/multi_agent_schema.py index c0d72cdd..8fba2929 100644 --- a/api/app/schemas/multi_agent_schema.py +++ b/api/app/schemas/multi_agent_schema.py @@ -4,7 +4,7 @@ import datetime from typing import Optional, List, Dict, Any, Union from pydantic import BaseModel, Field, ConfigDict, field_serializer -from app.schemas import ModelParameters +from app.schemas.app_schema import ModelParameters # ==================== 子 Agent 配置 ==================== diff --git a/api/app/services/master_agent_router.py b/api/app/services/master_agent_router.py index 3971aab7..87fdb22c 100644 --- a/api/app/services/master_agent_router.py +++ b/api/app/services/master_agent_router.py @@ -5,7 +5,7 @@ import uuid from typing import Dict, Any, List, Optional, Tuple from sqlalchemy.orm import Session -from app.schemas import ModelParameters +from app.schemas.app_schema import ModelParameters from app.services.conversation_state_manager import ConversationStateManager from app.models import ModelConfig, AgentConfig from app.core.logging_config import get_business_logger diff --git a/api/app/utils/app_config_utils.py b/api/app/utils/app_config_utils.py index ae41d8bf..06549989 100644 --- a/api/app/utils/app_config_utils.py +++ b/api/app/utils/app_config_utils.py @@ -57,7 +57,7 @@ def dict_to_model_parameters(data: Optional[Dict[str, Any]]) -> Optional[Any]: if data is None: return None - from app.schemas import ModelParameters + from app.schemas.app_schema import ModelParameters if isinstance(data, ModelParameters): return data From 87731090cab1097dbb30909e388b62774bb10283 Mon Sep 17 00:00:00 2001 From: Mark Date: Mon, 26 Jan 2026 19:19:41 +0800 Subject: [PATCH 6/7] [modify] migration script --- api/migrations/versions/325b759cd66b_2026011240.py | 1 + 1 file changed, 1 insertion(+) diff --git a/api/migrations/versions/325b759cd66b_2026011240.py b/api/migrations/versions/325b759cd66b_2026011240.py index 763b0289..3d7443a8 100644 --- a/api/migrations/versions/325b759cd66b_2026011240.py +++ b/api/migrations/versions/325b759cd66b_2026011240.py @@ -31,6 +31,7 @@ def upgrade() -> None: op.execute("UPDATE memory_config SET config_id = apply_id::uuid") op.alter_column('memory_config', 'config_id', nullable=False) op.create_primary_key('memory_config_pkey', 'memory_config', ['config_id']) + op.execute("ALTER TABLE memory_config ALTER COLUMN config_id_old DROP DEFAULT") op.execute("DROP SEQUENCE IF EXISTS data_config_config_id_seq") From c3ea3b751b07c9325f80000e9a4d93a0a4790fe4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E4=B9=90=E5=8A=9B=E9=BD=90?= <162269739+lanceyq@users.noreply.github.com> Date: Mon, 26 Jan 2026 20:30:07 +0800 Subject: [PATCH 7/7] delete benchmark-test (#204) * Refactor: Move evaluation folder to redbear-mem-benchmark submodule * [changes]Restore .gitmodules --- .../memory/evaluation/.env.evaluation.example | 224 --- api/app/core/memory/evaluation/.gitignore | 13 - api/app/core/memory/evaluation/__init__.py | 1 - api/app/core/memory/evaluation/benchmark.md | 748 --------- .../memory/evaluation/check_enduser_data.py | 371 ----- .../core/memory/evaluation/common/metrics.py | 100 -- .../memory/evaluation/dialogue_queries.py | 62 - .../memory/evaluation/extraction_utils.py | 444 ------ .../evaluation/locomo/locomo_benchmark.py | 770 ---------- .../evaluation/locomo/locomo_metrics.py | 225 --- .../memory/evaluation/locomo/locomo_test.py | 864 ----------- .../memory/evaluation/locomo/locomo_utils.py | 687 --------- .../evaluation/locomo/qwen_search_eval.py | 874 ----------- .../longmemeval/longmemeval_benchmark.py | 1339 ----------------- .../evaluation/longmemeval/test_eval.py | 1312 ---------------- .../evaluation/memsciqa/memsciqa-test.py | 559 ------- .../evaluation/memsciqa/memsciqa_benchmark.py | 369 ----- api/app/core/memory/evaluation/run_eval.py | 147 -- redbear-mem-benchmark | 2 +- 19 files changed, 1 insertion(+), 9110 deletions(-) delete mode 100644 api/app/core/memory/evaluation/.env.evaluation.example delete mode 100644 api/app/core/memory/evaluation/.gitignore delete mode 100644 api/app/core/memory/evaluation/__init__.py delete mode 100644 api/app/core/memory/evaluation/benchmark.md delete mode 100644 api/app/core/memory/evaluation/check_enduser_data.py delete mode 100644 api/app/core/memory/evaluation/common/metrics.py delete mode 100644 api/app/core/memory/evaluation/dialogue_queries.py delete mode 100644 api/app/core/memory/evaluation/extraction_utils.py delete mode 100644 api/app/core/memory/evaluation/locomo/locomo_benchmark.py delete mode 100644 api/app/core/memory/evaluation/locomo/locomo_metrics.py delete mode 100644 api/app/core/memory/evaluation/locomo/locomo_test.py delete mode 100644 api/app/core/memory/evaluation/locomo/locomo_utils.py delete mode 100644 api/app/core/memory/evaluation/locomo/qwen_search_eval.py delete mode 100644 api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py delete mode 100644 api/app/core/memory/evaluation/longmemeval/test_eval.py delete mode 100644 api/app/core/memory/evaluation/memsciqa/memsciqa-test.py delete mode 100644 api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py delete mode 100644 api/app/core/memory/evaluation/run_eval.py diff --git a/api/app/core/memory/evaluation/.env.evaluation.example b/api/app/core/memory/evaluation/.env.evaluation.example deleted file mode 100644 index be089eb4..00000000 --- a/api/app/core/memory/evaluation/.env.evaluation.example +++ /dev/null @@ -1,224 +0,0 @@ -# ============================================================================ -# 基准测试统一配置文件示例 -# ============================================================================ -# 复制此文件为 .env.evaluation 并根据需要修改 -# 支持的基准测试：LoCoMo、LongMemEval、MemSciQA -# ============================================================================ - -# ============================================================================ -# 通用配置（所有基准测试共用） -# ============================================================================ - -# ---------------------------------------------------------------------------- -# Neo4j 配置 -# ---------------------------------------------------------------------------- -# 默认 Group ID（建议各基准测试使用独立的 group） -EVAL_GROUP_ID=benchmark_default - -# ---------------------------------------------------------------------------- -# 模型配置（必需） -# ---------------------------------------------------------------------------- -# ⚠️ 必填：从数据库 models 表中选择有效的模型 ID -# -# 如何获取模型 ID： -# 1. 查询数据库：SELECT id, model_name FROM models WHERE is_active = true; -# 2. 或通过系统管理界面查看 -# 3. 确保模型可用且配置正确 - -# LLM 模型 ID（必填） -EVAL_LLM_ID=your_llm_model_id_here - -# Embedding 模型 ID（必填） -EVAL_EMBEDDING_ID=your_embedding_model_id_here - -# ---------------------------------------------------------------------------- -# 检索参数 -# ---------------------------------------------------------------------------- -# 检索类型: "keyword", "embedding", "hybrid" -EVAL_SEARCH_TYPE=hybrid - -# 检索结果数量限制（默认值） -EVAL_SEARCH_LIMIT=12 - -# 上下文最大字符数（默认值） -EVAL_MAX_CONTEXT_CHARS=8000 - -# ---------------------------------------------------------------------------- -# LLM 参数 -# ---------------------------------------------------------------------------- -# LLM 温度参数（0.0 = 确定性输出） -EVAL_LLM_TEMPERATURE=0.0 - -# LLM 最大生成 token 数 -EVAL_LLM_MAX_TOKENS=32 - -# LLM 超时时间（秒） -EVAL_LLM_TIMEOUT=10.0 - -# LLM 最大重试次数 -EVAL_LLM_MAX_RETRIES=1 - -# ---------------------------------------------------------------------------- -# 数据处理参数 -# ---------------------------------------------------------------------------- -# Chunker 策略 -EVAL_CHUNKER_STRATEGY=RecursiveChunker - -# 是否在导入前清空现有数据 -EVAL_RESET_ON_INGEST=true - -# 是否保存详细日志 -EVAL_SAVE_DETAILED_LOGS=true - -# ============================================================================ -# LoCoMo 基准测试专用配置 -# ============================================================================ -# 数据集：locomo10.json -# 运行：python locomo_benchmark.py --sample_size 20 -# ---------------------------------------------------------------------------- - -# Group ID（LoCoMo 专用） -LOCOMO_GROUP_ID=locomo_benchmark - -# 测试样本数量 -# 建议值：20（快速测试）、100（中等测试）、1986（完整测试） -LOCOMO_SAMPLE_SIZE=20 - -# 检索结果数量限制 -LOCOMO_SEARCH_LIMIT=12 - -# 上下文最大字符数 -LOCOMO_CONTEXT_CHAR_BUDGET=8000 - -# 导入的对话数量 -LOCOMO_MAX_DIALOGUES=1 - -# 跳过数据摄入（true=跳过，false=摄入） -# 首次运行设置为 false，后续运行可设置为 true 以节省时间 -LOCOMO_SKIP_INGEST=false - -# 结果保存目录 -LOCOMO_OUTPUT_DIR=locomo/results - -# ============================================================================ -# LongMemEval 基准测试专用配置 -# ============================================================================ -# 数据集：longmemeval_oracle_zh.json -# 运行：python longmemeval_benchmark.py --sample_size 3 -# 特点：支持时间推理问题的增强检索 -# ---------------------------------------------------------------------------- - -# Group ID（LongMemEval 专用） -LONGMEMEVAL_GROUP_ID=longmemeval_zh_bak_3 - -# 测试样本数量（<=0 表示全部样本） -LONGMEMEVAL_SAMPLE_SIZE=3 - -# 起始样本索引 -LONGMEMEVAL_START_INDEX=0 - -# 检索结果数量限制 -LONGMEMEVAL_SEARCH_LIMIT=8 - -# 上下文最大字符数 -LONGMEMEVAL_CONTEXT_CHAR_BUDGET=4000 - -# LLM 最大生成 token 数 -LONGMEMEVAL_LLM_MAX_TOKENS=16 - -# 每条样本最多摄入的上下文段数 -LONGMEMEVAL_MAX_CONTEXTS_PER_ITEM=2 - -# 是否保存分块结果 -LONGMEMEVAL_SAVE_CHUNK_OUTPUT=true - -# 自定义分块输出路径（留空使用默认） -LONGMEMEVAL_SAVE_CHUNK_OUTPUT_PATH= - -# 摄入前是否清空组数据 -LONGMEMEVAL_RESET_GROUP_BEFORE_INGEST=false - -# 是否跳过摄入，仅检索评估 -LONGMEMEVAL_SKIP_INGEST=false - -# 结果保存目录 -LONGMEMEVAL_OUTPUT_DIR=longmemeval/results - -# ============================================================================ -# MemSciQA 基准测试专用配置 -# ============================================================================ -# 数据集：msc_self_instruct.jsonl -# 运行：python memsciqa_benchmark.py --sample_size 1 -# 特点：对话记忆检索评估 -# ---------------------------------------------------------------------------- - -# Group ID（MemSciQA 专用，独立数据集） -MEMSCIQA_GROUP_ID=memsciqa_benchmark - -# 测试样本数量 -MEMSCIQA_SAMPLE_SIZE=1 # 0或者-1标识测试数据集中的所有样本 - -# 检索结果数量限制 -MEMSCIQA_SEARCH_LIMIT=8 - -# 上下文最大字符数 -MEMSCIQA_CONTEXT_CHAR_BUDGET=4000 - -# LLM 最大生成 token 数 -MEMSCIQA_LLM_MAX_TOKENS=64 - -# 跳过数据摄入（true=跳过，false=摄入） -# 首次运行设置为 false，后续运行可设置为 true 以节省时间 -MEMSCIQA_SKIP_INGEST=false - -# 结果保存目录（相对于 memsciqa 脚本所在目录） -# 使用 "results" 会保存到 api/app/core/memory/evaluation/memsciqa/results/ -MEMSCIQA_OUTPUT_DIR=results - -# ============================================================================ -# 高级配置（可选） -# ============================================================================ - -# BM25 权重（用于混合检索，0.0-1.0） -EVAL_RERANK_ALPHA=0.6 - -# 是否使用遗忘重排序 -EVAL_USE_FORGETTING_RERANK=false - -# 是否使用 LLM 重排序 -EVAL_USE_LLM_RERANK=false - -# 连接重置间隔（每 N 个问题重置一次） -EVAL_RESET_INTERVAL=5 - -# 性能阈值（低于此值触发重置） -EVAL_PERFORMANCE_THRESHOLD=0.6 - -# ============================================================================ -# 快速配置指南 -# ============================================================================ -# 1. 复制此文件为 .env.evaluation -# 2. 修改 EVAL_LLM_ID 和 EVAL_EMBEDDING_ID 为你的模型 ID -# 3. 根据需要修改各基准测试的专用配置 -# 4. 运行测试： -# - LoCoMo: python locomo/locomo_benchmark.py --sample_size 20 -# - LongMemEval: python longmemeval/longmemeval_benchmark.py --sample_size 3 --all -# - MemSciQA: python memsciqa/memsciqa_benchmark.py --sample_size 10 -# 配置优先级： -# 命令行参数 > 特定配置（如 LOCOMO_*）> 通用配置（EVAL_*）> 代码默认值 -# ============================================================================ - - -# 执行LoCoMo测试 -# 只摄入前5条消息，评估3个问题（最小测试） -# python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 3 --max_ingest_messages 5 -# -# 如果数据已经摄入，跳过摄入阶段直接测试 -# python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 5 --skip_ingest - - -# 执行longmemeval测试 -# python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --sample-size 10 --max-contexts-per-item 3 --reset-group-before-ingest - -# 执行memsciqa测试 -# python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --sample-size 1 diff --git a/api/app/core/memory/evaluation/.gitignore b/api/app/core/memory/evaluation/.gitignore deleted file mode 100644 index 38b1055a..00000000 --- a/api/app/core/memory/evaluation/.gitignore +++ /dev/null @@ -1,13 +0,0 @@ -# 忽略实际的评估配置文件（包含敏感信息） -.env.evaluation - -# 保留示例文件 -!.env.evaluation.example - -# 忽略测试结果文件 -*/results/*.json -*/results/*.log - -# 忽略数据集文件（文件过大，不应提交到 Git） -dataset/*.json -dataset/*.jsonl diff --git a/api/app/core/memory/evaluation/__init__.py b/api/app/core/memory/evaluation/__init__.py deleted file mode 100644 index e9d6aa6c..00000000 --- a/api/app/core/memory/evaluation/__init__.py +++ /dev/null @@ -1 +0,0 @@ -"""Evaluation package with dataset-specific pipelines and a unified runner.""" diff --git a/api/app/core/memory/evaluation/benchmark.md b/api/app/core/memory/evaluation/benchmark.md deleted file mode 100644 index 7c31cccd..00000000 --- a/api/app/core/memory/evaluation/benchmark.md +++ /dev/null @@ -1,748 +0,0 @@ -# 1.数据集下载地址 -Locomo10.json : https://github.com/snap-research/locomo/tree/main/data -LongMemEval_oracle.json : https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned -msc_self_instruct.jsonl : https://huggingface.co/datasets/MemGPT/MSC-Self-Instruct - -数据集下载之后保存至api\app\core\memory\evaluation\dataset目录下 -# 2.配置说明 -文件api\app\core\memory\evaluation\.env.evaluation.example对三个基准测试所需配置有着详细的说明 -**实际配置文件**：api\app\core\memory\evaluation\.env.evaluation -```python -# 当使用不带配置参数的命令行执行基准测试，基准测试所需的配置参数根据.env.evaluation中的参数执行 -python -m app.core.memory.evaluation.locomo.locomo_benchmark -``` -**检查neo4j指定的grou_id是否摄入数据** -```python -# 1. 进入交互模式 -python -m app.core.memory.evaluation.check_enduser_data - -# 2. 选择 "1" 检查指定 group -# 3. 输入 group_id，例如: locomo_benchmark -# 4. 选择是否显示详细统计 (y/n) -``` -# 3.locomo - -### (1)locomo执行命令 -```python -# 首先进入api目录 -cd api - -# 只摄入前5条消息，评估3个问题（最小测试） -python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 3 --max_ingest_messages 5 - -# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) -python -m app.core.memory.evaluation.locomo.locomo_benchmark --sample_size 5 --skip_ingest -``` -### (2)locomo结果说明 - -#### 结果示例 -```json -{ - "dataset": "locomo", - "sample_size": 0, - "timestamp": "2026-01-26T11:24:28.239156", - "params": { - "group_id": "locomo_benchmark", - "search_type": "hybrid", - "search_limit": 12, - "context_char_budget": 8000, - "llm_id": "2c9b0782-7a85-4740-ba84-4baf77f256c4", - "embedding_id": "e2a6392d-ca63-4d59-a523-647420b59cb2" - }, - "overall_metrics": { - "f1": 0.0, - "bleu1": 0.0, - "jaccard": 0.0, - "locomo_f1": 0.0 - }, - "by_category": {}, - "latency": { - "search": { - "mean": 0.0, - "p50": 0.0, - "p95": 0.0, - "iqr": 0.0 - }, - "llm": { - "mean": 0.0, - "p50": 0.0, - "p95": 0.0, - "iqr": 0.0 - } - }, - "context_stats": { - "avg_retrieved_docs": 0.0, - "avg_context_chars": 0.0, - "avg_context_tokens": 0.0 - }, - "samples": [] -} -``` - -#### 参数详解 - -##### 1. 核心评估指标 (overall_metrics) - -**🎯 关键进步指标：** - -- **`f1`** (F1 Score): 精确率和召回率的调和平均值 - - 范围：0.0 - 1.0 - - **越高越好**，衡量检索和生成答案的准确性 - - 这是最重要的综合性能指标 - - 优秀标准：> 0.85 - -- **`bleu1`** (BLEU-1): 单词级别的匹配度 - - 范围：0.0 - 1.0 - - **越高越好**，衡量生成答案与标准答案的词汇重叠度 - - 关注词汇层面的准确性 - -- **`jaccard`** (Jaccard 相似度): 集合相似度 - - 范围：0.0 - 1.0 - - **越高越好**，衡量答案集合的相似性 - - 计算公式：交集大小 / 并集大小 - -- **`locomo_f1`**: Locomo 特定的 F1 分数 - - 范围：0.0 - 1.0 - - **越高越好**，针对 Locomo 数据集优化的评估指标 - - 考虑了长对话记忆的特殊性 - -##### 2. 性能指标 (latency) - -**⚡ 关键效率指标：** - -- **`search`**: 检索延迟统计（单位：毫秒） - - `mean`: 平均延迟 - - `p50`: 中位数延迟（50%的请求在此时间内完成） - - `p95`: 95分位数延迟（95%的请求在此时间内完成） - - `iqr`: 四分位距（Q3-Q1，衡量稳定性） - - **越低越好**，衡量记忆检索速度 - - 优秀标准：p95 < 2000ms - -- **`llm`**: LLM 推理延迟统计（单位：毫秒） - - `mean`: 平均推理时间 - - `p50`: 中位数推理时间 - - `p95`: 95分位数推理时间 - - `iqr`: 四分位距（越小越稳定） - - **越低越好**，衡量答案生成速度 - - 优秀标准：p95 < 3000ms - -##### 3. 上下文统计 (context_stats) - -**📊 资源效率指标：** - -- **`avg_retrieved_docs`**: 平均检索文档数 - - 反映检索策略的广度 - - 需要平衡：太少可能信息不足，太多增加噪音和延迟 - - 建议范围：8-15 个文档 - -- **`avg_context_chars`**: 平均上下文字符数 - - 反映检索内容的总量 - - 应在满足准确性前提下尽量精简 - - 受 `context_char_budget` 参数限制 - -- **`avg_context_tokens`**: 平均上下文 token 数 - - **越低越好**（在保持准确性前提下） - - 直接影响 API 调用成本和推理速度 - - 成本效益比 = f1 / avg_context_tokens - -##### 4. 分类统计 (by_category) - -- 按问题类型分类的性能指标 -- 帮助识别系统在不同场景下的强弱项 -- 可针对性优化特定类型的问题 - -#### 系统进步衡量标准 - -**一级指标（最重要）：** -- `f1` 和 `locomo_f1` 提升 → 核心能力提升 -- 目标：f1 > 0.85 - -**二级指标（重要）：** -- `latency.p95` 降低 → 用户体验提升 -- 目标：search.p95 < 2000ms, llm.p95 < 3000ms - -**三级指标（辅助）：** -- `avg_context_tokens` 降低（在保持 f1 前提下）→ 成本优化 -- `iqr` 降低 → 性能稳定性提升 -# 4.longmemeval -支持时间推理问题的增强检索 -### (1)执行命令 -```python -# 首先进入api目录 -cd api - -# 不带参数运行 - 使用环境变量 -python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark - -# 命令行参数覆盖环境变量 -python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --sample-size 2 - -# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) -python -m app.core.memory.evaluation.longmemeval.longmemeval_benchmark --skip_ingest -``` -### (2)结果说明 - -#### 结果示例 -```json -{ - "dataset": "longmemeval", - "items": 1, - "accuracy_by_type": { - "single-session-user": 1.0 - }, - "f1_by_type": { - "single-session-user": 1.0 - }, - "jaccard_by_type": { - "single-session-user": 1.0 - }, - "samples": [ - { - "question": "What degree did I graduate with?", - "prediction": "Business Administration", - "answer": "Business Administration", - "question_type": "single-session-user", - "is_temporal": false, - "question_id": "e47becba", - "options": [], - "context_count": 13, - "context_chars": 1268, - "retrieved_dialogue_count": 0, - "retrieved_statement_count": 12, - "metrics": { - "exact_match": true, - "f1": 1.0, - "jaccard": 1.0 - }, - "timing": { - "search_ms": 1483.100175857544, - "llm_ms": 995.8682060241699 - } - } - ], - "latency": { - "search": { - "mean": 1483.100175857544, - "p50": 1483.100175857544, - "p95": 1483.100175857544, - "iqr": 0.0 - }, - "llm": { - "mean": 995.8682060241699, - "p50": 995.8682060241699, - "p95": 995.8682060241699, - "iqr": 0.0 - } - }, - "context": { - "avg_tokens": 204.0, - "avg_chars": 1268, - "count_avg": 13 - }, - "params": { - "group_id": "longmemeval_zh_bak_3", - "search_limit": 8, - "context_char_budget": 4000, - "search_type": "hybrid", - "llm_id": "6dc52e1b-9cec-4194-af66-a74c6307fc3f", - "embedding_id": "e2a6392d-ca63-4d59-a523-647420b59cb2", - "sample_size": 1, - "start_index": 0 - }, - "timestamp": "2026-01-24T21:36:10.818308", - "metric_summary": { - "score_accuracy": 100.0, - "latency_median_s": 2.478968381881714, - "latency_iqr_s": 0.0, - "avg_context_tokens_k": 0.204 - }, - "diagnostics": { - "duplicate_previews_top": [], - "unique_preview_count": 1 - } -} -``` - -#### 参数详解 - -##### 1. 核心评估指标 - -**🎯 关键进步指标：** - -- **`accuracy_by_type`**: 按问题类型分类的准确率 - - 范围：0.0 - 1.0 - - **越高越好**，1.0 表示 100% 准确 - - 问题类型包括： - - `single-session-user`: 单会话用户信息 - - `single-session-event`: 单会话事件信息 - - `multi-session-user`: 多会话用户信息 - - `multi-session-event`: 多会话事件信息 - - 可以识别系统在不同场景下的强弱项 - -- **`f1_by_type`**: 按问题类型的 F1 分数 - - 范围：0.0 - 1.0 - - **越高越好**，综合评估精确率和召回率 - - 比单纯的准确率更全面 - -- **`jaccard_by_type`**: 按问题类型的 Jaccard 相似度 - - 范围：0.0 - 1.0 - - **越高越好**，衡量答案集合匹配度 - - 对于集合类答案特别有用 - -##### 2. 样本级指标 (samples) - -**详细诊断指标：** - -- **`metrics.exact_match`**: 精确匹配（布尔值） - - **true 越多越好**，最严格的评估标准 - - 要求预测答案与标准答案完全一致 - -- **`metrics.f1`**: 单个样本的 F1 分数 - - 范围：0.0 - 1.0 - - **越高越好**，衡量单个问题的回答质量 - -- **`is_temporal`**: 是否为时间推理问题 - - 布尔值，标识问题是否涉及时间推理 - - 时间推理问题通常更具挑战性 - -- **`context_count`**: 检索到的上下文数量 - - 反映检索策略的有效性 - - 建议范围：8-15 个上下文片段 - -- **`retrieved_dialogue_count`**: 检索到的对话数 -- **`retrieved_statement_count`**: 检索到的陈述数 - - 这两个指标帮助理解检索的内容类型分布 - - 可用于优化检索策略 - -- **`timing.search_ms`**: 单个问题的检索延迟（毫秒） -- **`timing.llm_ms`**: 单个问题的 LLM 推理延迟（毫秒） - - **越低越好**，反映单次查询的响应速度 - -##### 3. 汇总指标 (metric_summary) - -**📊 关键 KPI：** - -- **`score_accuracy`**: 总体准确率百分比 - - 范围：0.0 - 100.0 - - **越高越好**，最直观的性能指标 - - 优秀标准：> 90.0 - -- **`latency_median_s`**: 中位延迟（秒） - - **越低越好**，反映真实响应速度 - - 优秀标准：< 3.0 秒 - -- **`latency_iqr_s`**: 延迟四分位距（秒） - - **越低越好**，反映性能稳定性 - - 越小说明响应时间越稳定 - -- **`avg_context_tokens_k`**: 平均上下文 token 数（千） - - **越低越好**（在保持准确性前提下） - - 直接影响 API 调用成本 - - 成本效益比 = score_accuracy / (avg_context_tokens_k * 1000) - -##### 4. 上下文统计 (context) - -- **`avg_tokens`**: 平均 token 数 -- **`avg_chars`**: 平均字符数 -- **`count_avg`**: 平均上下文片段数 - - 这些指标反映检索内容的规模 - - 需要在准确性和效率之间平衡 - -##### 5. 性能指标 (latency) - -**⚡ 效率指标：** - -- **`search`**: 检索延迟统计（单位：毫秒） - - `mean`: 平均延迟 - - `p50`: 中位数延迟 - - `p95`: 95分位数延迟 - - `iqr`: 四分位距 - - **越低越好**，衡量记忆检索速度 - -- **`llm`**: LLM 推理延迟统计（单位：毫秒） - - `mean`: 平均推理时间 - - `p50`: 中位数推理时间 - - `p95`: 95分位数推理时间 - - `iqr`: 四分位距 - - **越低越好**，衡量答案生成速度 - -##### 6. 诊断信息 (diagnostics) - -- **`duplicate_previews_top`**: 重复预览统计 - - 列出出现频率最高的重复内容 - - 帮助发现检索冗余问题 - - 应该尽量减少重复 - -- **`unique_preview_count`**: 唯一预览数量 - - 反映检索多样性 - - **越高越好**，说明检索到的内容更丰富 - -#### 系统进步衡量标准 - -**一级指标（最重要）：** -- `score_accuracy` 提升 → 核心能力提升 -- 目标：> 90.0% -- 各类型的 `accuracy_by_type` 均衡提升 → 全面能力提升 - -**二级指标（重要）：** -- `latency_median_s` 降低 → 用户体验提升 -- 目标：< 3.0 秒 -- `exact_match` 比例提升 → 精确度提升 - -**三级指标（辅助）：** -- `avg_context_tokens_k` 降低（在保持准确性前提下）→ 成本优化 -- `unique_preview_count` 提升 → 检索多样性提升 -- `latency_iqr_s` 降低 → 性能稳定性提升 - -**特殊关注：** -- 时间推理问题（`is_temporal: true`）的准确率 -- 多会话问题的准确率（通常更具挑战性） -# 5.memsciqa -对话记忆检索评估 -### (1)执行命令 -```python -# 首先进入api目录 -cd api - -# 不带参数运行 - 使用环境变量 -python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark - -# 命令行参数覆盖环境变量 -python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --sample-size 100 - -# 如果数据已经摄入，跳过摄入阶段直接测试(使用skip_ingest参数) -python -m app.core.memory.evaluation.memsciqa.memsciqa_benchmark --skip_ingest -``` -### (2)结果说明 - -#### 结果示例 -```json -{ - "dataset": "memsciqa", - "items": 1, - "metrics": { - "accuracy": 0.0, - "f1": 0.0, - "bleu1": 0.0, - "jaccard": 0.0 - }, - "latency": { - "search": { - "mean": 0.0, - "p50": 0.0, - "p95": 0.0, - "iqr": 0.0 - }, - "llm": { - "mean": 3067.7285194396973, - "p50": 3067.7285194396973, - "p95": 3067.7285194396973, - "iqr": 0.0 - } - }, - "avg_context_tokens": 4.0 -} -``` - -#### 参数详解 - -##### 1. 核心评估指标 (metrics) - -**🎯 关键进步指标：** - -- **`accuracy`**: 准确率 - - 范围：0.0 - 1.0 - - **越高越好**，最直接的性能指标 - - 衡量系统回答正确的问题比例 - - 优秀标准：> 0.85 - -- **`f1`**: F1 分数 - - 范围：0.0 - 1.0 - - **越高越好**，平衡精确率和召回率 - - 计算公式：2 * (precision * recall) / (precision + recall) - - 比单纯的准确率更全面，特别适合不平衡数据集 - -- **`bleu1`**: BLEU-1 分数 - - 范围：0.0 - 1.0 - - **越高越好**，衡量词汇级别的匹配度 - - 关注生成答案与标准答案的单词重叠 - - 源自机器翻译评估，适用于自然语言生成 - -- **`jaccard`**: Jaccard 相似度 - - 范围：0.0 - 1.0 - - **越高越好**，衡量集合相似性 - - 计算公式：|A ∩ B| / |A ∪ B| - - 对于多答案或集合类问题特别有用 - -##### 2. 性能指标 (latency) - -**⚡ 效率指标：** - -- **`search`**: 检索延迟统计（单位：毫秒） - - `mean`: 平均检索延迟 - - `p50`: 中位数延迟（50%的请求在此时间内完成） - - `p95`: 95分位数延迟（95%的请求在此时间内完成） - - `iqr`: 四分位距（Q3-Q1，衡量稳定性） - - **越低越好**，衡量记忆检索效率 - - 优秀标准：p95 < 2000ms - -- **`llm`**: LLM 推理延迟统计（单位：毫秒） - - `mean`: 平均推理时间 - - `p50`: 中位数推理时间 - - `p95`: 95分位数推理时间 - - `iqr`: 四分位距（越小越稳定） - - **越低越好**，衡量答案生成速度 - - 优秀标准：p95 < 3000ms - - 注意：LLM 延迟通常占总延迟的大部分 - -##### 3. 资源指标 - -- **`avg_context_tokens`**: 平均上下文 token 数 - - **越低越好**（在保持准确性前提下） - - 直接影响： - - API 调用成本（按 token 计费） - - 推理速度（token 越多越慢） - - 上下文窗口占用 - - 成本效益比 = accuracy / avg_context_tokens - - 建议范围：根据模型上下文窗口和成本预算调整 - -##### 4. 数据集特点 - -- **`items`**: 评估的问题数量 - - 样本量越大，评估结果越可靠 - - 建议至少 100 个样本以获得稳定的评估结果 - -- **对话记忆特性**： - - MemSciQA 专注于对话历史中的记忆检索 - - 评估系统从多轮对话中提取和回忆信息的能力 - - 模拟真实的对话场景 - -#### 系统进步衡量标准 - -**一级指标（最重要）：** -- `accuracy` 提升 → 核心能力提升 -- 目标：> 0.85 -- `f1` 提升 → 综合性能提升 -- 目标：> 0.80 - -**二级指标（重要）：** -- `latency.p95` 降低 → 用户体验提升 - - search.p95 目标：< 2000ms - - llm.p95 目标：< 3000ms -- `iqr` 降低 → 性能稳定性提升 - -**三级指标（辅助）：** -- `avg_context_tokens` 降低（在保持准确性前提下）→ 成本优化 -- `bleu1` 和 `jaccard` 提升 → 答案质量提升 - -**综合评估：** -- 成本效益比 = accuracy / avg_context_tokens - - 该比值越高，说明系统在相同成本下性能越好 -- 总延迟 = search.p95 + llm.p95 - - 应控制在 5 秒以内以保证良好的用户体验 - -#### 优化建议 - -**提升准确性：** -- 优化检索算法（调整 hybrid search 参数） -- 改进 embedding 模型质量 -- 增加检索上下文数量（`search_limit`） -- 优化 prompt 工程 - -**提升效率：** -- 减少不必要的检索文档 -- 使用更快的 LLM 模型或量化版本 -- 实施缓存策略（相似问题复用结果） -- 优化数据库索引 - -**平衡性能：** -- 监控 accuracy vs latency 的权衡 -- 监控 accuracy vs cost (tokens) 的权衡 -- 根据业务需求调整优先级 - - ---- - -# 6. 三个基准测试对比总结 - -## 6.1 测试特点对比 - -| 基准测试 | 主要评估目标 | 数据集特点 | 适用场景 | -|---------|------------|-----------|---------| -| **Locomo** | 长对话记忆检索 | 长对话历史，多轮交互 | 评估长期记忆保持和检索能力 | -| **LongMemEval** | 时间推理和多会话记忆 | 支持时间推理，多会话场景 | 评估时间感知和跨会话记忆能力 | -| **MemSciQA** | 对话记忆问答 | 对话历史问答 | 评估对话上下文理解和记忆提取 | - -## 6.2 核心指标对比 - -### 准确性指标 - -| 指标 | Locomo | LongMemEval | MemSciQA | 说明 | -|-----|--------|-------------|----------|------| -| **F1 Score** | ✅ | ✅ | ✅ | 所有测试都使用，最重要的综合指标 | -| **Accuracy** | ❌ | ✅ | ✅ | 直观的准确率指标 | -| **BLEU-1** | ✅ | ❌ | ✅ | 词汇级别匹配度 | -| **Jaccard** | ✅ | ✅ | ✅ | 集合相似度 | -| **Exact Match** | ❌ | ✅ | ❌ | 最严格的评估标准 | - -### 性能指标 - -所有三个测试都包含： -- **检索延迟** (search latency): mean, p50, p95, iqr -- **LLM 延迟** (llm latency): mean, p50, p95, iqr -- **上下文统计**: token 数、字符数、文档数 - -## 6.3 关键进步指标优先级 - -### 🥇 一级指标（必须关注） - -1. **准确性指标** - - Locomo: `f1`, `locomo_f1` - - LongMemEval: `score_accuracy`, `accuracy_by_type` - - MemSciQA: `accuracy`, `f1` - - **目标**: > 85% 或 > 0.85 - -2. **综合性能** - - 所有测试的 F1 分数应保持一致性 - - 不同类型问题的准确率应均衡 - -### 🥈 二级指标（重要） - -3. **响应延迟** - - `latency.p95` (95分位数延迟) - - **目标**: - - search.p95 < 2000ms - - llm.p95 < 3000ms - - 总延迟 < 5000ms - -4. **性能稳定性** - - `iqr` (四分位距) - - **目标**: 越小越好，说明性能稳定 - -### 🥉 三级指标（优化） - -5. **成本效率** - - `avg_context_tokens` - - **目标**: 在保持准确性前提下最小化 - - 成本效益比 = accuracy / avg_context_tokens - -6. **检索质量** - - `avg_retrieved_docs` 的合理性 - - `unique_preview_count` (LongMemEval) - - 检索内容的多样性和相关性 - -## 6.4 系统优化路径 - -### 阶段一：提升准确性（优先级最高） - -**目标**: 所有测试的准确率 > 85% - -**优化方向**: -1. 改进 embedding 模型质量 -2. 优化检索算法（hybrid search 参数） -3. 增加检索上下文数量（`search_limit`） -4. 优化 prompt 工程 -5. 改进记忆存储结构 - -**监控指标**: -- Locomo: `f1`, `locomo_f1` -- LongMemEval: `score_accuracy`, `exact_match` 比例 -- MemSciQA: `accuracy`, `f1` - -### 阶段二：优化性能（准确性达标后） - -**目标**: p95 延迟 < 5 秒，性能稳定 - -**优化方向**: -1. 优化数据库索引和查询 -2. 实施缓存策略 -3. 使用更快的 LLM 模型 -4. 并行化检索和推理 -5. 减少不必要的检索 - -**监控指标**: -- `latency.p50`, `latency.p95` -- `iqr` (稳定性) -- 各阶段耗时分布 - -### 阶段三：降低成本（性能达标后） - -**目标**: 在保持准确性和性能前提下，最小化成本 - -**优化方向**: -1. 精简检索上下文 -2. 优化 context 选择策略 -3. 使用更小的 LLM 模型 -4. 实施智能缓存 -5. 批处理优化 - -**监控指标**: -- `avg_context_tokens` -- 成本效益比 = accuracy / avg_context_tokens -- API 调用成本 - -## 6.5 评估最佳实践 - -### 测试执行建议 - -1. **初始测试**: 使用小样本快速验证 - ```bash - --sample_size 10 - ``` - -2. **完整评估**: 使用足够大的样本量 - ```bash - --sample_size 100 # 或更多 - ``` - -3. **增量测试**: 数据已摄入时跳过摄入阶段 - ```bash - --skip_ingest - ``` - -4. **参数调优**: 系统性地调整参数并记录结果 - - 调整 `search_limit`: 4, 8, 12, 16 - - 调整 `context_char_budget`: 2000, 4000, 8000 - - 尝试不同的 `search_type`: vector, keyword, hybrid - -### 结果分析建议 - -1. **横向对比**: 比较三个测试的结果，识别系统的强弱项 -2. **纵向对比**: 跟踪同一测试在不同版本的表现 -3. **分类分析**: 关注不同问题类型的性能差异 -4. **异常诊断**: 分析失败案例，找出根本原因 - -### 持续监控 - -建议建立监控仪表板，跟踪： -- 核心指标趋势（准确率、延迟） -- 成本效益比趋势 -- 不同问题类型的性能分布 -- 异常样本和失败模式 - -## 6.6 性能基准参考 - -### 优秀水平（Production Ready） - -- **准确性**: accuracy/f1 > 0.90 -- **延迟**: p95 < 3 秒 -- **稳定性**: iqr < 500ms -- **成本效益**: accuracy/tokens > 0.0001 - -### 良好水平（Acceptable） - -- **准确性**: accuracy/f1 > 0.85 -- **延迟**: p95 < 5 秒 -- **稳定性**: iqr < 1000ms -- **成本效益**: accuracy/tokens > 0.00005 - -### 需要改进（Below Target） - -- **准确性**: accuracy/f1 < 0.85 -- **延迟**: p95 > 5 秒 -- **稳定性**: iqr > 1000ms -- **成本效益**: accuracy/tokens < 0.00005 - ---- - -**注**: 以上标准仅供参考，实际目标应根据具体业务需求和资源约束调整。 diff --git a/api/app/core/memory/evaluation/check_enduser_data.py b/api/app/core/memory/evaluation/check_enduser_data.py deleted file mode 100644 index 18ecbb34..00000000 --- a/api/app/core/memory/evaluation/check_enduser_data.py +++ /dev/null @@ -1,371 +0,0 @@ -""" -交互式 Neo4j End User 数据检查工具 - -用于查询指定 end_user_id 在 Neo4j 中是否存在数据，以及数据的详细统计信息。 - -使用方法: - python check_group_data.py - python check_group_data.py --group-id locomo_benchmark - python check_group_data.py --group-id memsciqa_benchmark --detailed -""" - -import asyncio -import argparse -import os -from pathlib import Path -from typing import Dict, Any -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}\n") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector - - -async def check_group_exists(end_user_id: str) -> Dict[str, Any]: - """ - 检查指定 end_user_id 是否存在数据 - - Args: - end_user_id: 要检查的 end_user ID - - Returns: - 包含统计信息的字典 - """ - connector = Neo4jConnector() - - try: - # 查询该 end_user 的节点总数 - query_total = """ - MATCH (n {end_user_id: $end_user_id}) - RETURN count(n) as total_nodes - """ - result_total = await connector.execute_query(query_total, end_user_id=end_user_id) - total_nodes = result_total[0]["total_nodes"] if result_total else 0 - - # 查询各类型节点的数量 - query_by_type = """ - MATCH (n {end_user_id: $end_user_id}) - RETURN labels(n) as labels, count(n) as count - ORDER BY count DESC - """ - result_by_type = await connector.execute_query(query_by_type, end_user_id=end_user_id) - - # 查询关系数量 - query_relationships = """ - MATCH (n {end_user_id: $end_user_id})-[r]-() - RETURN count(DISTINCT r) as total_relationships - """ - result_rel = await connector.execute_query(query_relationships, end_user_id=end_user_id) - total_relationships = result_rel[0]["total_relationships"] if result_rel else 0 - - return { - "exists": total_nodes > 0, - "total_nodes": total_nodes, - "total_relationships": total_relationships, - "nodes_by_type": result_by_type - } - - finally: - await connector.close() - - -async def get_detailed_stats(end_user_id: str) -> Dict[str, Any]: - """ - 获取详细的统计信息 - - Args: - end_user_id: 要检查的 end_user ID - - Returns: - 详细统计信息字典 - """ - connector = Neo4jConnector() - - try: - stats = {} - - # Chunk 节点统计 - query_chunks = """ - MATCH (c:Chunk {end_user_id: $end_user_id}) - RETURN count(c) as count, - avg(size(c.content)) as avg_content_length - """ - result_chunks = await connector.execute_query(query_chunks, end_user_id=end_user_id) - if result_chunks and result_chunks[0]["count"] > 0: - stats["chunks"] = { - "count": result_chunks[0]["count"], - "avg_content_length": int(result_chunks[0]["avg_content_length"]) if result_chunks[0]["avg_content_length"] else 0 - } - - # Statement 节点统计 - query_statements = """ - MATCH (s:Statement {end_user_id: $end_user_id}) - RETURN count(s) as count - """ - result_statements = await connector.execute_query(query_statements, end_user_id=end_user_id) - if result_statements and result_statements[0]["count"] > 0: - stats["statements"] = { - "count": result_statements[0]["count"] - } - - # Entity 节点统计 - query_entities = """ - MATCH (e:Entity {end_user_id: $end_user_id}) - RETURN count(e) as count, - count(DISTINCT e.entity_type) as unique_types - """ - result_entities = await connector.execute_query(query_entities, end_user_id=end_user_id) - if result_entities and result_entities[0]["count"] > 0: - stats["entities"] = { - "count": result_entities[0]["count"], - "unique_types": result_entities[0]["unique_types"] - } - - # Dialogue 节点统计 - query_dialogues = """ - MATCH (d:Dialogue {end_user_id: $end_user_id}) - RETURN count(d) as count - """ - result_dialogues = await connector.execute_query(query_dialogues, end_user_id=end_user_id) - if result_dialogues and result_dialogues[0]["count"] > 0: - stats["dialogues"] = { - "count": result_dialogues[0]["count"] - } - - # Summary 节点统计 - query_summaries = """ - MATCH (s:Summary {end_user_id: $end_user_id}) - RETURN count(s) as count - """ - result_summaries = await connector.execute_query(query_summaries, end_user_id=end_user_id) - if result_summaries and result_summaries[0]["count"] > 0: - stats["summaries"] = { - "count": result_summaries[0]["count"] - } - - return stats - - finally: - await connector.close() - - -async def list_all_end_users() -> list: - """ - 列出数据库中所有的 end_user_id - - Returns: - end_user_id 列表及其节点数量 - """ - connector = Neo4jConnector() - - try: - query = """ - MATCH (n) - WHERE n.end_user_id IS NOT NULL - RETURN DISTINCT n.end_user_id as end_user_id, count(n) as node_count - ORDER BY node_count DESC - """ - results = await connector.execute_query(query) - return results - - finally: - await connector.close() - - -def print_results(end_user_id: str, stats: Dict[str, Any], detailed_stats: Dict[str, Any] = None): - """ - 打印查询结果 - - Args: - end_user_id: End User ID - stats: 基本统计信息 - detailed_stats: 详细统计信息（可选） - """ - print(f"\n{'='*60}") - print(f"📊 End User ID: {end_user_id}") - print(f"{'='*60}\n") - - if not stats["exists"]: - print("❌ 该 end_user_id 不存在数据") - print("\n💡 提示: 请先运行基准测试以摄入数据") - return - - print(f"✅ 该 end_user_id 存在数据\n") - print(f"📈 基本统计:") - print(f" 总节点数: {stats['total_nodes']}") - print(f" 总关系数: {stats['total_relationships']}") - - if stats["nodes_by_type"]: - print(f"\n📋 节点类型分布:") - for item in stats["nodes_by_type"]: - labels = ", ".join(item["labels"]) - count = item["count"] - print(f" {labels}: {count}") - - if detailed_stats: - print(f"\n🔍 详细统计:") - - if "chunks" in detailed_stats: - print(f" Chunks: {detailed_stats['chunks']['count']} 个") - print(f" 平均内容长度: {detailed_stats['chunks']['avg_content_length']} 字符") - - if "statements" in detailed_stats: - print(f" Statements: {detailed_stats['statements']['count']} 个") - - if "entities" in detailed_stats: - print(f" Entities: {detailed_stats['entities']['count']} 个") - print(f" 唯一类型数: {detailed_stats['entities']['unique_types']}") - - if "dialogues" in detailed_stats: - print(f" Dialogues: {detailed_stats['dialogues']['count']} 个") - - if "summaries" in detailed_stats: - print(f" Summaries: {detailed_stats['summaries']['count']} 个") - - print(f"\n{'='*60}\n") - - -async def interactive_mode(): - """ - 交互式模式 - """ - print("\n" + "="*60) - print("🔍 Neo4j End User 数据检查工具 - 交互模式") - print("="*60 + "\n") - - while True: - print("\n请选择操作:") - print(" 1. 检查指定 end_user_id") - print(" 2. 列出所有 end_user_id") - print(" 3. 退出") - - choice = input("\n请输入选项 (1-3): ").strip() - - if choice == "1": - end_user_id = input("\n请输入 end_user_id: ").strip() - if not end_user_id: - print("❌ end_user_id 不能为空") - continue - - detailed = input("是否显示详细统计? (y/n, 默认 n): ").strip().lower() == 'y' - - print("\n🔄 正在查询...") - stats = await check_group_exists(end_user_id) - - detailed_stats = None - if detailed and stats["exists"]: - detailed_stats = await get_detailed_stats(end_user_id) - - print_results(end_user_id, stats, detailed_stats) - - elif choice == "2": - print("\n🔄 正在查询所有 end_user_id...") - end_users = await list_all_end_users() - - if not end_users: - print("\n❌ 数据库中没有任何 end_user 数据") - else: - print(f"\n{'='*60}") - print(f"📋 数据库中的所有 End User ID") - print(f"{'='*60}\n") - - for idx, end_user in enumerate(end_users, 1): - print(f" {idx}. {end_user['end_user_id']}") - print(f" 节点数: {end_user['node_count']}") - - print(f"\n{'='*60}\n") - - elif choice == "3": - print("\n👋 再见！") - break - - else: - print("\n❌ 无效的选项，请重新选择") - - -async def main(): - """ - 主函数 - """ - parser = argparse.ArgumentParser( - description="检查 Neo4j 中指定 end_user_id 的数据情况", - formatter_class=argparse.RawDescriptionHelpFormatter, - epilog=""" -示例: - # 交互模式 - python check_group_data.py - - # 检查指定 end_user - python check_group_data.py --end-user-id locomo_benchmark - - # 检查并显示详细统计 - python check_group_data.py --end-user-id memsciqa_benchmark --detailed - - # 列出所有 end_user - python check_group_data.py --list-all - """ - ) - - parser.add_argument( - "--end-user-id", - type=str, - help="要检查的 end_user ID" - ) - - parser.add_argument( - "--detailed", - action="store_true", - help="显示详细统计信息" - ) - - parser.add_argument( - "--list-all", - action="store_true", - help="列出所有 end_user_id" - ) - - args = parser.parse_args() - - # 如果没有提供任何参数，进入交互模式 - if not args.end_user_id and not args.list_all: - await interactive_mode() - return - - # 列出所有 end_user - if args.list_all: - print("\n🔄 正在查询所有 end_user_id...") - end_users = await list_all_end_users() - - if not end_users: - print("\n❌ 数据库中没有任何 end_user 数据") - else: - print(f"\n{'='*60}") - print(f"📋 数据库中的所有 End User ID") - print(f"{'='*60}\n") - - for idx, end_user in enumerate(end_users, 1): - print(f" {idx}. {end_user['end_user_id']}") - print(f" 节点数: {end_user['node_count']}") - - print(f"\n{'='*60}\n") - return - - # 检查指定 end_user - if args.end_user_id: - print(f"\n🔄 正在查询 end_user_id: {args.end_user_id}...") - stats = await check_group_exists(args.end_user_id) - - detailed_stats = None - if args.detailed and stats["exists"]: - print("🔄 正在获取详细统计...") - detailed_stats = await get_detailed_stats(args.end_user_id) - - print_results(args.end_user_id, stats, detailed_stats) - - -if __name__ == "__main__": - asyncio.run(main()) diff --git a/api/app/core/memory/evaluation/common/metrics.py b/api/app/core/memory/evaluation/common/metrics.py deleted file mode 100644 index 961ce7f0..00000000 --- a/api/app/core/memory/evaluation/common/metrics.py +++ /dev/null @@ -1,100 +0,0 @@ -import math -import re -from typing import List, Dict - -# 评估指标的实现 -def _normalize(text: str) -> List[str]: - """Lowercase, strip punctuation, and split into tokens.""" - text = text.lower().strip() - # Python's re doesn't support \p classes; use a simple non-word filter - text = re.sub(r"[^\w\s]", " ", text) - tokens = [t for t in text.split() if t] - return tokens - - -def exact_match(pred: str, ref: str) -> float: - return float(_normalize(pred) == _normalize(ref)) - - -def jaccard(pred: str, ref: str) -> float: - p = set(_normalize(pred)) - r = set(_normalize(ref)) - if not p and not r: - return 1.0 - if not p or not r: - return 0.0 - return len(p & r) / len(p | r) - - -def f1_score(pred: str, ref: str) -> float: - p_tokens = _normalize(pred) - r_tokens = _normalize(ref) - if not p_tokens and not r_tokens: - return 1.0 - if not p_tokens or not r_tokens: - return 0.0 - p_set = set(p_tokens) - r_set = set(r_tokens) - tp = len(p_set & r_set) - precision = tp / len(p_set) if p_set else 0.0 - recall = tp / len(r_set) if r_set else 0.0 - if precision + recall == 0: - return 0.0 - return 2 * precision * recall / (precision + recall) - - -def bleu1(pred: str, ref: str) -> float: - """Unigram BLEU (BLEU-1) with clipping and brevity penalty.""" - p_tokens = _normalize(pred) - r_tokens = _normalize(ref) - if not p_tokens: - return 0.0 - # Clipped count - r_counts: Dict[str, int] = {} - for t in r_tokens: - r_counts[t] = r_counts.get(t, 0) + 1 - clipped = 0 - p_counts: Dict[str, int] = {} - for t in p_tokens: - p_counts[t] = p_counts.get(t, 0) + 1 - for t, c in p_counts.items(): - clipped += min(c, r_counts.get(t, 0)) - precision = clipped / max(len(p_tokens), 1) - # Brevity penalty - ref_len = len(r_tokens) - pred_len = len(p_tokens) - if pred_len > ref_len or pred_len == 0: - bp = 1.0 - else: - bp = math.exp(1 - ref_len / max(pred_len, 1)) - return bp * precision - - -def percentile(values: List[float], p: float) -> float: - if not values: - return 0.0 - vals = sorted(values) - k = (len(vals) - 1) * p - f = math.floor(k) - c = math.ceil(k) - if f == c: - return vals[int(k)] - return vals[f] + (k - f) * (vals[c] - vals[f]) - - -def latency_stats(latencies_ms: List[float]) -> Dict[str, float]: - """Return basic latency stats: mean, p50, p95, iqr (p75-p25).""" - if not latencies_ms: - return {"mean": 0.0, "p50": 0.0, "p95": 0.0, "iqr": 0.0} - p25 = percentile(latencies_ms, 0.25) - p50 = percentile(latencies_ms, 0.50) - p75 = percentile(latencies_ms, 0.75) - p95 = percentile(latencies_ms, 0.95) - mean = sum(latencies_ms) / max(len(latencies_ms), 1) - return {"mean": mean, "p50": p50, "p95": p95, "iqr": p75 - p25} - - -def avg_context_tokens(contexts: List[str]) -> float: - if not contexts: - return 0.0 - return sum(len(_normalize(c)) for c in contexts) / len(contexts) diff --git a/api/app/core/memory/evaluation/dialogue_queries.py b/api/app/core/memory/evaluation/dialogue_queries.py deleted file mode 100644 index 0aace0ec..00000000 --- a/api/app/core/memory/evaluation/dialogue_queries.py +++ /dev/null @@ -1,62 +0,0 @@ -""" -Dialogue search queries for evaluation purposes. -This file contains Cypher queries for searching dialogues, entities, and chunks. -Placed in evaluation directory to avoid circular imports with src modules. -""" - -# 应该是neo4j browser的cypher语句，需要修改文件名 - -# Entity search queries -SEARCH_ENTITIES_BY_NAME = """ -MATCH (e:ExtractedEntity) -WHERE e.name = $name -RETURN e -""" - -SEARCH_ENTITIES_BY_NAME_FALLBACK = """ -MATCH (e:ExtractedEntity) -WHERE e.name CONTAINS $name -RETURN e -""" - -# Chunk search queries -SEARCH_CHUNKS_BY_CONTENT = """ -MATCH (c:Chunk) -WHERE c.content CONTAINS $content -RETURN c -""" - -# Dialogue search queries -SEARCH_DIALOGUE_BY_DIALOG_ID = """ -MATCH (d:Dialogue) -WHERE d.dialog_id = $dialog_id -RETURN d -""" - -SEARCH_DIALOGUES_BY_CONTENT = """ -MATCH (d:Dialogue) -WHERE d.content CONTAINS $q -RETURN d -""" - -DIALOGUE_EMBEDDING_SEARCH = """ -WITH $embedding AS q -MATCH (d:Dialogue) -WHERE d.dialog_embedding IS NOT NULL - AND ($end_user_id IS NULL OR d.end_user_id = $end_user_id) -WITH d, q, d.dialog_embedding AS v -WITH d, - reduce(dot = 0.0, i IN range(0, size(q)-1) | dot + toFloat(q[i]) * toFloat(v[i])) AS dot, - sqrt(reduce(qs = 0.0, i IN range(0, size(q)-1) | qs + toFloat(q[i]) * toFloat(q[i]))) AS qnorm, - sqrt(reduce(vs = 0.0, i IN range(0, size(v)-1) | vs + toFloat(v[i]) * toFloat(v[i]))) AS vnorm -WITH d, CASE WHEN qnorm = 0 OR vnorm = 0 THEN 0.0 ELSE dot / (qnorm * vnorm) END AS score -WHERE score > $threshold -RETURN d.id AS dialog_id, - d.end_user_id AS end_user_id, - d.content AS content, - d.created_at AS created_at, - d.expired_at AS expired_at, - score -ORDER BY score DESC -LIMIT $limit -""" diff --git a/api/app/core/memory/evaluation/extraction_utils.py b/api/app/core/memory/evaluation/extraction_utils.py deleted file mode 100644 index 43ef6fe0..00000000 --- a/api/app/core/memory/evaluation/extraction_utils.py +++ /dev/null @@ -1,444 +0,0 @@ -import os -import asyncio -import json -from typing import List, Dict, Any, Optional -from datetime import datetime -from uuid import UUID -import re - -from app.core.memory.llm_tools.openai_client import LLMClient -from app.core.memory.storage_services.extraction_engine.knowledge_extraction.chunk_extraction import DialogueChunker -from app.core.memory.models.message_models import DialogData, ConversationContext, ConversationMessage -import os -import sys -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent / "app" / "core" / "memory" / "evaluation" / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.utils.llm.llm_utils import get_llm_client - -# 使用新的模块化架构 -from app.core.memory.storage_services.extraction_engine.extraction_orchestrator import ExtractionOrchestrator - -# Import from database module -from app.repositories.neo4j.graph_saver import save_dialog_and_statements_to_neo4j - -# Cypher queries for evaluation -# Note: Entity, chunk, and dialogue search queries have been moved to evaluation/dialogue_queries.py - - -async def ingest_contexts_via_full_pipeline( - contexts: List[str], - end_user_id: str, - chunker_strategy: str | None = None, - embedding_name: str | None = None, - save_chunk_output: bool = False, - save_chunk_output_path: str | None = None, - reset_group: bool = False, -) -> bool: - """ - 使用新的 ExtractionOrchestrator 运行完整的提取流水线 - - Run the full extraction pipeline on provided dialogue contexts and save to Neo4j. - This function uses the new ExtractionOrchestrator architecture for better maintainability. - - Args: - contexts: List of dialogue texts, each containing lines like "role: message". - end_user_id: Group ID to assign to generated DialogData and graph nodes. - chunker_strategy: Optional chunker strategy; defaults to SELECTED_CHUNKER_STRATEGY. - embedding_name: Optional embedding model ID; defaults to SELECTED_EMBEDDING_ID. - save_chunk_output: If True, write chunked DialogData list to a JSON file for debugging. - save_chunk_output_path: Optional output path; defaults to src/chunker_test_output.txt. - reset_group: If True, clear existing data for this group before ingestion. - Returns: - True if data saved successfully, False otherwise. - """ - chunker_strategy = chunker_strategy or os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker") - embedding_name = embedding_name or os.getenv("EVAL_EMBEDDING_ID") - - # Check if we should reset from environment variable if not explicitly set - if not reset_group: - reset_group = os.getenv("EVAL_RESET_ON_INGEST", "false").lower() in ("true", "1", "yes") - - # Step 0: Reset group if requested - if reset_group: - print(f"[Ingestion] 🗑️ 清空 end_user '{end_user_id}' 的现有数据...") - try: - from app.repositories.neo4j.neo4j_connector import Neo4jConnector - connector = Neo4jConnector() - try: - # 删除该 end_user 的所有节点和关系 - query = """ - MATCH (n {end_user_id: $end_user_id}) - DETACH DELETE n - """ - await connector.execute_query(query, end_user_id=end_user_id) - print(f"[Ingestion] ✅ End User '{end_user_id}' 已清空") - finally: - await connector.close() - except Exception as e: - print(f"[Ingestion] ⚠️ 清空 end_user 失败: {e}") - # 继续执行，不中断摄入流程 - - # Step 1: Initialize LLM client - llm_client = None - try: - # 使用评估配置中的 LLM ID - llm_id = os.getenv("EVAL_LLM_ID") - if not llm_id: - print("[Ingestion] ❌ EVAL_LLM_ID not set in .env.evaluation") - return False - - from app.db import get_db - - db = next(get_db()) - try: - llm_client = get_llm_client(llm_id, db) - finally: - db.close() - except Exception as e: - print(f"[Ingestion] LLM client unavailable: {e}") - return False - - # Step 2: Parse contexts and create DialogData with chunks - print(f"[Ingestion] Parsing {len(contexts)} contexts...") - chunker = DialogueChunker(chunker_strategy) - dialog_data_list: List[DialogData] = [] - - for idx, ctx in enumerate(contexts): - messages: List[ConversationMessage] = [] - - # Improved parsing: capture multi-line message blocks, normalize roles - pattern = r"^\s*(用户|AI|assistant|user)\s*[：:]\s*(.+?)(?=\n\s*(?:用户|AI|assistant|user)\s*[：:]|\Z)" - matches = list(re.finditer(pattern, ctx, flags=re.MULTILINE | re.DOTALL)) - - if matches: - for m in matches: - raw_role = m.group(1).strip() - content = m.group(2).strip() - norm_role = "AI" if raw_role.lower() in ("ai", "assistant") else "用户" - messages.append(ConversationMessage(role=norm_role, msg=content)) - else: - # Fallback: line-by-line parsing - for raw in ctx.split("\n"): - line = raw.strip() - if not line: - continue - m = re.match(r'^\s*([^:：]+)\s*[：:]\s*(.+)', line) - if m: - role = m.group(1).strip() - msg = m.group(2).strip() - norm_role = "AI" if role.lower() in ("ai", "assistant") else "用户" - messages.append(ConversationMessage(role=norm_role, msg=msg)) - else: - # Final fallback: treat as user message - default_role = "AI" if re.match(r'^\s*(assistant|AI)\b', line, flags=re.IGNORECASE) else "用户" - messages.append(ConversationMessage(role=default_role, msg=line)) - - context_model = ConversationContext(msgs=messages) - dialog = DialogData( - context=context_model, - ref_id=f"pipeline_item_{idx}", - end_user_id=end_user_id, - user_id="default_user", - apply_id="default_application", - ) - # Generate chunks - dialog.chunks = await chunker.process_dialogue(dialog) - dialog_data_list.append(dialog) - - if not dialog_data_list: - print("[Ingestion] No dialogs to process.") - return False - - print(f"[Ingestion] Parsed {len(dialog_data_list)} dialogs with chunks") - - # Step 3: Optionally save chunking outputs for debugging - if save_chunk_output: - try: - def _serialize_datetime(obj): - if isinstance(obj, datetime): - return obj.isoformat() - raise TypeError(f"Object of type {obj.__class__.__name__} is not JSON serializable") - - from app.core.config import settings - settings.ensure_memory_output_dir() - default_path = settings.get_memory_output_path("chunker_test_output.txt") - out_path = save_chunk_output_path or default_path - - combined_output = [dd.model_dump() for dd in dialog_data_list] - with open(out_path, "w", encoding="utf-8") as f: - json.dump(combined_output, f, ensure_ascii=False, indent=4, default=_serialize_datetime) - print(f"[Ingestion] Saved chunking results to: {out_path}") - except Exception as e: - print(f"[Ingestion] Failed to save chunking results: {e}") - - # Step 4: Initialize embedder client - from app.core.models.base import RedBearModelConfig - from app.core.memory.utils.config.config_utils import get_embedder_config - from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient - from app.db import get_db - - try: - db = next(get_db()) - try: - embedder_config_dict = get_embedder_config(embedding_name, db) - embedder_config = RedBearModelConfig(**embedder_config_dict) - embedder_client = OpenAIEmbedderClient(embedder_config) - finally: - db.close() - except Exception as e: - print(f"[Ingestion] Failed to initialize embedder client: {e}") - return False - - # Step 5: Initialize Neo4j connector - connector = Neo4jConnector() - - # Step 6: 构建 MemoryConfig（从环境变量直接构建，不依赖数据库） - print("[Ingestion] 构建 MemoryConfig from environment variables...") - from app.schemas.memory_config_schema import MemoryConfig - - try: - # 从环境变量获取配置参数 - llm_id = os.getenv("EVAL_LLM_ID") - embedding_id = os.getenv("EVAL_EMBEDDING_ID") - chunker_strategy_env = os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker") - - if not llm_id or not embedding_id: - print("[Ingestion] ❌ EVAL_LLM_ID or EVAL_EMBEDDING_ID is not set in .env.evaluation") - print("[Ingestion] Please set both EVAL_LLM_ID and EVAL_EMBEDDING_ID") - await connector.close() - return False - - # 从数据库获取模型信息（仅用于显示名称） - from app.db import get_db - db = next(get_db()) - try: - from sqlalchemy import text - # 获取 LLM 模型信息（从 model_configs 表） - llm_result = db.execute( - text("SELECT name FROM model_configs WHERE id = :id"), - {"id": llm_id} - ).fetchone() - llm_model_name = llm_result[0] if llm_result else "Unknown LLM" - - # 获取 Embedding 模型信息（从 model_configs 表） - emb_result = db.execute( - text("SELECT name FROM model_configs WHERE id = :id"), - {"id": embedding_id} - ).fetchone() - embedding_model_name = emb_result[0] if emb_result else "Unknown Embedding" - except Exception as e: - # 如果查询失败，使用默认名称 - print(f"[Ingestion] Warning: Failed to query model names from database: {e}") - llm_model_name = f"LLM ({llm_id[:8]}...)" - embedding_model_name = f"Embedding ({embedding_id[:8]}...)" - finally: - db.close() - - # 构建 MemoryConfig 对象（使用最小必需配置） - from uuid import uuid4 - memory_config = MemoryConfig( - config_id=0, # 评估环境不需要真实的 config_id - config_name="evaluation_config", - workspace_id=uuid4(), # 临时 workspace_id - workspace_name="evaluation_workspace", - tenant_id=uuid4(), # 临时 tenant_id - llm_model_id=UUID(llm_id), - llm_model_name=llm_model_name, - embedding_model_id=UUID(embedding_id), - embedding_model_name=embedding_model_name, - storage_type="neo4j", - chunker_strategy=chunker_strategy_env, - reflexion_enabled=False, - reflexion_iteration_period=3, - reflexion_range="partial", - reflexion_baseline="TIME", - loaded_at=datetime.now(), - # 可选字段使用默认值 - rerank_model_id=None, - rerank_model_name=None, - llm_params={}, - embedding_params={}, - config_version="2.0", - ) - - print(f"[Ingestion] ✅ 构建 MemoryConfig 成功") - print(f"[Ingestion] LLM: {llm_model_name}") - print(f"[Ingestion] Embedding: {embedding_model_name}") - print(f"[Ingestion] Chunker: {chunker_strategy_env}") - - except Exception as e: - print(f"[Ingestion] ❌ Failed to build MemoryConfig: {e}") - print(f"[Ingestion] Please check:") - print(f"[Ingestion] 1. EVAL_LLM_ID and EVAL_EMBEDDING_ID are set in .env.evaluation") - print(f"[Ingestion] 2. Model IDs exist in the models table") - print(f"[Ingestion] 3. Database connection is working") - await connector.close() - return False - - # Step 7: Initialize and run ExtractionOrchestrator - print("[Ingestion] Running extraction pipeline with ExtractionOrchestrator...") - from app.services.memory_config_service import MemoryConfigService - config = MemoryConfigService.get_pipeline_config(memory_config) - - orchestrator = ExtractionOrchestrator( - llm_client=llm_client, - embedder_client=embedder_client, - connector=connector, - config=config, - embedding_id=str(memory_config.embedding_model_id), # 传递 embedding_id - ) - - try: - # Run the complete extraction pipeline - result = await orchestrator.run(dialog_data_list, is_pilot_run=False) - - # Handle different return formats: - # - Pilot mode: 7 values (without dedup_details) - # - Normal mode: 8 values (with dedup_details at the end) - if len(result) == 8: - # Normal mode: includes dedup_details - ( - dialogue_nodes, - chunk_nodes, - statement_nodes, - entity_nodes, - statement_chunk_edges, - statement_entity_edges, - entity_entity_edges, - _, # dedup_details - not needed here - ) = result - elif len(result) == 7: - # Pilot mode or older version: no dedup_details - ( - dialogue_nodes, - chunk_nodes, - statement_nodes, - entity_nodes, - statement_chunk_edges, - statement_entity_edges, - entity_entity_edges, - ) = result - else: - raise ValueError(f"Unexpected number of return values: {len(result)}") - - print(f"[Ingestion] Extraction completed: {len(statement_nodes)} statements, {len(entity_nodes)} entities") - - except ValueError as e: - # If unpacking fails, provide helpful error message - print(f"[Ingestion] Extraction pipeline result unpacking failed: {e}") - print(f"[Ingestion] Result type: {type(result)}, length: {len(result) if hasattr(result, '__len__') else 'N/A'}") - if hasattr(result, '__len__') and len(result) > 0: - print(f"[Ingestion] First element type: {type(result[0])}") - await connector.close() - return False - except Exception as e: - print(f"[Ingestion] Extraction pipeline failed: {e}") - import traceback - traceback.print_exc() - await connector.close() - return False - - # Step 7: Generate memory summaries - print("[Ingestion] Generating memory summaries...") - try: - from app.core.memory.storage_services.extraction_engine.knowledge_extraction.memory_summary import ( - memory_summary_generation, - ) - from app.repositories.neo4j.add_nodes import add_memory_summary_nodes - from app.repositories.neo4j.add_edges import add_memory_summary_statement_edges - - summaries = await memory_summary_generation( - chunked_dialogs=dialog_data_list, - llm_client=llm_client, - embedder_client=embedder_client - ) - print(f"[Ingestion] Generated {len(summaries)} memory summaries") - except Exception as e: - print(f"[Ingestion] Warning: Failed to generate memory summaries: {e}") - summaries = [] - - # Step 8: Save to Neo4j - print("[Ingestion] Saving to Neo4j...") - try: - success = await save_dialog_and_statements_to_neo4j( - dialogue_nodes=dialogue_nodes, - chunk_nodes=chunk_nodes, - statement_nodes=statement_nodes, - entity_nodes=entity_nodes, - entity_edges=entity_entity_edges, - statement_chunk_edges=statement_chunk_edges, - statement_entity_edges=statement_entity_edges, - connector=connector - ) - - # Save memory summaries separately - if summaries: - try: - await add_memory_summary_nodes(summaries, connector) - await add_memory_summary_statement_edges(summaries, connector) - print(f"[Ingestion] Saved {len(summaries)} memory summary nodes to Neo4j") - except Exception as e: - print(f"[Ingestion] Warning: Failed to save summary nodes: {e}") - - await connector.close() - - if success: - print("[Ingestion] Successfully saved all data to Neo4j!") - else: - print("[Ingestion] Failed to save data to Neo4j") - return success - - except Exception as e: - print(f"[Ingestion] Failed to save data to Neo4j: {e}") - await connector.close() - return False - - -async def handle_context_processing(args): - """Handle context-based processing from command line arguments.""" - contexts = [] - - if args.contexts: - contexts.extend(args.contexts) - - if args.context_file: - try: - with open(args.context_file, 'r', encoding='utf-8') as f: - contexts.extend(line.strip() for line in f if line.strip()) - except Exception as e: - print(f"Error reading context file: {e}") - return False - - if not contexts: - print("No contexts provided for processing.") - return False - - return await main_from_contexts(contexts, args.context_end_user_id) - - -async def main_from_contexts(contexts: List[str], end_user_id: str): - """Run the pipeline from provided dialogue contexts instead of test data.""" - print("=== Running pipeline from provided contexts ===") - - success = await ingest_contexts_via_full_pipeline( - contexts=contexts, - end_user_id=end_user_id, - chunker_strategy=SELECTED_CHUNKER_STRATEGY, - embedding_name=SELECTED_EMBEDDING_ID, - save_chunk_output=True - ) - - if success: - print("Successfully processed and saved contexts to Neo4j!") - else: - print("Failed to process contexts.") - - return success diff --git a/api/app/core/memory/evaluation/locomo/locomo_benchmark.py b/api/app/core/memory/evaluation/locomo/locomo_benchmark.py deleted file mode 100644 index eed75016..00000000 --- a/api/app/core/memory/evaluation/locomo/locomo_benchmark.py +++ /dev/null @@ -1,770 +0,0 @@ -""" -LoCoMo Benchmark Script - -This module provides the main entry point for running LoCoMo benchmark evaluations. -It orchestrates data loading, ingestion, retrieval, LLM inference, and metric calculation -in a clean, maintainable way. - -Usage: - python locomo_benchmark.py --sample_size 20 --search_type hybrid -""" - -import argparse -import asyncio -import json -import os -import time -from datetime import datetime -from typing import List, Dict, Any, Optional -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.models.base import RedBearModelConfig -from app.core.memory.utils.config.config_utils import get_embedder_config -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.common.metrics import ( - f1_score, - bleu1, - jaccard, - latency_stats, - avg_context_tokens -) -from app.core.memory.evaluation.locomo.locomo_metrics import ( - locomo_f1_score, - locomo_multi_f1, - get_category_name -) -from app.core.memory.evaluation.locomo.locomo_utils import ( - load_locomo_data, - extract_conversations, - resolve_temporal_references, - select_and_format_information, - retrieve_relevant_information, -) -from app.core.memory.utils.llm.llm_utils import MemoryClientFactory -from app.db import get_db_context -from app.services.memory_config_service import MemoryConfigService - -# Get configuration from environment variables -PROJECT_ROOT = str(Path(__file__).resolve().parents[5]) # api directory -SELECTED_EMBEDDING_ID = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") -SELECTED_end_user_id = os.getenv("LOCOMO_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "locomo_benchmark") -SELECTED_LLM_ID = os.getenv("EVAL_LLM_ID", "2c9b0782-7a85-4740-ba84-4baf77f256c4") - - -# ============================================================================ -# Step 1: Data Loading -# ============================================================================ - -def step_load_data(data_path: str, sample_size: int) -> List[Dict[str, Any]]: - """ - Load QA pairs from LoCoMo dataset. - - Args: - data_path: Path to locomo10.json file - sample_size: Number of QA pairs to load (0 for all) - - Returns: - List of QA items from the first conversation - """ - print("📂 Loading LoCoMo data...") - - # Load the dataset - qa_items = load_locomo_data(data_path, sample_size) - - print(f"✅ Loaded {len(qa_items)} QA pairs from first conversation\n") - return qa_items - - -# ============================================================================ -# Step 2: Data Ingestion -# ============================================================================ - -async def ingest_conversations_if_needed( - conversations: List[str], - end_user_id: str, - reset: bool = False -) -> bool: - """ - Ingest conversations into Neo4j database. - - Args: - conversations: List of conversation strings (already formatted) - end_user_id: Database end_user ID - reset: Whether to reset the group before ingestion - - Returns: - True if successful, False otherwise - """ - try: - from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline - - # Conversations are already formatted as strings, use them directly - await ingest_contexts_via_full_pipeline(conversations, end_user_id) - return True - - except Exception as e: - print(f"⚠️ Ingestion error: {e}") - import traceback - traceback.print_exc() - return False - - -async def step_ingest_data( - data_path: str, - end_user_id: str, - skip_ingest: bool, - reset_group: bool, - max_messages: Optional[int] = None -) -> bool: - """ - Ingest conversations into Neo4j database if needed. - - Args: - data_path: Path to locomo10.json file - end_user_id: Database end_user ID - skip_ingest: Whether to skip ingestion - reset_group: Whether to reset the group before ingestion - max_messages: Maximum messages per dialogue to ingest (for testing) - - Returns: - True if ingestion succeeded or was skipped, False otherwise - """ - if skip_ingest: - print("⏭️ Skipping data ingestion (using existing data in Neo4j)") - print(f" End User ID: {end_user_id}\n") - else: - print("💾 Checking database ingestion...") - try: - # Extract conversations with optional message limit - conversations = extract_conversations( - data_path, - max_dialogues=1, - max_messages_per_dialogue=max_messages - ) - print(f"📝 Extracted {len(conversations)} conversations") - - # Always ingest for now (ingestion check not implemented) - print(f"🔄 Ingesting conversations into end_user '{end_user_id}'...") - success = await ingest_conversations_if_needed( - conversations=conversations, - end_user_id=end_user_id, - reset=reset_group - ) - - if success: - print("✅ Ingestion completed successfully\n") - else: - print("⚠️ Ingestion may have failed, continuing anyway\n") - - except Exception as e: - print(f"❌ Ingestion failed: {e}") - import traceback - traceback.print_exc() - print("⚠️ Continuing with evaluation (database may be empty)\n") - - return True - - -# ============================================================================ -# Step 3: Initialize Clients -# ============================================================================ - -def step_initialize_clients(llm_id: str, embedding_id: str): - """ - Initialize Neo4j connector, LLM client, and embedder. - - Args: - llm_id: LLM model ID - embedding_id: Embedding model ID - - Returns: - Tuple of (connector, llm_client, embedder) - """ - print("🔧 Initializing clients...") - - connector = Neo4jConnector() - - # Get database session - from app.db import get_db - db = next(get_db()) - try: - llm_client = get_llm_client(llm_id, db) - cfg_dict = get_embedder_config(embedding_id, db) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - finally: - db.close() - - print("✅ Clients initialized\n") - return connector, llm_client, embedder - - -# ============================================================================ -# Step 4: Process Questions -# ============================================================================ - -async def step_process_all_questions( - qa_items: List[Dict[str, Any]], - end_user_id: str, - search_type: str, - search_limit: int, - context_char_budget: int, - connector: Neo4jConnector, - embedder: OpenAIEmbedderClient, - llm_client: Any -) -> List[Dict[str, Any]]: - """Process all QA items: retrieve, generate, and calculate metrics.""" - print(f"🔍 Processing {len(qa_items)} questions...") - print(f"{'='*60}\n") - - samples: List[Dict[str, Any]] = [] - anchor_date = datetime(2023, 5, 8) - - for idx, item in enumerate(qa_items, 1): - question = item.get("question", "") - ground_truth = item.get("answer", "") - category = get_category_name(item) - ground_truth_str = str(ground_truth) if ground_truth is not None else "" - - print(f"[{idx}/{len(qa_items)}] Category: {category}") - print(f"❓ Question: {question}") - print(f"✅ Ground Truth: {ground_truth_str}") - - # Retrieve - t_search_start = time.time() - try: - retrieved_info = await retrieve_relevant_information( - question=question, - end_user_id=end_user_id, - search_type=search_type, - search_limit=search_limit, - connector=connector, - embedder=embedder - ) - search_latency = (time.time() - t_search_start) * 1000 - print(f"🔍 Retrieved {len(retrieved_info)} documents ({search_latency:.1f}ms)") - except Exception as e: - print(f"❌ Retrieval failed: {e}") - retrieved_info = [] - search_latency = 0.0 - - # Format context - context_text = select_and_format_information( - retrieved_info=retrieved_info, - question=question, - max_chars=context_char_budget - ) - context_text = resolve_temporal_references(context_text, anchor_date) - if context_text: - context_text = f"Reference date: {anchor_date.date().isoformat()}\n\n{context_text}" - else: - context_text = "No relevant context found." - - print(f"📝 Context: {len(context_text)} chars, {len(retrieved_info)} docs") - - # Generate answer - messages = [ - { - "role": "system", - "content": ( - "You are a precise QA assistant. Answer following these rules:\n" - "1) Extract the EXACT information mentioned in the context\n" - "2) For time questions: calculate actual dates from relative times\n" - "3) Return ONLY the answer text in simplest form\n" - "4) For dates, use format 'DD Month YYYY' (e.g., '7 May 2023')\n" - "5) If no clear answer found, respond with 'Unknown'" - ) - }, - { - "role": "user", - "content": f"Question: {question}\n\nContext:\n{context_text}" - } - ] - - t_llm_start = time.time() - try: - response = await llm_client.chat(messages=messages) - llm_latency = (time.time() - t_llm_start) * 1000 - if hasattr(response, 'content'): - prediction = response.content.strip() - elif isinstance(response, dict): - prediction = response["choices"][0]["message"]["content"].strip() - else: - prediction = "Unknown" - print(f"🤖 Prediction: {prediction} ({llm_latency:.1f}ms)") - except Exception as e: - print(f"❌ LLM failed: {e}") - prediction = "Unknown" - llm_latency = 0.0 - - # Calculate metrics - f1_val = f1_score(prediction, ground_truth_str) - bleu1_val = bleu1(prediction, ground_truth_str) - jaccard_val = jaccard(prediction, ground_truth_str) - if item.get("category") == 1: - locomo_f1_val = locomo_multi_f1(prediction, ground_truth_str) - else: - locomo_f1_val = locomo_f1_score(prediction, ground_truth_str) - - print(f"📊 Metrics - F1: {f1_val:.3f}, BLEU-1: {bleu1_val:.3f}, " - f"Jaccard: {jaccard_val:.3f}, LoCoMo F1: {locomo_f1_val:.3f}") - print() - - samples.append({ - "question": question, - "ground_truth": ground_truth_str, - "prediction": prediction, - "category": category, - "metrics": { - "f1": f1_val, - "bleu1": bleu1_val, - "jaccard": jaccard_val, - "locomo_f1": locomo_f1_val - }, - "retrieval": { - "num_docs": len(retrieved_info), - "context_length": len(context_text) - }, - "context_tokens": len(context_text.split()), - "timing": { - "search_ms": search_latency, - "llm_ms": llm_latency - } - }) - - return samples - - -# ============================================================================ -# Step 5: Aggregate Results -# ============================================================================ - -def step_aggregate_results(samples: List[Dict[str, Any]]) -> Dict[str, Any]: - """Aggregate metrics from all samples.""" - print(f"\n{'='*60}") - print("📊 Aggregating Results") - print(f"{'='*60}\n") - - if not samples: - return { - "overall_metrics": {}, - "by_category": {}, - "latency": {}, - "context_stats": {} - } - - # Extract metrics - f1_scores = [s["metrics"]["f1"] for s in samples] - bleu1_scores = [s["metrics"]["bleu1"] for s in samples] - jaccard_scores = [s["metrics"]["jaccard"] for s in samples] - locomo_f1_scores = [s["metrics"]["locomo_f1"] for s in samples] - - # Extract timing - latencies_search = [s["timing"]["search_ms"] for s in samples] - latencies_llm = [s["timing"]["llm_ms"] for s in samples] - - # Extract context stats - context_counts = [s["retrieval"]["num_docs"] for s in samples] - context_chars = [s["retrieval"]["context_length"] for s in samples] - context_tokens = [s["context_tokens"] for s in samples] - - # Overall metrics - overall_metrics = { - "f1": sum(f1_scores) / len(f1_scores) if f1_scores else 0.0, - "bleu1": sum(bleu1_scores) / len(bleu1_scores) if bleu1_scores else 0.0, - "jaccard": sum(jaccard_scores) / len(jaccard_scores) if jaccard_scores else 0.0, - "locomo_f1": sum(locomo_f1_scores) / len(locomo_f1_scores) if locomo_f1_scores else 0.0 - } - - # Per-category metrics - category_data: Dict[str, Dict[str, List[float]]] = {} - for sample in samples: - cat = sample["category"] - if cat not in category_data: - category_data[cat] = { - "f1": [], - "bleu1": [], - "jaccard": [], - "locomo_f1": [] - } - category_data[cat]["f1"].append(sample["metrics"]["f1"]) - category_data[cat]["bleu1"].append(sample["metrics"]["bleu1"]) - category_data[cat]["jaccard"].append(sample["metrics"]["jaccard"]) - category_data[cat]["locomo_f1"].append(sample["metrics"]["locomo_f1"]) - - by_category: Dict[str, Dict[str, Any]] = {} - for cat, metrics_lists in category_data.items(): - by_category[cat] = { - "count": len(metrics_lists["f1"]), - "f1": sum(metrics_lists["f1"]) / len(metrics_lists["f1"]), - "bleu1": sum(metrics_lists["bleu1"]) / len(metrics_lists["bleu1"]), - "jaccard": sum(metrics_lists["jaccard"]) / len(metrics_lists["jaccard"]), - "locomo_f1": sum(metrics_lists["locomo_f1"]) / len(metrics_lists["locomo_f1"]) - } - - # Latency statistics - latency = { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm) - } - - # Context statistics - context_stats = { - "avg_retrieved_docs": sum(context_counts) / len(context_counts) if context_counts else 0.0, - "avg_context_chars": sum(context_chars) / len(context_chars) if context_chars else 0.0, - "avg_context_tokens": sum(context_tokens) / len(context_tokens) if context_tokens else 0.0 - } - - return { - "overall_metrics": overall_metrics, - "by_category": by_category, - "latency": latency, - "context_stats": context_stats - } - - -# ============================================================================ -# Step 6: Result Saving -# ============================================================================ - -def step_save_results( - result: Dict[str, Any], - output_dir: Optional[str] -) -> str: - """ - Save evaluation results to JSON file. - - Args: - result: Complete result dictionary - output_dir: Directory to save results (uses default if None) - - Returns: - Path to saved file - """ - if output_dir is None: - # Use absolute path to ensure results are saved in the correct location - script_dir = Path(__file__).resolve().parent - output_dir = script_dir / "results" - else: - # Convert to Path object - output_dir = Path(output_dir) - # If relative path, make it relative to script directory - if not output_dir.is_absolute(): - script_dir = Path(__file__).resolve().parent - output_dir = script_dir / output_dir - - # Create directory if it doesn't exist - output_dir.mkdir(parents=True, exist_ok=True) - - timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S") - output_path = output_dir / f"locomo_{timestamp_str}.json" - - try: - with open(output_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"✅ Results saved to: {output_path}\n") - return str(output_path) - except Exception as e: - print(f"❌ Failed to save results: {e}") - print("📊 Printing results to console instead:\n") - print(json.dumps(result, ensure_ascii=False, indent=2)) - return "" - - -# ============================================================================ -# Main Orchestration Function -# ============================================================================ - - -async def run_locomo_benchmark( - sample_size: int = 20, - end_user_id: Optional[str] = None, - search_type: str = "hybrid", - search_limit: int = 12, - context_char_budget: int = 8000, - reset_group: bool = False, - skip_ingest: bool = False, - output_dir: Optional[str] = None, - max_ingest_messages: Optional[int] = None -) -> Dict[str, Any]: - """ - Run LoCoMo benchmark evaluation. - - This function orchestrates the complete evaluation pipeline by calling - well-defined step functions: - 1. Load LoCoMo dataset (only QA pairs from first conversation) - 2. Ingest conversations into database (unless skip_ingest=True) - 3. Initialize clients (Neo4j, LLM, Embedder) - 4. Process all questions (retrieve, generate, calculate metrics) - 5. Aggregate results - 6. Save results to file - - Note: By default, only the first conversation is ingested into the database, - and only QA pairs from that conversation are evaluated. This ensures that - all questions have corresponding memory in the database for retrieval. - - Args: - sample_size: Number of QA pairs to evaluate (from first conversation) - end_user_id: Database end_user ID for retrieval (uses default if None) - search_type: "keyword", "embedding", or "hybrid" - search_limit: Max documents to retrieve per query - context_char_budget: Max characters for context - reset_group: Whether to clear and re-ingest data - skip_ingest: If True, skip data ingestion and use existing data in Neo4j - output_dir: Directory to save results (uses default if None) - max_ingest_messages: Max messages per dialogue to ingest (for testing, None = all) - - Returns: - Dictionary with evaluation results including metrics, timing, and samples - """ - # Use default end_user_id if not provided - # 优先级：命令行参数 > LOCOMO_END_USER_ID > EVAL_END_USER_ID > 默认值 - if end_user_id is None: - end_user_id = os.getenv("LOCOMO_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "locomo_benchmark") - - # Get model IDs from config - llm_id = os.getenv("EVAL_LLM_ID", "6dc52e1b-9cec-4194-af66-a74c6307fc3f") - embedding_id = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") - - # Determine data path - dataset_dir = Path(__file__).resolve().parent.parent / "dataset" - data_path = dataset_dir / "locomo10.json" - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 locomo10.json 放置在: {dataset_dir}" - ) - - # Print configuration - print(f"\n{'='*60}") - print("🚀 Starting LoCoMo Benchmark Evaluation") - print(f"{'='*60}") - print("📊 Configuration:") - print(f" Sample size: {sample_size}") - print(f" End User ID: {end_user_id}") - print(f" Search type: {search_type}") - print(f" Search limit: {search_limit}") - print(f" Context budget: {context_char_budget} chars") - print(f" Data path: {data_path}") - if max_ingest_messages: - print(f" Max ingest messages: {max_ingest_messages} (testing mode)") - print(f"{'='*60}\n") - - # Step 1: Load LoCoMo data （加载数据） - try: - qa_items = step_load_data(data_path, sample_size) - except Exception as e: - print(f"❌ Failed to load data: {e}") - return { - "error": f"Data loading failed: {e}", - "timestamp": datetime.now().isoformat() - } - - # Step 2: Ingest data if needed（数据摄入） - await step_ingest_data(data_path, end_user_id, skip_ingest, reset_group, max_ingest_messages) - - # Step 3: Initialize clients （初始化客户端） - connector, llm_client, embedder = step_initialize_clients(llm_id, embedding_id) - - # Step 4: Process all questions （处理所有问题） - try: - samples = await step_process_all_questions( - qa_items=qa_items, - end_user_id=end_user_id, - search_type=search_type, - search_limit=search_limit, - context_char_budget=context_char_budget, - connector=connector, - embedder=embedder, - llm_client=llm_client - ) - finally: - await connector.close() - - # Step 5: Aggregate results （聚合答案） - aggregated = step_aggregate_results(samples) - - # Build final result dictionary - result = { - "dataset": "locomo", - "sample_size": len(qa_items), - "timestamp": datetime.now().isoformat(), - "params": { - "end_user_id": end_user_id, - "search_type": search_type, - "search_limit": search_limit, - "context_char_budget": context_char_budget, - "llm_id": llm_id, - "embedding_id": embedding_id - }, - "overall_metrics": aggregated["overall_metrics"], - "by_category": aggregated["by_category"], - "latency": aggregated["latency"], - "context_stats": aggregated["context_stats"], - "samples": samples - } - - # Step 6: Save results （保存结果） - step_save_results(result, output_dir) - - return result - - -def main(): - """ - Parse command-line arguments and run benchmark. - - This function provides a CLI interface for running LoCoMo benchmarks - with configurable parameters. - - Configuration priority: Command-line args > Environment variables > Code defaults - """ - # Load environment variables first - load_dotenv() - - # Get defaults from environment variables - env_sample_size = os.getenv("LOCOMO_SAMPLE_SIZE") - env_search_limit = os.getenv("LOCOMO_SEARCH_LIMIT") - env_context_budget = os.getenv("LOCOMO_CONTEXT_CHAR_BUDGET") - env_output_dir = os.getenv("LOCOMO_OUTPUT_DIR") - env_skip_ingest = os.getenv("LOCOMO_SKIP_INGEST", "false").lower() in ("true", "1", "yes") - - # Convert to appropriate types with fallback to code defaults - default_sample_size = int(env_sample_size) if env_sample_size else 20 - default_search_limit = int(env_search_limit) if env_search_limit else 12 - default_context_budget = int(env_context_budget) if env_context_budget else 8000 - default_output_dir = env_output_dir if env_output_dir else None - - parser = argparse.ArgumentParser( - description="Run LoCoMo benchmark evaluation", - formatter_class=argparse.ArgumentDefaultsHelpFormatter - ) - - parser.add_argument( - "--sample_size", - type=int, - default=default_sample_size, - help=f"Number of QA pairs to evaluate (env: LOCOMO_SAMPLE_SIZE={env_sample_size or 'not set'}, 0 for all)" - ) - parser.add_argument( - "--end_user_id", - type=str, - default=None, - help="Database end user ID for retrieval (uses LOCOMO_END_USER_ID or EVAL_END_USER_ID if not specified)" - ) - parser.add_argument( - "--search_type", - type=str, - default="hybrid", - choices=["keyword", "embedding", "hybrid"], - help="Search strategy to use" - ) - parser.add_argument( - "--search_limit", - type=int, - default=default_search_limit, - help=f"Maximum number of documents to retrieve per query (env: LOCOMO_SEARCH_LIMIT={env_search_limit or 'not set'})" - ) - parser.add_argument( - "--context_char_budget", - type=int, - default=default_context_budget, - help=f"Maximum characters for context (env: LOCOMO_CONTEXT_CHAR_BUDGET={env_context_budget or 'not set'})" - ) - parser.add_argument( - "--reset_group", - action="store_true", - help="Clear and re-ingest data (not implemented)" - ) - parser.add_argument( - "--skip_ingest", - action="store_true", - default=env_skip_ingest, - help=f"Skip data ingestion and use existing data in Neo4j (env: LOCOMO_SKIP_INGEST={os.getenv('LOCOMO_SKIP_INGEST', 'false')})" - ) - parser.add_argument( - "--output_dir", - type=str, - default=default_output_dir, - help=f"Directory to save results (env: LOCOMO_OUTPUT_DIR={env_output_dir or 'not set'})" - ) - parser.add_argument( - "--max_ingest_messages", - type=int, - default=None, - help="Maximum messages per dialogue to ingest (for testing, default: all messages)" - ) - - args = parser.parse_args() - - # Run benchmark - result = asyncio.run(run_locomo_benchmark( - sample_size=args.sample_size, - end_user_id=args.end_user_id, - search_type=args.search_type, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - reset_group=args.reset_group, - skip_ingest=args.skip_ingest, - output_dir=args.output_dir, - max_ingest_messages=args.max_ingest_messages - )) - - # Print summary - print(f"\n{'='*60}") - - # Check if there was an error - if 'error' in result: - print("❌ Benchmark Failed!") - print(f"{'='*60}") - print(f"Error: {result['error']}") - return - - print("🎉 Benchmark Complete!") - print(f"{'='*60}") - print("📊 Final Results:") - print(f" Sample size: {result.get('sample_size', 0)}") - print(f" F1: {result['overall_metrics']['f1']:.3f}") - print(f" BLEU-1: {result['overall_metrics']['bleu1']:.3f}") - print(f" Jaccard: {result['overall_metrics']['jaccard']:.3f}") - print(f" LoCoMo F1: {result['overall_metrics']['locomo_f1']:.3f}") - - if result.get('context_stats'): - print("\n📈 Context Statistics:") - print(f" Avg retrieved docs: {result['context_stats']['avg_retrieved_docs']:.1f}") - print(f" Avg context chars: {result['context_stats']['avg_context_chars']:.0f}") - print(f" Avg context tokens: {result['context_stats']['avg_context_tokens']:.0f}") - - if result.get('latency'): - print("\n⏱️ Latency Statistics:") - print(f" Search - Mean: {result['latency']['search']['mean']:.1f}ms, " - f"P50: {result['latency']['search']['p50']:.1f}ms, " - f"P95: {result['latency']['search']['p95']:.1f}ms") - print(f" LLM - Mean: {result['latency']['llm']['mean']:.1f}ms, " - f"P50: {result['latency']['llm']['p50']:.1f}ms, " - f"P95: {result['latency']['llm']['p95']:.1f}ms") - - if result.get('by_category'): - print("\n📂 Results by Category:") - for cat, metrics in result['by_category'].items(): - print(f" {cat}:") - print(f" Count: {metrics['count']}") - print(f" F1: {metrics['f1']:.3f}") - print(f" LoCoMo F1: {metrics['locomo_f1']:.3f}") - print(f" Jaccard: {metrics['jaccard']:.3f}") - - print(f"\n{'='*60}\n") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/locomo/locomo_metrics.py b/api/app/core/memory/evaluation/locomo/locomo_metrics.py deleted file mode 100644 index 20d5f2b5..00000000 --- a/api/app/core/memory/evaluation/locomo/locomo_metrics.py +++ /dev/null @@ -1,225 +0,0 @@ -""" -LoCoMo-specific metric calculations. - -This module provides clean, simplified implementations of metrics used for -LoCoMo benchmark evaluation, including text normalization and F1 score variants. -""" - -import re -from typing import Dict, Any - - -def normalize_text(text: str) -> str: - """ - Normalize text for LoCoMo evaluation. - - Normalization steps: - - Convert to lowercase - - Remove commas - - Remove stop words (a, an, the, and) - - Remove punctuation - - Normalize whitespace - - Args: - text: Input text to normalize - - Returns: - Normalized text string with consistent formatting - - Examples: - >>> normalize_text("The cat, and the dog") - 'cat dog' - >>> normalize_text("Hello, World!") - 'hello world' - """ - # Ensure input is a string - text = str(text) if text is not None else "" - - # Convert to lowercase - text = text.lower() - - # Remove commas - text = re.sub(r"[\,]", " ", text) - - # Remove stop words - text = re.sub(r"\b(a|an|the|and)\b", " ", text) - - # Remove punctuation (keep only word characters and whitespace) - text = re.sub(r"[^\w\s]", " ", text) - - # Normalize whitespace (collapse multiple spaces to single space) - text = " ".join(text.split()) - - return text - - -def locomo_f1_score(prediction: str, ground_truth: str) -> float: - """ - Calculate LoCoMo F1 score for single-answer questions. - - Uses token-level precision and recall based on normalized text. - Treats tokens as sets (no duplicate counting). - - Args: - prediction: Model's predicted answer - ground_truth: Correct answer - - Returns: - F1 score between 0.0 and 1.0 - - Examples: - >>> locomo_f1_score("Paris", "Paris") - 1.0 - >>> locomo_f1_score("The cat", "cat") - 1.0 - >>> locomo_f1_score("dog", "cat") - 0.0 - """ - # Ensure inputs are strings - pred_str = str(prediction) if prediction is not None else "" - truth_str = str(ground_truth) if ground_truth is not None else "" - - # Normalize and tokenize - pred_tokens = normalize_text(pred_str).split() - truth_tokens = normalize_text(truth_str).split() - - # Handle empty cases - if not pred_tokens or not truth_tokens: - return 0.0 - - # Convert to sets for comparison - pred_set = set(pred_tokens) - truth_set = set(truth_tokens) - - # Calculate true positives (intersection) - true_positives = len(pred_set & truth_set) - - # Calculate precision and recall - precision = true_positives / len(pred_set) if pred_set else 0.0 - recall = true_positives / len(truth_set) if truth_set else 0.0 - - # Calculate F1 score - if precision + recall == 0: - return 0.0 - - f1 = 2 * precision * recall / (precision + recall) - return f1 - - -def locomo_multi_f1(prediction: str, ground_truth: str) -> float: - """ - Calculate LoCoMo F1 score for multi-answer questions. - - Handles comma-separated answers by: - 1. Splitting both prediction and ground truth by commas - 2. For each ground truth answer, finding the best matching prediction - 3. Averaging the F1 scores across all ground truth answers - - Args: - prediction: Model's predicted answer (may contain multiple comma-separated answers) - ground_truth: Correct answer (may contain multiple comma-separated answers) - - Returns: - Average F1 score across all ground truth answers (0.0 to 1.0) - - Examples: - >>> locomo_multi_f1("Paris, London", "Paris, London") - 1.0 - >>> locomo_multi_f1("Paris", "Paris, London") - 0.5 - >>> locomo_multi_f1("Paris, Berlin", "Paris, London") - 0.5 - """ - # Ensure inputs are strings - pred_str = str(prediction) if prediction is not None else "" - truth_str = str(ground_truth) if ground_truth is not None else "" - - # Split by commas and strip whitespace - predictions = [p.strip() for p in pred_str.split(',') if p.strip()] - ground_truths = [g.strip() for g in truth_str.split(',') if g.strip()] - - # Handle empty cases - if not predictions or not ground_truths: - return 0.0 - - # For each ground truth, find the best matching prediction - f1_scores = [] - for gt in ground_truths: - # Calculate F1 with each prediction and take the maximum - best_f1 = max(locomo_f1_score(pred, gt) for pred in predictions) - f1_scores.append(best_f1) - - # Return average F1 across all ground truths - return sum(f1_scores) / len(f1_scores) - - -def get_category_name(item: Dict[str, Any]) -> str: - """ - Extract and normalize category name from QA item. - - Handles both numeric categories (1-4) and string categories with various formats. - Supports multiple field names: "cat", "category", "type". - - Category mapping: - - 1 or "multi-hop" -> "Multi-Hop" - - 2 or "temporal" -> "Temporal" - - 3 or "open domain" -> "Open Domain" - - 4 or "single-hop" -> "Single-Hop" - - Args: - item: QA item dictionary containing category information - - Returns: - Standardized category name or "unknown" if not found - - Examples: - >>> get_category_name({"category": 1}) - 'Multi-Hop' - >>> get_category_name({"cat": "temporal"}) - 'Temporal' - >>> get_category_name({"type": "Single-Hop"}) - 'Single-Hop' - """ - # Numeric category mapping - CATEGORY_MAP = { - 1: "Multi-Hop", - 2: "Temporal", - 3: "Open Domain", - 4: "Single-Hop", - } - - # String category aliases (case-insensitive) - TYPE_ALIASES = { - "single-hop": "Single-Hop", - "singlehop": "Single-Hop", - "single hop": "Single-Hop", - "multi-hop": "Multi-Hop", - "multihop": "Multi-Hop", - "multi hop": "Multi-Hop", - "open domain": "Open Domain", - "opendomain": "Open Domain", - "temporal": "Temporal", - } - - # Try "cat" field first (string category) - cat = item.get("cat") - if isinstance(cat, str) and cat.strip(): - name = cat.strip() - lower = name.lower() - return TYPE_ALIASES.get(lower, name) - - # Try "category" field (can be int or string) - cat_num = item.get("category") - if isinstance(cat_num, int): - return CATEGORY_MAP.get(cat_num, "unknown") - elif isinstance(cat_num, str) and cat_num.strip(): - lower = cat_num.strip().lower() - return TYPE_ALIASES.get(lower, cat_num.strip()) - - # Try "type" field as fallback - cat_type = item.get("type") - if isinstance(cat_type, str) and cat_type.strip(): - lower = cat_type.strip().lower() - return TYPE_ALIASES.get(lower, cat_type.strip()) - - return "unknown" diff --git a/api/app/core/memory/evaluation/locomo/locomo_test.py b/api/app/core/memory/evaluation/locomo/locomo_test.py deleted file mode 100644 index 2cb0664c..00000000 --- a/api/app/core/memory/evaluation/locomo/locomo_test.py +++ /dev/null @@ -1,864 +0,0 @@ -# file name: check_neo4j_connection_fixed.py -import asyncio -import os -import sys -import json -import time -import math -import re -from datetime import datetime, timedelta -from typing import List, Dict, Any -from pathlib import Path -from dotenv import load_dotenv - -# Load main .env -load_dotenv() - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -# Get group_id from config -group_id = os.getenv("EVAL_GROUP_ID", "locomo_test") -print(f"✅ 使用配置的 group_id: {group_id}") - -# 首先定义 _loc_normalize 函数，因为其他函数依赖它 -def _loc_normalize(text: str) -> str: - text = str(text) if text is not None else "" - text = text.lower() - text = re.sub(r"[\,]", " ", text) - text = re.sub(r"\b(a|an|the|and)\b", " ", text) - text = re.sub(r"[^\w\s]", " ", text) - text = " ".join(text.split()) - return text - -# 尝试从 metrics.py 导入基础指标 -try: - from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard - print("✅ 从 metrics.py 导入基础指标成功") -except ImportError as e: - print(f"❌ 从 metrics.py 导入失败: {e}") - # 回退到本地实现 - def f1_score(pred: str, ref: str) -> float: - pred_str = str(pred) if pred is not None else "" - ref_str = str(ref) if ref is not None else "" - - p_tokens = _loc_normalize(pred_str).split() - r_tokens = _loc_normalize(ref_str).split() - if not p_tokens and not r_tokens: - return 1.0 - if not p_tokens or not r_tokens: - return 0.0 - p_set = set(p_tokens) - r_set = set(r_tokens) - tp = len(p_set & r_set) - precision = tp / len(p_set) if p_set else 0.0 - recall = tp / len(r_set) if r_set else 0.0 - if precision + recall == 0: - return 0.0 - return 2 * precision * recall / (precision + recall) - - def bleu1(pred: str, ref: str) -> float: - pred_str = str(pred) if pred is not None else "" - ref_str = str(ref) if ref is not None else "" - - p_tokens = _loc_normalize(pred_str).split() - r_tokens = _loc_normalize(ref_str).split() - if not p_tokens: - return 0.0 - - r_counts = {} - for t in r_tokens: - r_counts[t] = r_counts.get(t, 0) + 1 - - clipped = 0 - p_counts = {} - for t in p_tokens: - p_counts[t] = p_counts.get(t, 0) + 1 - - for t, c in p_counts.items(): - clipped += min(c, r_counts.get(t, 0)) - - precision = clipped / max(len(p_tokens), 1) - ref_len = len(r_tokens) - pred_len = len(p_tokens) - - if pred_len > ref_len or pred_len == 0: - bp = 1.0 - else: - bp = math.exp(1 - ref_len / max(pred_len, 1)) - - return bp * precision - - def jaccard(pred: str, ref: str) -> float: - pred_str = str(pred) if pred is not None else "" - ref_str = str(ref) if ref is not None else "" - - p = set(_loc_normalize(pred_str).split()) - r = set(_loc_normalize(ref_str).split()) - if not p and not r: - return 1.0 - if not p or not r: - return 0.0 - return len(p & r) / len(p | r) - -# 尝试从 qwen_search_eval.py 导入 LoCoMo 特定指标 -try: - from app.core.memory.evaluation.locomo.qwen_search_eval import loc_f1_score, loc_multi_f1, _resolve_relative_times - print("✅ 从 qwen_search_eval 导入 LoCoMo 特定指标成功") -except ImportError as e: - print(f"❌ 从 qwen_search_eval.py 导入失败: {e}") - # 回退到本地实现 LoCoMo 特定函数 - def _resolve_relative_times(text: str, anchor: datetime) -> str: - t = str(text) if text is not None else "" - t = re.sub(r"\btoday\b", anchor.date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\byesterday\b", (anchor - timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\btomorrow\b", (anchor + timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - - def _ago_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor - timedelta(days=n)).date().isoformat() - def _in_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor + timedelta(days=n)).date().isoformat() - - t = re.sub(r"\b(\d+)\s+days\s+ago\b", _ago_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\bin\s+(\d+)\s+days\b", _in_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\blast\s+week\b", (anchor - timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\bnext\s+week\b", (anchor + timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - return t - - def loc_f1_score(prediction: str, ground_truth: str) -> float: - p_tokens = _loc_normalize(prediction).split() - g_tokens = _loc_normalize(ground_truth).split() - if not p_tokens or not g_tokens: - return 0.0 - p = set(p_tokens) - g = set(g_tokens) - tp = len(p & g) - precision = tp / len(p) if p else 0.0 - recall = tp / len(g) if g else 0.0 - return (2 * precision * recall / (precision + recall)) if (precision + recall) > 0 else 0.0 - - def loc_multi_f1(prediction: str, ground_truth: str) -> float: - predictions = [p.strip() for p in str(prediction).split(',') if p.strip()] - ground_truths = [g.strip() for g in str(ground_truth).split(',') if g.strip()] - if not predictions or not ground_truths: - return 0.0 - def _f1(a: str, b: str) -> float: - return loc_f1_score(a, b) - vals = [] - for gt in ground_truths: - vals.append(max(_f1(pred, gt) for pred in predictions)) - return sum(vals) / len(vals) - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 8000) -> str: - """基于问题关键词智能选择上下文""" - if not contexts: - return "" - - # 提取问题关键词（只保留有意义的词） - question_lower = question.lower() - stop_words = {'what', 'when', 'where', 'who', 'why', 'how', 'did', 'do', 'does', 'is', 'are', 'was', 'were', 'the', 'a', 'an', 'and', 'or', 'but'} - question_words = set(re.findall(r'\b\w+\b', question_lower)) - question_words = {word for word in question_words if word not in stop_words and len(word) > 2} - - print(f"🔍 问题关键词: {question_words}") - - # 给每个上下文打分 - scored_contexts = [] - for i, context in enumerate(contexts): - context_lower = context.lower() - score = 0 - - # 关键词匹配得分 - keyword_matches = 0 - for word in question_words: - if word in context_lower: - keyword_matches += 1 - # 关键词出现次数越多，得分越高 - score += context_lower.count(word) * 2 - - # 上下文长度得分（适中的长度更好） - context_len = len(context) - if 100 < context_len < 2000: # 理想长度范围 - score += 5 - elif context_len >= 2000: # 太长可能包含无关信息 - score += 2 - - # 如果是前几个上下文，给予额外分数（通常相关性更高） - if i < 3: - score += 3 - - scored_contexts.append((score, context, keyword_matches)) - - # 按得分排序 - scored_contexts.sort(key=lambda x: x[0], reverse=True) - - # 选择高得分的上下文，直到达到字符限制 - selected = [] - total_chars = 0 - selected_count = 0 - - print("📊 上下文相关性分析:") - for score, context, matches in scored_contexts[:5]: # 只显示前5个 - print(f" - 得分: {score}, 关键词匹配: {matches}, 长度: {len(context)}") - - for score, context, matches in scored_contexts: - if total_chars + len(context) <= max_chars: - selected.append(context) - total_chars += len(context) - selected_count += 1 - else: - # 如果这个上下文得分很高但放不下，尝试截取 - if score > 10 and total_chars < max_chars - 500: - remaining = max_chars - total_chars - # 找到包含关键词的部分 - lines = context.split('\n') - relevant_lines = [] - current_chars = 0 - - for line in lines: - line_lower = line.lower() - line_relevance = any(word in line_lower for word in question_words) - - if line_relevance and current_chars < remaining - 100: - relevant_lines.append(line) - current_chars += len(line) - - if relevant_lines: - truncated = '\n'.join(relevant_lines) - if len(truncated) > 100: # 确保有足够内容 - selected.append(truncated + "\n[相关内容截断...]") - total_chars += len(truncated) - selected_count += 1 - break # 不再尝试添加更多上下文 - - result = "\n\n".join(selected) - print(f"✅ 智能选择: {selected_count}个上下文, 总长度: {total_chars}字符") - return result - - -def get_dynamic_search_params(question: str, question_index: int, total_questions: int): - """根据问题复杂度和进度动态调整检索参数""" - - # 分析问题复杂度 - word_count = len(question.split()) - has_temporal = any(word in question.lower() for word in ['when', 'date', 'time', 'ago']) - has_multi_hop = any(word in question.lower() for word in ['and', 'both', 'also', 'while']) - - # 根据进度调整 - 后期问题可能需要更精确的检索 - progress_factor = question_index / total_questions - - base_limit = 12 - if has_temporal and has_multi_hop: - base_limit = 20 - elif word_count > 8: - base_limit = 16 - - # 随着测试进行，逐渐收紧检索范围 - adjusted_limit = max(8, int(base_limit * (1 - progress_factor * 0.3))) - - # 动态调整最大字符数 - max_chars = 8000 + 4000 * (1 - progress_factor) - - return { - "limit": adjusted_limit, - "max_chars": int(max_chars) - } - - -class EnhancedEvaluationMonitor: - def __init__(self, reset_interval=5, performance_threshold=0.6): - self.question_count = 0 - self.reset_interval = reset_interval - self.performance_threshold = performance_threshold - self.consecutive_low_scores = 0 - self.performance_history = [] - self.recent_f1_scores = [] - - def should_reset_connections(self, current_f1=None): - """基于计数和性能双重判断""" - # 定期重置 - if self.question_count % self.reset_interval == 0: - return True - - # 性能驱动的重置 - if current_f1 is not None and current_f1 < self.performance_threshold: - self.consecutive_low_scores += 1 - if self.consecutive_low_scores >= 2: # 连续2个低分就重置 - print("🚨 连续低分，触发紧急重置") - self.consecutive_low_scores = 0 - return True - else: - self.consecutive_low_scores = 0 - - return False - - def record_performance(self, question_index, metrics, context_length, retrieved_docs): - """记录性能指标，检测衰减""" - self.performance_history.append({ - 'index': question_index, - 'metrics': metrics, - 'context_length': context_length, - 'retrieved_docs': retrieved_docs, - 'timestamp': time.time() - }) - - # 记录最近的F1分数 - self.recent_f1_scores.append(metrics['f1']) - if len(self.recent_f1_scores) > 5: - self.recent_f1_scores.pop(0) - - def get_recent_performance(self): - """获取近期平均性能""" - if not self.recent_f1_scores: - return 0.5 - return sum(self.recent_f1_scores) / len(self.recent_f1_scores) - - def get_performance_trend(self): - """分析性能趋势""" - if len(self.performance_history) < 2: - return "stable" - - recent_metrics = [item['metrics']['f1'] for item in self.performance_history[-5:]] - earlier_metrics = [item['metrics']['f1'] for item in self.performance_history[-10:-5]] - - if len(recent_metrics) < 2 or len(earlier_metrics) < 2: - return "stable" - - recent_avg = sum(recent_metrics) / len(recent_metrics) - earlier_avg = sum(earlier_metrics) / len(earlier_metrics) - - if recent_avg < earlier_avg * 0.8: - return "degrading" - elif recent_avg > earlier_avg * 1.1: - return "improving" - else: - return "stable" - - -def get_enhanced_search_params(question: str, question_index: int, total_questions: int, recent_performance: float): - """基于问题复杂度和近期性能动态调整检索参数""" - - # 基础参数 - base_params = get_dynamic_search_params(question, question_index, total_questions) - - # 性能自适应调整 - if recent_performance < 0.5: # 近期表现差 - # 增加检索范围，尝试获取更多上下文 - base_params["limit"] = min(base_params["limit"] + 5, 25) - base_params["max_chars"] = min(base_params["max_chars"] + 2000, 12000) - print(f"📈 性能自适应：增加检索范围 (limit={base_params['limit']}, max_chars={base_params['max_chars']})") - - elif recent_performance > 0.8: # 近期表现好 - # 收紧检索，提高精度 - base_params["limit"] = max(base_params["limit"] - 2, 8) - base_params["max_chars"] = max(base_params["max_chars"] - 1000, 6000) - print(f"🎯 性能自适应：提高检索精度 (limit={base_params['limit']}, max_chars={base_params['max_chars']})") - - # 中间阶段特殊处理 - mid_sequence_factor = abs(question_index / total_questions - 0.5) - if mid_sequence_factor < 0.2: # 在中间30%的问题 - print("🎯 中间阶段：使用更精确的检索策略") - base_params["limit"] = max(base_params["limit"] - 2, 10) # 减少数量，提高质量 - base_params["max_chars"] = max(base_params["max_chars"] - 1000, 7000) - - return base_params - - -def enhanced_context_selection(contexts: List[str], question: str, question_index: int, total_questions: int, max_chars: int = 8000) -> str: - """考虑问题序列位置的智能选择""" - - if not contexts: - return "" - - # 在序列中间阶段使用更严格的筛选 - mid_sequence_factor = abs(question_index / total_questions - 0.5) # 距离中心的距离 - - if mid_sequence_factor < 0.2: # 在中间30%的问题 - print("🎯 中间阶段：使用严格上下文筛选") - - # 提取问题关键词 - question_lower = question.lower() - stop_words = {'what', 'when', 'where', 'who', 'why', 'how', 'did', 'do', 'does', 'is', 'are', 'was', 'were', 'the', 'a', 'an', 'and', 'or', 'but'} - question_words = set(re.findall(r'\b\w+\b', question_lower)) - question_words = {word for word in question_words if word not in stop_words and len(word) > 2} - - # 只保留高度相关的上下文 - filtered_contexts = [] - for context in contexts: - context_lower = context.lower() - relevance_score = sum(3 if word in context_lower else 0 for word in question_words) - - # 额外加分给包含数字、日期的上下文（对事实性问题更重要） - if any(char.isdigit() for char in context): - relevance_score += 2 - - # 提高阈值：只有得分>=3的上下文才保留 - if relevance_score >= 3: - filtered_contexts.append(context) - else: - print(f" - 过滤低分上下文: 得分={relevance_score}") - - contexts = filtered_contexts - print(f"🔍 严格筛选后保留 {len(contexts)} 个上下文") - - # 使用原有的智能选择逻辑 - return smart_context_selection(contexts, question, max_chars) - - -async def run_enhanced_evaluation(): - """使用增强方法进行完整评估 - 解决中间性能衰减问题""" - from dotenv import load_dotenv - from uuid import UUID - from datetime import datetime - from dataclasses import dataclass - - # 修正导入路径：使用 app.core.memory.src 前缀 - from app.repositories.neo4j.neo4j_connector import Neo4jConnector - from app.repositories.neo4j.graph_search import search_graph_by_embedding - from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient - from app.core.models.base import RedBearModelConfig - from app.core.memory.utils.llm.llm_utils import get_llm_client - from app.core.memory.utils.config.config_utils import get_embedder_config - from app.schemas.memory_config_schema import MemoryConfig - from app.services.memory_config_service import MemoryConfigService - - # Get model IDs from config - llm_id = os.getenv("EVAL_LLM_ID", "6dc52e1b-9cec-4194-af66-a74c6307fc3f") - embedding_id = os.getenv("EVAL_EMBEDDING_ID", "e2a6392d-ca63-4d59-a523-647420b59cb2") - - # 加载数据 - 使用统一的 dataset 目录 - data_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "dataset", "locomo10.json") - - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 locomo10.json 放置在: api/app/core/memory/evaluation/dataset/" - ) - - print(f"✅ 找到数据文件: {data_path}") - - with open(data_path, "r", encoding="utf-8") as f: - raw = json.load(f) - - qa_items = [] - if isinstance(raw, list): - for entry in raw: - qa_items.extend(entry.get("qa", [])) - else: - qa_items.extend(raw.get("qa", [])) - - # 测试多少个问题 - 可通过环境变量设置 - sample_size = int(os.getenv("LOCOMO_SAMPLE_SIZE", "20")) - items = qa_items[:sample_size] - print(f"📊 将测试 {len(items)} 个问题（总共 {len(qa_items)} 个可用）") - - # 初始化增强监控器 - monitor = EnhancedEvaluationMonitor(reset_interval=5, performance_threshold=0.6) - - # 获取数据库会话并初始化 LLM 客户端 - from app.db import get_db - db = next(get_db()) - - try: - llm = get_llm_client(llm_id, db) - - # 初始化embedder - cfg_dict = get_embedder_config(embedding_id, db) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - - # 🔧 创建 MemoryConfig 对象用于搜索 - # 方案1：如果有配置ID，从数据库加载 - config_id = os.getenv("EVAL_CONFIG_ID") - if config_id: - print(f"📋 从数据库加载配置 ID: {config_id}") - memory_config_service = MemoryConfigService(db) - memory_config = memory_config_service.load_memory_config(config_id, service_name="locomo_test") - else: - # 方案2：创建临时配置对象用于测试 - print(f"📋 创建临时测试配置") - from uuid import UUID - from datetime import datetime - - # 将字符串 ID 转换为 UUID - try: - embedding_uuid = UUID(embedding_id) - llm_uuid = UUID(llm_id) - except ValueError as e: - raise ValueError(f"无效的 UUID 格式: {e}") - - memory_config = MemoryConfig( - config_id=1, # 临时 ID - config_name="locomo_test_config", - workspace_id=UUID("00000000-0000-0000-0000-000000000000"), # 临时 workspace - workspace_name="test_workspace", - tenant_id=UUID("00000000-0000-0000-0000-000000000000"), # 临时 tenant - embedding_model_id=embedding_uuid, - embedding_model_name="test_embedding", - llm_model_id=llm_uuid, - llm_model_name="test_llm", - storage_type="neo4j", - chunker_strategy="RecursiveChunker", - reflexion_enabled=False, - reflexion_iteration_period=3, - reflexion_range="partial", - reflexion_baseline="Time", - loaded_at=datetime.now() - ) - - print(f"✅ MemoryConfig 已准备: embedding_id={memory_config.embedding_model_id}, llm_id={memory_config.llm_model_id}") - - # 初始化连接器 - connector = Neo4jConnector() - - # 初始化结果字典 - results = { - "questions": [], - "overall_metrics": {"f1": 0.0, "b1": 0.0, "j": 0.0, "loc_f1": 0.0}, - "category_metrics": {}, - "retrieval_stats": {"total_questions": len(items), "avg_context_length": 0, "avg_retrieved_docs": 0}, - "performance_trend": "stable", - "timestamp": datetime.now().isoformat(), - "enhanced_strategy": True - } - - total_f1 = 0.0 - total_bleu1 = 0.0 - total_jaccard = 0.0 - total_loc_f1 = 0.0 - total_context_length = 0 - total_retrieved_docs = 0 - category_stats = {} - - try: - for i, item in enumerate(items): - monitor.question_count += 1 - - # 获取近期性能用于重置判断 - recent_performance = monitor.get_recent_performance() - - # 增强的重置判断 - should_reset = monitor.should_reset_connections(current_f1=recent_performance) - if should_reset and i > 0: - print(f"🔄 重置Neo4j连接 (问题 {i+1}/{len(items)}, 近期性能: {recent_performance:.3f})...") - await connector.close() - connector = Neo4jConnector() # 创建新连接 - print("✅ 连接重置完成") - - q = item.get("question", "") - ref = item.get("answer", "") - ref_str = str(ref) if ref is not None else "" - - print(f"\n🔍 [{i+1}/{len(items)}] 问题: {q}") - print(f"✅ 真实答案: {ref_str}") - - # 分类别统计 - category = "Unknown" - if item.get("category") == 1: - category = "Multi-Hop" - elif item.get("category") == 2: - category = "Temporal" - elif item.get("category") == 3: - category = "Open Domain" - elif item.get("category") == 4: - category = "Single-Hop" - - # 增强的检索参数 - search_params = get_enhanced_search_params(q, i, len(items), recent_performance) - search_limit = search_params["limit"] - max_chars = search_params["max_chars"] - - print(f"🏷️ 类别: {category}, 检索参数: limit={search_limit}, max_chars={max_chars}") - - # 使用项目标准的混合检索方法 - t0 = time.time() - contexts_all = [] - - try: - # 使用旧版本的搜索服务（重构前的版本） - from app.core.memory.src.search import run_hybrid_search - - print(f"🔀 使用混合搜索服务（旧版本）...") - print(f"📍 检索参数: group_id={group_id}, limit=20, search_type=hybrid") - print(f"📍 查询文本: {q}") - - search_results = await run_hybrid_search( - query_text=q, - search_type="hybrid", - end_user_id="locomo_sk", - limit=20, - include=["statements", "chunks", "entities", "summaries"], - output_path=None, - memory_config=memory_config, # 🔧 添加必需的 memory_config 参数 - rerank_alpha=0.6, # BM25权重 - use_forgetting_rerank=False, - use_llm_rerank=False - ) - - # 处理搜索结果 - 旧版本返回包含 reranked_results 的结构 - # 对于 hybrid 搜索，使用 reranked_results - if "reranked_results" in search_results: - reranked = search_results["reranked_results"] - chunks = reranked.get("chunks", []) - statements = reranked.get("statements", []) - entities = reranked.get("entities", []) - summaries = reranked.get("summaries", []) - else: - # 单一搜索类型的结果 - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - print(f"✅ 混合检索成功: {len(chunks)} chunks, {len(statements)} 条陈述, {len(entities)} 个实体, {len(summaries)} 个摘要") - - # 构建上下文：优先使用 chunks、statements 和 summaries - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # 实体摘要：最多加入前3个高分实体，避免噪声 - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + ' '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - print(f"📊 有效上下文数量: {len(contexts_all)}") - except Exception as e: - print(f"❌ 检索失败: {e}") - import traceback - print(f"详细错误信息:\n{traceback.format_exc()}") - contexts_all = [] - - t1 = time.time() - search_time = (t1 - t0) * 1000 - - # 增强的上下文选择 - context_text = "" - if contexts_all: - # 使用增强的上下文选择 - context_text = enhanced_context_selection(contexts_all, q, i, len(items), max_chars=max_chars) - - # 如果智能选择后仍然过长，进行最终保护性截断 - if len(context_text) > max_chars: - print(f"⚠️ 智能选择后仍然过长 ({len(context_text)}字符)，进行最终截断") - context_text = context_text[:max_chars] + "\n\n[最终截断...]" - - # 时间解析 - anchor_date = datetime(2023, 5, 8) # 使用固定日期确保一致性 - context_text = _resolve_relative_times(context_text, anchor_date) - - context_text = f"Reference date: {anchor_date.date().isoformat()}\n\n" + context_text - - print(f"📝 最终上下文长度: {len(context_text)} 字符") - - # 显示不同上下文的预览（不只是第一条） - print("🔍 上下文预览:") - for j, context in enumerate(contexts_all[:3]): # 显示前3个上下文 - preview = context[:150].replace('\n', ' ') - print(f" 上下文{j+1}: {preview}...") - - # 🔍 调试：检查答案是否在上下文中 - if ref_str and ref_str.strip(): - answer_found = any(ref_str.lower() in ctx.lower() for ctx in contexts_all) - print(f"🔍 调试：答案 '{ref_str}' 是否在检索到的上下文中？ {'✅ 是' if answer_found else '❌ 否'}") - - else: - print("❌ 没有检索到有效上下文") - context_text = "No relevant context found." - - # LLM 回答 - messages = [ - {"role": "system", "content": ( - "You are a precise QA assistant. Answer following these rules:\n" - "1) Extract the EXACT information mentioned in the context\n" - "2) For time questions: calculate actual dates from relative times\n" - "3) Return ONLY the answer text in simplest form\n" - "4) For dates, use format 'DD Month YYYY' (e.g., '7 May 2023')\n" - "5) If no clear answer found, respond with 'Unknown'" - )}, - {"role": "user", "content": f"Question: {q}\n\nContext:\n{context_text}"}, - ] - - t2 = time.time() - try: - # 使用异步调用 - resp = await llm.chat(messages=messages) - # 兼容不同的响应格式 - pred = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else "Unknown") - except Exception as e: - print(f"❌ LLM 生成失败: {e}") - pred = "Unknown" - t3 = time.time() - llm_time = (t3 - t2) * 1000 - - # 计算指标 - 使用导入的指标函数 - f1_val = f1_score(pred, ref_str) - bleu1_val = bleu1(pred, ref_str) - jaccard_val = jaccard(pred, ref_str) - loc_f1_val = loc_f1_score(pred, ref_str) - - print(f"🤖 LLM 回答: {pred}") - print(f"📈 指标 - F1: {f1_val:.3f}, BLEU-1: {bleu1_val:.3f}, Jaccard: {jaccard_val:.3f}, LoCoMo F1: {loc_f1_val:.3f}") - print(f"⏱️ 时间 - 检索: {search_time:.1f}ms, LLM: {llm_time:.1f}ms") - - # 更新统计 - total_f1 += f1_val - total_bleu1 += bleu1_val - total_jaccard += jaccard_val - total_loc_f1 += loc_f1_val - total_context_length += len(context_text) - total_retrieved_docs += len(contexts_all) - - if category not in category_stats: - category_stats[category] = {"count": 0, "f1_sum": 0.0, "b1_sum": 0.0, "j_sum": 0.0, "loc_f1_sum": 0.0} - - category_stats[category]["count"] += 1 - category_stats[category]["f1_sum"] += f1_val - category_stats[category]["b1_sum"] += bleu1_val - category_stats[category]["j_sum"] += jaccard_val - category_stats[category]["loc_f1_sum"] += loc_f1_val - - # 记录性能指标 - metrics = {"f1": f1_val, "bleu1": bleu1_val, "jaccard": jaccard_val, "loc_f1": loc_f1_val} - monitor.record_performance(i, metrics, len(context_text), len(contexts_all)) - - # 保存结果 - question_result = { - "question": q, - "ground_truth": ref_str, - "prediction": pred, - "category": category, - "metrics": metrics, - "retrieval": { - "retrieved_documents": len(contexts_all), - "context_length": len(context_text), - "search_limit": search_limit, - "max_chars": max_chars, - "recent_performance": recent_performance - }, - "timing": { - "search_ms": search_time, - "llm_ms": llm_time - } - } - - results["questions"].append(question_result) - - print("="*60) - - except Exception as e: - print(f"❌ 评估过程中发生错误: {e}") - # 即使出错，也返回已有的结果 - import traceback - traceback.print_exc() - - finally: - await connector.close() - - finally: - db.close() # 关闭数据库会话 - - # 计算总体指标 - n = len(items) - if n > 0: - results["overall_metrics"] = { - "f1": total_f1 / n, - "b1": total_bleu1 / n, - "j": total_jaccard / n, - "loc_f1": total_loc_f1 / n - } - - for category, stats in category_stats.items(): - count = stats["count"] - results["category_metrics"][category] = { - "count": count, - "f1": stats["f1_sum"] / count, - "bleu1": stats["b1_sum"] / count, - "jaccard": stats["j_sum"] / count, - "loc_f1": stats["loc_f1_sum"] / count - } - - results["retrieval_stats"]["avg_context_length"] = total_context_length / n - results["retrieval_stats"]["avg_retrieved_docs"] = total_retrieved_docs / n - - # 分析性能趋势 - results["performance_trend"] = monitor.get_performance_trend() - results["reset_interval"] = monitor.reset_interval - results["total_questions_processed"] = monitor.question_count - - return results - - -if __name__ == "__main__": - print("🚀 运行增强版完整评估（解决中间性能衰减问题）...") - print("📋 增强特性:") - print(" - 双重重置策略：定期重置 + 性能驱动重置") - print(" - 动态检索参数：基于近期性能自适应调整") - print(" - 中间阶段严格筛选：提高上下文质量要求") - print(" - 连续性能监控：实时检测性能衰减") - - result = asyncio.run(run_enhanced_evaluation()) - - print("\n📊 最终评估结果:") - print("总体指标:") - print(f" F1: {result['overall_metrics']['f1']:.4f}") - print(f" BLEU-1: {result['overall_metrics']['b1']:.4f}") - print(f" Jaccard: {result['overall_metrics']['j']:.4f}") - print(f" LoCoMo F1: {result['overall_metrics']['loc_f1']:.4f}") - - print("\n分类别指标:") - for category, metrics in result['category_metrics'].items(): - print(f" {category}: F1={metrics['f1']:.4f}, BLEU-1={metrics['bleu1']:.4f}, Jaccard={metrics['jaccard']:.4f}, LoCoMo F1={metrics['loc_f1']:.4f} (样本数: {metrics['count']})") - - print("\n检索统计:") - stats = result['retrieval_stats'] - print(f" 平均上下文长度: {stats['avg_context_length']:.0f} 字符") - print(f" 平均检索文档数: {stats['avg_retrieved_docs']:.1f}") - - print(f"\n性能趋势: {result['performance_trend']}") - print(f"重置间隔: 每{result['reset_interval']}个问题") - print(f"处理问题总数: {result['total_questions_processed']}") - print(f"增强策略: {'启用' if result.get('enhanced_strategy', False) else '未启用'}") - - - # 保存结果到指定目录 - # 使用代码文件所在目录的绝对路径 - current_file_dir = os.path.dirname(os.path.abspath(__file__)) - output_dir = os.path.join(current_file_dir, "results") - os.makedirs(output_dir, exist_ok=True) - output_file = os.path.join(output_dir, "enhanced_evaluation_results.json") - with open(output_file, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n详细结果已保存到: {output_file}") diff --git a/api/app/core/memory/evaluation/locomo/locomo_utils.py b/api/app/core/memory/evaluation/locomo/locomo_utils.py deleted file mode 100644 index 6ad68470..00000000 --- a/api/app/core/memory/evaluation/locomo/locomo_utils.py +++ /dev/null @@ -1,687 +0,0 @@ -""" -LoCoMo Utilities Module - -This module provides helper functions for the LoCoMo benchmark evaluation: -- Data loading from JSON files -- Conversation extraction for ingestion -- Temporal reference resolution -- Context selection and formatting -- Retrieval wrapper functions -- Ingestion wrapper functions -""" - -import os -import json -import re -from datetime import datetime, timedelta -from typing import List, Dict, Any, Optional -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - -from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline - - -def load_locomo_data( - data_path: str, - sample_size: int, - conversation_index: int = 0 -) -> List[Dict[str, Any]]: - """ - Load LoCoMo dataset from JSON file. - - The LoCoMo dataset structure is a list of conversation objects, where each - object contains a "qa" list of question-answer pairs. - - Args: - data_path: Path to locomo10.json file - sample_size: Number of QA pairs to load (limits total QA items returned) - conversation_index: Which conversation to load QA pairs from (default: 0 for first) - - Returns: - List of QA item dictionaries, each containing: - - question: str - - answer: str - - category: int (1-4) - - evidence: List[str] - - Raises: - FileNotFoundError: If data_path does not exist - json.JSONDecodeError: If file is not valid JSON - IndexError: If conversation_index is out of range - """ - if not os.path.exists(data_path): - raise FileNotFoundError(f"LoCoMo data file not found: {data_path}") - - with open(data_path, "r", encoding="utf-8") as f: - raw = json.load(f) - - # LoCoMo data structure: list of objects, each with a "qa" list - qa_items: List[Dict[str, Any]] = [] - - if isinstance(raw, list): - # Only load QA pairs from the specified conversation - if conversation_index < len(raw): - entry = raw[conversation_index] - if isinstance(entry, dict) and "qa" in entry: - qa_items.extend(entry.get("qa", [])) - else: - raise IndexError( - f"Conversation index {conversation_index} out of range. " - f"Dataset has {len(raw)} conversations." - ) - else: - # Fallback: single object with qa list - if conversation_index == 0: - qa_items.extend(raw.get("qa", [])) - else: - raise IndexError( - f"Conversation index {conversation_index} out of range. " - f"Dataset has only 1 conversation." - ) - - # Return only the requested sample size - return qa_items[:sample_size] - - -def extract_conversations(data_path: str, max_dialogues: int = 1, max_messages_per_dialogue: Optional[int] = None) -> List[str]: - """ - Extract conversation texts from LoCoMo data for ingestion. - - This function extracts the raw conversation dialogues from the LoCoMo dataset - so they can be ingested into the memory system. Each conversation is formatted - as a multi-line string with "role: message" format. - - Args: - data_path: Path to locomo10.json file - max_dialogues: Maximum number of dialogues to extract (default: 1) - max_messages_per_dialogue: Maximum messages per dialogue (default: None = all messages) - - Returns: - List of conversation strings formatted for ingestion. - Each string contains multiple lines in format "role: message" - - Example output: - [ - "User: I went to the store yesterday.\\nAI: What did you buy?\\n...", - "User: I love hiking.\\nAI: Where do you like to hike?\\n..." - ] - """ - if not os.path.exists(data_path): - raise FileNotFoundError(f"LoCoMo data file not found: {data_path}") - - with open(data_path, "r", encoding="utf-8") as f: - raw = json.load(f) - - # Ensure we have a list of entries - entries = raw if isinstance(raw, list) else [raw] - - contents: List[str] = [] - - for i, entry in enumerate(entries[:max_dialogues]): - if not isinstance(entry, dict): - continue - - conv = entry.get("conversation", {}) - - if not isinstance(conv, dict): - continue - - lines: List[str] = [] - - # Collect all session_* messages - for key, val in sorted(conv.items()): - if isinstance(val, list) and key.startswith("session_"): - for msg in val: - if not isinstance(msg, dict): - continue - - role = msg.get("speaker") or "User" - text = msg.get("text") or "" - text = str(text).strip() - - if not text: - continue - - lines.append(f"{role}: {text}") - - # Limit messages if specified - if max_messages_per_dialogue and len(lines) >= max_messages_per_dialogue: - break - - # Break outer loop if we've reached the message limit - if max_messages_per_dialogue and len(lines) >= max_messages_per_dialogue: - break - - if lines: - contents.append("\n".join(lines)) - - return contents - -# 时间解析：将相对时间表达转换为绝对日期 -def resolve_temporal_references(text: str, anchor_date: datetime) -> str: - """ - Resolve relative temporal references to absolute dates. - - This function converts relative time expressions (like "today", "yesterday", - "3 days ago") into absolute ISO date strings based on an anchor date. - - Supported patterns: - - today, yesterday, tomorrow - - X days ago, in X days - - last week, next week - - Args: - text: Text containing temporal references - anchor_date: Reference date for resolution (datetime object) - - Returns: - Text with temporal references replaced by ISO dates (YYYY-MM-DD format) - - Example: - >>> anchor = datetime(2023, 5, 8) - >>> resolve_temporal_references("I saw him yesterday", anchor) - "I saw him 2023-05-07" - """ - # Ensure input is a string - t = str(text) if text is not None else "" - - # today / yesterday / tomorrow - t = re.sub( - r"\btoday\b", - anchor_date.date().isoformat(), - t, - flags=re.IGNORECASE - ) - t = re.sub( - r"\byesterday\b", - (anchor_date - timedelta(days=1)).date().isoformat(), - t, - flags=re.IGNORECASE - ) - t = re.sub( - r"\btomorrow\b", - (anchor_date + timedelta(days=1)).date().isoformat(), - t, - flags=re.IGNORECASE - ) - - # X days ago - def _ago_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor_date - timedelta(days=n)).date().isoformat() - - # in X days - def _in_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor_date + timedelta(days=n)).date().isoformat() - - t = re.sub( - r"\b(\d+)\s+days?\s+ago\b", - _ago_repl, - t, - flags=re.IGNORECASE - ) - t = re.sub( - r"\bin\s+(\d+)\s+days?\b", - _in_repl, - t, - flags=re.IGNORECASE - ) - - # last week / next week (approximate as 7 days) - t = re.sub( - r"\blast\s+week\b", - (anchor_date - timedelta(days=7)).date().isoformat(), - t, - flags=re.IGNORECASE - ) - - # 中文支持 - t = re.sub( - r"\bnext\s+week\b", - (anchor_date + timedelta(days=7)).date().isoformat(), - t, - flags=re.IGNORECASE - ) - - return t - - -def select_and_format_information( - retrieved_info: List[str], - question: str, - max_chars: int = 8000 -) -> str: - """ - Intelligently select and format most relevant retrieved information for LLM prompt. - - This function scores each piece of retrieved information based on keyword matching - with the question, then selects the highest-scoring pieces up to the character limit. - - Scoring criteria: - - Keyword matches (higher weight for multiple occurrences) - - Context length (moderate length preferred) - - Position (earlier contexts get bonus points) - - Args: - retrieved_info: List of retrieved information strings (chunks, statements, entities) - question: Question being answered - max_chars: Maximum total characters to include in final prompt - - Returns: - Formatted string combining the most relevant information for LLM prompt. - Contexts are separated by double newlines. - - Example: - >>> contexts = ["Alice went to Paris", "Bob likes pizza", "Alice visited the Eiffel Tower"] - >>> question = "Where did Alice go?" - >>> select_and_format_information(contexts, question, max_chars=100) - "Alice went to Paris\\n\\nAlice visited the Eiffel Tower" - """ - if not retrieved_info: - return "" - - # Extract question keywords (filter out stop words and short words) - question_lower = question.lower() - stop_words = { - 'what', 'when', 'where', 'who', 'why', 'how', - 'did', 'do', 'does', 'is', 'are', 'was', 'were', - 'the', 'a', 'an', 'and', 'or', 'but', 'in', 'on', 'at' - } - question_words = set(re.findall(r'\b\w+\b', question_lower)) - question_words = { - word for word in question_words - if word not in stop_words and len(word) > 2 - } - - # Score each context - scored_contexts = [] - for i, context in enumerate(retrieved_info): - context_lower = context.lower() - score = 0 - - # Keyword matching score - keyword_matches = 0 - for word in question_words: - if word in context_lower: - keyword_matches += 1 - # Multiple occurrences increase score - score += context_lower.count(word) * 2 - - # Length score (prefer moderate length) - context_len = len(context) - if 100 < context_len < 2000: - score += 5 - elif context_len >= 2000: - score += 2 - - # Position bonus (earlier contexts often more relevant) - if i < 3: - score += 3 - - scored_contexts.append((score, context, keyword_matches)) - - # Sort by score (descending) - scored_contexts.sort(key=lambda x: x[0], reverse=True) - - # Select contexts up to character limit - selected = [] - total_chars = 0 - - for score, context, matches in scored_contexts: - if total_chars + len(context) <= max_chars: - selected.append(context) - total_chars += len(context) - else: - # Try to include high-scoring context by truncating - if score > 10 and total_chars < max_chars - 500: - remaining = max_chars - total_chars - # Find lines with keywords - lines = context.split('\n') - relevant_lines = [] - current_chars = 0 - - for line in lines: - line_lower = line.lower() - line_relevance = any(word in line_lower for word in question_words) - - if line_relevance and current_chars < remaining - 100: - relevant_lines.append(line) - current_chars += len(line) - - if relevant_lines and len('\n'.join(relevant_lines)) > 100: - truncated = '\n'.join(relevant_lines) - selected.append(truncated + "\n[Content truncated...]") - total_chars += len(truncated) - break - - return "\n\n".join(selected) - -# 记忆系统核心能力：写入与读取 -async def ingest_conversations_if_needed( - conversations: List[str], - end_user_id: str, - reset: bool = False -) -> bool: - """ - Wrapper for conversation ingestion using external extraction pipeline. - - This function populates the Neo4j database with processed conversation data - (chunks, statements, entities) so that the retrieval system has memory to search. - - The ingestion process: - 1. Parses conversation text into dialogue messages - 2. Chunks the dialogues into semantic units - 3. Extracts statements and entities using LLM - 4. Generates embeddings for all content - 5. Stores everything in Neo4j graph database - - Args: - conversations: List of raw conversation texts from LoCoMo dataset - Example: ["User: I went to Paris. AI: When was that?", ...] - end_user_id: Target end_user ID for database storage - reset: Whether to clear existing data first (not implemented in wrapper) - - Returns: - True if successful, False otherwise - - Note: - The external function uses "contexts" to mean "conversation texts". - This runs the full extraction pipeline: chunking → entity extraction → - statement extraction → embedding → Neo4j storage. - """ - try: - success = await ingest_contexts_via_full_pipeline( - contexts=conversations, - end_user_id=end_user_id, - save_chunk_output=True, - reset_group=reset - ) - return success - except Exception as e: - print(f"[Ingestion] Failed to ingest conversations: {e}") - return False - -async def retrieve_relevant_information( - question: str, - end_user_id: str, - search_type: str, - search_limit: int, - connector: Any, - embedder: Any -) -> List[str]: - """ - Retrieve relevant information from memory graph for a question. - - This function searches the Neo4j memory graph (populated during ingestion) and - returns relevant chunks, statements, and entity information that might help - answer the question. - - The function supports three search types: - - "keyword": Full-text search using Cypher queries - - "embedding": Vector similarity search using embeddings - - "hybrid": Combination of keyword and embedding search with reranking - - Args: - question: Question to search for - end_user_id: Database group ID (identifies which conversation memory to search) - search_type: "keyword", "embedding", or "hybrid" - search_limit: Max memory pieces to retrieve - connector: Neo4j connector instance - embedder: Embedder client instance - - Returns: - List of text strings (chunks, statements, entity summaries) from memory graph. - Each string represents a piece of retrieved information. - - Raises: - Exception: If search fails (caught and returns empty list) - """ - from app.repositories.neo4j.graph_search import ( - search_graph, - search_graph_by_embedding - ) - from app.core.memory.src.search import run_hybrid_search - - contexts_all: List[str] = [] - - try: - if search_type == "embedding": - # Embedding-based search - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], - ) - - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - # Build context from chunks - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - - # Add statements - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - # Add summaries - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # Add top entities (limit to 3 to avoid noise) - if entities: - scored = [e for e in entities if e.get("score") is not None] - top_entities = ( - sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] - if scored else entities[:3] - ) - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append( - f"EntitySummary: {name}" - f"{(' [' + '; '.join(meta) + ']') if meta else ''}" - ) - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - elif search_type == "keyword": - # Keyword-based search - search_results = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit - ) - - dialogs = search_results.get("dialogues", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - - # Build context from dialogues - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - - # Add statements - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - # Add entity names - if entities: - entity_names = [ - str(e.get("name", "")).strip() - for e in entities[:5] - if e.get("name") - ] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - else: # hybrid - # Hybrid search with fallback to embedding - try: - search_results = await run_hybrid_search( - query_text=question, - search_type=search_type, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], - output_path=None, - ) - - # Handle flat structure (new API format) - if search_results and isinstance(search_results, dict): - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - # Check if we got results - if not (chunks or statements or entities or summaries): - # Try nested structure (backward compatibility) - reranked = search_results.get("reranked_results", {}) - if reranked and isinstance(reranked, dict): - chunks = reranked.get("chunks", []) - statements = reranked.get("statements", []) - entities = reranked.get("entities", []) - summaries = reranked.get("summaries", []) - else: - raise ValueError("Hybrid search returned empty results") - else: - raise ValueError("Hybrid search returned empty results") - - except Exception as e: - # Fallback to embedding search - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], - ) - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - # Build context (same for both hybrid and fallback) - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # Add top entities - if entities: - scored = [e for e in entities if e.get("score") is not None] - top_entities = ( - sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] - if scored else entities[:3] - ) - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append( - f"EntitySummary: {name}" - f"{(' [' + '; '.join(meta) + ']') if meta else ''}" - ) - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - except Exception as e: - # Return empty list on error - contexts_all = [] - - return contexts_all - - -async def ingest_conversations_if_needed( - conversations: List[str], - end_user_id: str, - reset: bool = False -) -> bool: - """ - Wrapper for conversation ingestion using external extraction pipeline. - - This function populates the Neo4j database with processed conversation data - (chunks, statements, entities) so that the retrieval system has memory to search. - - The ingestion process: - 1. Parses conversation text into dialogue messages - 2. Chunks the dialogues into semantic units - 3. Extracts statements and entities using LLM - 4. Generates embeddings for all content - 5. Stores everything in Neo4j graph database - - Args: - conversations: List of raw conversation texts from LoCoMo dataset - Example: ["User: I went to Paris. AI: When was that?", ...] - end_user_id: Target group ID for database storage - reset: Whether to clear existing data first (not implemented in wrapper) - - Returns: - True if successful, False otherwise - - Note: - The external function uses "contexts" to mean "conversation texts". - This runs the full extraction pipeline: chunking → entity extraction → - statement extraction → embedding → Neo4j storage. - """ - try: - success = await ingest_contexts_via_full_pipeline( - contexts=conversations, - end_user_id=end_user_id, - save_chunk_output=True - ) - return success - except Exception as e: - print(f"[Ingestion] Failed to ingest conversations: {e}") - return False diff --git a/api/app/core/memory/evaluation/locomo/qwen_search_eval.py b/api/app/core/memory/evaluation/locomo/qwen_search_eval.py deleted file mode 100644 index 889c5065..00000000 --- a/api/app/core/memory/evaluation/locomo/qwen_search_eval.py +++ /dev/null @@ -1,874 +0,0 @@ -import argparse -import asyncio -import json -import os -import time -from datetime import datetime, timedelta -from typing import List, Dict, Any -import statistics -import re -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.models.base import RedBearModelConfig -from app.core.memory.utils.config.config_utils import get_embedder_config -from app.core.memory.src.search import run_hybrid_search # 使用旧版本（重构前） -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline -from app.core.memory.evaluation.common.metrics import f1_score as common_f1, bleu1, jaccard, latency_stats, avg_context_tokens - - -# 参考 evaluation/locomo/evaluation.py 的 F1 计算逻辑（移除外部依赖，内联实现） -def _loc_normalize(text: str) -> str: - import re - # 确保输入是字符串 - text = str(text) if text is not None else "" - text = text.lower() - text = re.sub(r"[\,]", " ", text) # 去掉逗号 - text = re.sub(r"\b(a|an|the|and)\b", " ", text) - text = re.sub(r"[^\w\s]", " ", text) - text = " ".join(text.split()) - return text - -# 追加：相对时间归一化为绝对日期（有限支持：today/yesterday/tomorrow/X days ago/in X days/last week/next week） -def _resolve_relative_times(text: str, anchor: datetime) -> str: - import re - # 确保输入是字符串 - t = str(text) if text is not None else "" - # today / yesterday / tomorrow - t = re.sub(r"\btoday\b", anchor.date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\byesterday\b", (anchor - timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\btomorrow\b", (anchor + timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - # X days ago / in X days - def _ago_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor - timedelta(days=n)).date().isoformat() - def _in_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor + timedelta(days=n)).date().isoformat() - t = re.sub(r"\b(\d+)\s+days\s+ago\b", _ago_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\bin\s+(\d+)\s+days\b", _in_repl, t, flags=re.IGNORECASE) - # last week / next week（以7天近似） - t = re.sub(r"\blast\s+week\b", (anchor - timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\bnext\s+week\b", (anchor + timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - return t - -def loc_f1_score(prediction: str, ground_truth: str) -> float: - # 单答案 F1：按词集合计算（近似原始实现，去除词干依赖） - # 确保输入是字符串 - pred_str = str(prediction) if prediction is not None else "" - truth_str = str(ground_truth) if ground_truth is not None else "" - - p_tokens = _loc_normalize(pred_str).split() - g_tokens = _loc_normalize(truth_str).split() - if not p_tokens or not g_tokens: - return 0.0 - p = set(p_tokens) - g = set(g_tokens) - tp = len(p & g) - precision = tp / len(p) if p else 0.0 - recall = tp / len(g) if g else 0.0 - return (2 * precision * recall / (precision + recall)) if (precision + recall) > 0 else 0.0 - -def loc_multi_f1(prediction: str, ground_truth: str) -> float: - # 多答案 F1：prediction 与 ground_truth 以逗号分隔，逐一匹配取最大，再对多个 GT 取平均 - # 确保输入是字符串 - pred_str = str(prediction) if prediction is not None else "" - truth_str = str(ground_truth) if ground_truth is not None else "" - - predictions = [p.strip() for p in str(pred_str).split(',') if p.strip()] - ground_truths = [g.strip() for g in str(truth_str).split(',') if g.strip()] - if not predictions or not ground_truths: - return 0.0 - def _f1(a: str, b: str) -> float: - return loc_f1_score(a, b) - vals = [] - for gt in ground_truths: - vals.append(max(_f1(pred, gt) for pred in predictions)) - return sum(vals) / len(vals) - -# 标准化 LoCoMo 类别名：支持数字 category 与字符串 cat/type -CATEGORY_MAP_NUM_TO_NAME = { - 4: "Single-Hop", - 1: "Multi-Hop", - 3: "Open Domain", - 2: "Temporal", -} - -_TYPE_ALIASES = { - "single-hop": "Single-Hop", - "singlehop": "Single-Hop", - "single hop": "Single-Hop", - "multi-hop": "Multi-Hop", - "multihop": "Multi-Hop", - "multi hop": "Multi-Hop", - "open domain": "Open Domain", - "opendomain": "Open Domain", - "temporal": "Temporal", -} - -def get_category_label(item: Dict[str, Any]) -> str: - # 1) 直接用字符串 cat - cat = item.get("cat") - if isinstance(cat, str) and cat.strip(): - name = cat.strip() - lower = name.lower() - return _TYPE_ALIASES.get(lower, name) - # 2) 数字 category 转名称 - cat_num = item.get("category") - if isinstance(cat_num, int): - return CATEGORY_MAP_NUM_TO_NAME.get(cat_num, "unknown") - # 3) 备用 type 字段 - t = item.get("type") - if isinstance(t, str) and t.strip(): - lower = t.strip().lower() - return _TYPE_ALIASES.get(lower, t.strip()) - return "unknown" - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 12000) -> str: - """基于问题关键词智能选择上下文""" - if not contexts: - return "" - - # 提取问题关键词（只保留有意义的词） - question_lower = question.lower() - stop_words = {'what', 'when', 'where', 'who', 'why', 'how', 'did', 'do', 'does', 'is', 'are', 'was', 'were', 'the', 'a', 'an', 'and', 'or', 'but'} - question_words = set(re.findall(r'\b\w+\b', question_lower)) - question_words = {word for word in question_words if word not in stop_words and len(word) > 2} - - print(f"🔍 问题关键词: {question_words}") - - # 给每个上下文打分 - scored_contexts = [] - for i, context in enumerate(contexts): - context_lower = context.lower() - score = 0 - - # 关键词匹配得分 - keyword_matches = 0 - for word in question_words: - if word in context_lower: - keyword_matches += 1 - # 关键词出现次数越多，得分越高 - score += context_lower.count(word) * 2 - - # 上下文长度得分（适中的长度更好） - context_len = len(context) - if 100 < context_len < 2000: # 理想长度范围 - score += 5 - elif context_len >= 2000: # 太长可能包含无关信息 - score += 2 - - # 如果是前几个上下文，给予额外分数（通常相关性更高） - if i < 3: - score += 3 - - scored_contexts.append((score, context, keyword_matches)) - - # 按得分排序 - scored_contexts.sort(key=lambda x: x[0], reverse=True) - - # 选择高得分的上下文，直到达到字符限制 - selected = [] - total_chars = 0 - selected_count = 0 - - print("📊 上下文相关性分析:") - for score, context, matches in scored_contexts[:5]: # 只显示前5个 - print(f" - 得分: {score}, 关键词匹配: {matches}, 长度: {len(context)}") - - for score, context, matches in scored_contexts: - if total_chars + len(context) <= max_chars: - selected.append(context) - total_chars += len(context) - selected_count += 1 - else: - # 如果这个上下文得分很高但放不下，尝试截取 - if score > 10 and total_chars < max_chars - 500: - remaining = max_chars - total_chars - # 找到包含关键词的部分 - lines = context.split('\n') - relevant_lines = [] - current_chars = 0 - - for line in lines: - line_lower = line.lower() - line_relevance = any(word in line_lower for word in question_words) - - if line_relevance and current_chars < remaining - 100: - relevant_lines.append(line) - current_chars += len(line) - - if relevant_lines: - truncated = '\n'.join(relevant_lines) - if len(truncated) > 100: # 确保有足够内容 - selected.append(truncated + "\n[相关内容截断...]") - total_chars += len(truncated) - selected_count += 1 - break # 不再尝试添加更多上下文 - - result = "\n\n".join(selected) - print(f"✅ 智能选择: {selected_count}个上下文, 总长度: {total_chars}字符") - return result - - -def get_search_params_by_category(category: str): - """根据问题类别调整检索参数""" - params_map = { - "Multi-Hop": {"limit": 20, "max_chars": 15000}, - "Temporal": {"limit": 16, "max_chars": 10000}, - "Open Domain": {"limit": 24, "max_chars": 18000}, - "Single-Hop": {"limit": 12, "max_chars": 8000}, - } - return params_map.get(category, {"limit": 16, "max_chars": 12000}) - - -async def run_locomo_eval( - sample_size: int = 1, - end_user_id: str | None = None, - search_limit: int = 8, - context_char_budget: int = 4000, # 保持默认值不变 - llm_temperature: float = 0.0, - llm_max_tokens: int = 32, - search_type: str = "hybrid", # 保持默认值不变 - output_path: str | None = None, - skip_ingest_if_exists: bool = True, - llm_timeout: float = 10.0, - llm_max_retries: int = 1 -) -> Dict[str, Any]: - - # 函数内部使用三路检索逻辑，但保持参数签名不变 - end_user_id = end_user_id or SELECTED_end_user_id - data_path = os.path.join(PROJECT_ROOT, "data", "locomo10.json") - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 locomo10.json 放置在: {dataset_dir}" - ) - with open(data_path, "r", encoding="utf-8") as f: - raw = json.load(f) - # LoCoMo 数据结构：顶层为若干对象，每个对象下有 qa 列表 - qa_items: List[Dict[str, Any]] = [] - if isinstance(raw, list): - for entry in raw: - qa_items.extend(entry.get("qa", [])) - else: - qa_items.extend(raw.get("qa", [])) - items: List[Dict[str, Any]] = qa_items[:sample_size] - - # === 保持原来的数据摄入逻辑 === - entries = raw if isinstance(raw, list) else [raw] - - # 只摄入前1条对话（保持原样） - max_dialogues_to_ingest = 1 - contents: List[str] = [] - print(f"📊 找到 {len(entries)} 个对话对象，只摄入前 {max_dialogues_to_ingest} 条") - - for i, entry in enumerate(entries[:max_dialogues_to_ingest]): - if not isinstance(entry, dict): - continue - - conv = entry.get("conversation", {}) - sample_id = entry.get("sample_id", f"unknown_{i}") - - print(f"🔍 处理对话 {i+1}: {sample_id}") - - lines: List[str] = [] - if isinstance(conv, dict): - # 收集所有 session_* 的消息 - session_count = 0 - for key, val in conv.items(): - if isinstance(val, list) and key.startswith("session_"): - session_count += 1 - for msg in val: - role = msg.get("speaker") or "用户" - text = msg.get("text") or "" - text = str(text).strip() - if not text: - continue - lines.append(f"{role}: {text}") - - print(f" - 包含 {session_count} 个session, {len(lines)} 条消息") - - if not lines: - print(f"⚠️ 警告: 对话 {sample_id} 没有对话内容，跳过摄入") - continue - - contents.append("\n".join(lines)) - - print(f"📥 总共摄入 {len(contents)} 个对话的conversation内容") - - # 选择要评测的QA对（从所有对话中选取） - indexed_items: List[tuple[int, Dict[str, Any]]] = [] - if isinstance(raw, list): - for e_idx, entry in enumerate(raw): - for qa in entry.get("qa", []): - indexed_items.append((e_idx, qa)) - else: - for qa in raw.get("qa", []): - indexed_items.append((0, qa)) - - # 这里使用sample_size来限制评测的QA数量 - selected = indexed_items[:sample_size] - items: List[Dict[str, Any]] = [qa for _, qa in selected] - - print(f"🎯 将评测 {len(items)} 个QA对，数据库中只包含 {len(contents)} 个对话") - # === 修改结束 === - - connector = Neo4jConnector() - - # 关键修复：强制重新摄入纯净的对话数据 - print("🔄 强制重新摄入纯净的对话数据...") - await ingest_contexts_via_full_pipeline(contents, end_user_id, save_chunk_output=True) - - # 使用异步LLM客户端 - llm_client = get_llm_client(llm_id) - # 初始化embedder用于直接调用 - cfg_dict = get_embedder_config(embedding_id) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - - # connector initialized above - latencies_llm: List[float] = [] - latencies_search: List[float] = [] - # 上下文诊断收集 - per_query_context_counts: List[int] = [] - per_query_context_avg_tokens: List[float] = [] - per_query_context_chars: List[int] = [] - per_query_context_tokens_total: List[int] = [] - # 详细样本调试信息 - samples: List[Dict[str, Any]] = [] - # 通用指标 - f1s: List[float] = [] - b1s: List[float] = [] - jss: List[float] = [] - # 参考 LoCoMo 评测的类别专用 F1（multi-hop 使用多答案 F1） - loc_f1s: List[float] = [] - # Per-category aggregation - cat_counts: Dict[str, int] = {} - cat_f1s: Dict[str, List[float]] = {} - cat_b1s: Dict[str, List[float]] = {} - cat_jss: Dict[str, List[float]] = {} - cat_loc_f1s: Dict[str, List[float]] = {} - try: - for item in items: - q = item.get("question", "") - ref = item.get("answer", "") - # 确保答案是字符串 - ref_str = str(ref) if ref is not None else "" - cat = get_category_label(item) - - print(f"\n=== 处理问题: {q} ===") - - # 根据类别调整检索参数 - search_params = get_search_params_by_category(cat) - adjusted_limit = search_params["limit"] - max_chars = search_params["max_chars"] - - print(f"🏷️ 类别: {cat}, 检索参数: limit={adjusted_limit}, max_chars={max_chars}") - - # 改进的检索逻辑：使用三路检索（statements, dialogues, entities） - t0 = time.time() - contexts_all: List[str] = [] - search_results = None # 保存完整的检索结果 - - try: - if search_type == "embedding": - # 直接调用嵌入检索，包含三路数据 - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=q, - end_user_id=end_user_id, - limit=adjusted_limit, - include=["chunks", "statements", "entities", "summaries"], # 修复：使用正确的类型 - ) - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - print(f"✅ 嵌入检索成功: {len(chunks)} chunks, {len(statements)} 条陈述, {len(entities)} 个实体, {len(summaries)} 个摘要") - - # 构建上下文：优先使用 chunks、statements 和 summaries - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # 实体摘要：最多加入前3个高分实体，避免噪声 - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - elif search_type == "keyword": - # 直接调用关键词检索 - search_results = await search_graph( - connector=connector, - q=q, - end_user_id=end_user_id, - limit=adjusted_limit - ) - dialogs = search_results.get("dialogues", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - print(f"🔤 关键词检索找到 {len(dialogs)} 条对话, {len(statements)} 条陈述, {len(entities)} 个实体") - - # 构建上下文 - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - # 实体处理（关键词检索的实体可能没有分数） - if entities: - entity_names = [str(e.get("name", "")).strip() for e in entities[:5] if e.get("name")] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - else: # hybrid - # 使用旧版本的混合检索（重构前） - print("🔀 使用混合检索（旧版本）...") - try: - search_results = await run_hybrid_search( - query_text=q, - search_type=search_type, - end_user_id=end_user_id, - limit=adjusted_limit, - include=["chunks", "statements", "entities", "summaries"], - output_path=None, - rerank_alpha=0.6, - use_forgetting_rerank=False, - use_llm_rerank=False - ) - - # 处理旧版本的返回结构（包含 reranked_results） - if search_results and isinstance(search_results, dict): - # 对于 hybrid 搜索，使用 reranked_results - if "reranked_results" in search_results: - reranked = search_results["reranked_results"] - chunks = reranked.get("chunks", []) - statements = reranked.get("statements", []) - entities = reranked.get("entities", []) - summaries = reranked.get("summaries", []) - else: - # 单一搜索类型的结果 - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - # 检查是否有有效结果 - if chunks or statements or entities or summaries: - print(f"✅ 混合检索成功: {len(chunks)} chunks, {len(statements)} 陈述, {len(entities)} 实体, {len(summaries)} 摘要") - else: - # 如果顶层没有结果，尝试旧的嵌套结构（向后兼容） - reranked = search_results.get("reranked_results", {}) - if reranked and isinstance(reranked, dict): - chunks = reranked.get("chunks", []) - statements = reranked.get("statements", []) - entities = reranked.get("entities", []) - summaries = reranked.get("summaries", []) - print(f"✅ 混合检索成功（使用旧格式reranked结果）: {len(chunks)} chunks, {len(statements)} 陈述") - else: - raise ValueError("混合检索返回空结果") - else: - raise ValueError("混合检索返回空结果") - - except Exception as e: - print(f"❌ 混合检索失败: {e}，回退到嵌入检索") - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=q, - end_user_id=end_user_id, - limit=adjusted_limit, - include=["chunks", "statements", "entities", "summaries"], - ) - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - print(f"✅ 回退嵌入检索成功: {len(chunks)} chunks, {len(statements)} 陈述") - - # 🎯 统一处理：构建上下文（所有检索类型共用） - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # 实体摘要：最多加入前3个高分实体 - if entities: - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - # 关键修复：过滤掉包含当前问题答案的上下文 - filtered_contexts = [] - for context in contexts_all: - content = str(context) - # 排除包含当前问题标准答案的上下文 - if ref_str and ref_str.strip() and ref_str.strip() in content: - print("🚫 过滤掉包含标准答案的上下文") - continue - filtered_contexts.append(context) - - print(f"📊 过滤后保留 {len(filtered_contexts)} 个上下文 (原 {len(contexts_all)} 个)") - contexts_all = filtered_contexts - - # 输出完整的检索结果信息 - print("🔍 检索结果详情:") - if search_results: - output_data = { - "statements": [ - { - "statement": s.get("statement", "")[:200] + "..." if len(s.get("statement", "")) > 200 else s.get("statement", ""), - "score": s.get("score", 0.0) - } - for s in (statements[:2] if 'statements' in locals() else []) - ], - "dialogues": [ - { - "uuid": d.get("uuid", ""), - "end_user_id": d.get("end_user_id", ""), - "content": d.get("content", "")[:200] + "..." if len(d.get("content", "")) > 200 else d.get("content", ""), - "score": d.get("score", 0.0) - } - for d in (dialogs[:2] if 'dialogs' in locals() else []) - ], - "entities": [ - { - "name": e.get("name", ""), - "entity_type": e.get("entity_type", ""), - "score": e.get("score", 0.0) - } - for e in (entities[:2] if 'entities' in locals() else []) - ] - } - print(json.dumps(output_data, ensure_ascii=False, indent=2)) - else: - print(" 无检索结果") - - except Exception as e: - print(f"❌ {search_type}检索失败: {e}") - contexts_all = [] - search_results = None - - t1 = time.time() - latencies_search.append((t1 - t0) * 1000) - - # 使用智能上下文选择 - context_text = "" - if contexts_all: - context_text = smart_context_selection(contexts_all, q, max_chars=max_chars) - - # 如果智能选择后仍然过长，进行最终保护性截断 - if len(context_text) > max_chars: - print(f"⚠️ 智能选择后仍然过长 ({len(context_text)}字符)，进行最终截断") - context_text = context_text[:max_chars] + "\n\n[最终截断...]" - - # 时间解析 - anchor_date = datetime(2023, 5, 8) # 使用固定日期确保一致性 - context_text = _resolve_relative_times(context_text, anchor_date) - - context_text = f"Reference date: {anchor_date.date().isoformat()}\n\n" + context_text - - print(f"📝 最终上下文长度: {len(context_text)} 字符") - - # 显示不同上下文的预览 - print("🔍 上下文预览:") - for j, context in enumerate(contexts_all[:3]): # 显示前3个上下文 - preview = context[:150].replace('\n', ' ') - print(f" 上下文{j+1}: {preview}...") - - else: - print("❌ 没有检索到有效上下文") - context_text = "No relevant context found." - - # 记录上下文诊断信息 - per_query_context_counts.append(len(contexts_all)) - per_query_context_avg_tokens.append(avg_context_tokens([context_text])) - per_query_context_chars.append(len(context_text)) - per_query_context_tokens_total.append(len(_loc_normalize(context_text).split())) - - # LLM 提示词 - messages = [ - {"role": "system", "content": ( - "You are a precise QA assistant. Answer following these rules:\n" - "1) Extract the EXACT information mentioned in the context\n" - "2) For time questions: calculate actual dates from relative times\n" - "3) Return ONLY the answer text in simplest form\n" - "4) For dates, use format 'DD Month YYYY' (e.g., '7 May 2023')\n" - "5) If no clear answer found, respond with 'Unknown'" - )}, - {"role": "user", "content": f"Question: {q}\n\nContext:\n{context_text}"}, - ] - - t2 = time.time() - # 使用异步调用 - resp = await llm_client.chat(messages=messages) - t3 = time.time() - latencies_llm.append((t3 - t2) * 1000) - - # 兼容不同的响应格式 - pred = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else "Unknown") - - # 计算指标（确保使用字符串） - f1_val = common_f1(str(pred), ref_str) - b1_val = bleu1(str(pred), ref_str) - j_val = jaccard(str(pred), ref_str) - - f1s.append(f1_val) - b1s.append(b1_val) - jss.append(j_val) - - # Accumulate by category - cat_counts[cat] = cat_counts.get(cat, 0) + 1 - cat_f1s.setdefault(cat, []).append(f1_val) - cat_b1s.setdefault(cat, []).append(b1_val) - cat_jss.setdefault(cat, []).append(j_val) - - # LoCoMo 专用 F1：multi-hop(1) 使用多答案 F1，其它(2/3/4)使用单答案 F1 - if item.get("category") in [2, 3, 4]: - loc_val = loc_f1_score(str(pred), ref_str) - elif item.get("category") in [1]: - loc_val = loc_multi_f1(str(pred), ref_str) - else: - loc_val = loc_f1_score(str(pred), ref_str) - loc_f1s.append(loc_val) - cat_loc_f1s.setdefault(cat, []).append(loc_val) - - # 保存完整的检索结果信息 - samples.append({ - "question": q, - "answer": ref_str, - "category": cat, - "prediction": pred, - "metrics": { - "f1": f1_val, - "b1": b1_val, - "j": j_val, - "loc_f1": loc_val - }, - "retrieval": { - "retrieved_documents": len(contexts_all), - "context_length": len(context_text), - "search_limit": adjusted_limit, - "max_chars": max_chars - }, - "timing": { - "search_ms": (t1 - t0) * 1000, - "llm_ms": (t3 - t2) * 1000 - } - }) - - print(f"🤖 LLM 回答: {pred}") - print(f"✅ 正确答案: {ref_str}") - print(f"📈 当前指标 - F1: {f1_val:.3f}, BLEU-1: {b1_val:.3f}, Jaccard: {j_val:.3f}, LoCoMo F1: {loc_val:.3f}") - - # Compute per-category averages and dispersion (std, iqr) - def _percentile(sorted_vals: List[float], p: float) -> float: - if not sorted_vals: - return 0.0 - if len(sorted_vals) == 1: - return sorted_vals[0] - k = (len(sorted_vals) - 1) * p - f = int(k) - c = f + 1 if f + 1 < len(sorted_vals) else f - if f == c: - return sorted_vals[f] - return sorted_vals[f] + (sorted_vals[c] - sorted_vals[f]) * (k - f) - - by_category: Dict[str, Dict[str, float | int]] = {} - for c in cat_counts: - f_list = cat_f1s.get(c, []) - b_list = cat_b1s.get(c, []) - j_list = cat_jss.get(c, []) - lf_list = cat_loc_f1s.get(c, []) - j_sorted = sorted(j_list) - j_std = statistics.stdev(j_list) if len(j_list) > 1 else 0.0 - j_q75 = _percentile(j_sorted, 0.75) - j_q25 = _percentile(j_sorted, 0.25) - by_category[c] = { - "count": cat_counts[c], - "f1": (sum(f_list) / max(len(f_list), 1)) if f_list else 0.0, - "b1": (sum(b_list) / max(len(b_list), 1)) if b_list else 0.0, - "j": (sum(j_list) / max(len(j_list), 1)) if j_list else 0.0, - "j_std": j_std, - "j_iqr": (j_q75 - j_q25) if j_list else 0.0, - # 参考 LoCoMo 评测的类别专用 F1 - "loc_f1": (sum(lf_list) / max(len(lf_list), 1)) if lf_list else 0.0, - } - - # 累加命中（cum accuracy by category）：与 evaluation_stats.py 输出形式相仿 - cum_accuracy_by_category = {c: sum(cat_loc_f1s.get(c, [])) for c in cat_counts} - - result = { - "dataset": "locomo", - "items": len(items), - "metrics": { - "f1": sum(f1s) / max(len(f1s), 1), - "b1": sum(b1s) / max(len(b1s), 1), - "j": sum(jss) / max(len(jss), 1), - # LoCoMo 类别专用 F1 的总体 - "loc_f1": sum(loc_f1s) / max(len(loc_f1s), 1), - }, - "by_category": by_category, - "category_counts": cat_counts, - "cum_accuracy_by_category": cum_accuracy_by_category, - "context": { - "avg_tokens": (sum(per_query_context_avg_tokens) / max(len(per_query_context_avg_tokens), 1)) if per_query_context_avg_tokens else 0.0, - "avg_chars": (sum(per_query_context_chars) / max(len(per_query_context_chars), 1)) if per_query_context_chars else 0.0, - "count_avg": (sum(per_query_context_counts) / max(len(per_query_context_counts), 1)) if per_query_context_counts else 0.0, - "avg_memory_tokens": (sum(per_query_context_tokens_total) / max(len(per_query_context_tokens_total), 1)) if per_query_context_tokens_total else 0.0, - }, - "latency": { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm), - }, - "samples": samples, - "params": { - "end_user_id": end_user_id, - "search_limit": search_limit, - "context_char_budget": context_char_budget, - "search_type": search_type, - "llm_id": llm_id, - "retrieval_embedding_id": embedding_id, - "chunker_strategy": os.getenv("EVAL_CHUNKER_STRATEGY", "RecursiveChunker"), - "skip_ingest_if_exists": skip_ingest_if_exists, - "llm_timeout": llm_timeout, - "llm_max_retries": llm_max_retries, - "llm_temperature": llm_temperature, - "llm_max_tokens": llm_max_tokens - }, - "timestamp": datetime.now().isoformat() - } - if output_path: - try: - os.makedirs(os.path.dirname(output_path), exist_ok=True) - with open(output_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"✅ 结果已保存到: {output_path}") - except Exception as e: - print(f"❌ 保存结果失败: {e}") - return result - finally: - await connector.close() - - -def main(): - parser = argparse.ArgumentParser(description="Run LoCoMo evaluation with Qwen search") - parser.add_argument("--sample_size", type=int, default=1, help="Number of samples to evaluate") - parser.add_argument("--end_user_id", type=str, default=None, help="Group ID for retrieval") - parser.add_argument("--search_limit", type=int, default=8, help="Search limit per query") - parser.add_argument("--context_char_budget", type=int, default=12000, help="Max characters for context") - parser.add_argument("--llm_temperature", type=float, default=0.0, help="LLM temperature") - parser.add_argument("--llm_max_tokens", type=int, default=32, help="LLM max tokens") - parser.add_argument("--search_type", type=str, default="embedding", choices=["keyword", "embedding", "hybrid"], help="Search type") - parser.add_argument("--output_path", type=str, default=None, help="Output path for results") - parser.add_argument("--skip_ingest_if_exists", action="store_true", help="Skip ingest if group exists") - parser.add_argument("--llm_timeout", type=float, default=10.0, help="LLM timeout in seconds") - parser.add_argument("--llm_max_retries", type=int, default=1, help="LLM max retries") - args = parser.parse_args() - - load_dotenv() - - result = asyncio.run(run_locomo_eval( - sample_size=args.sample_size, - end_user_id=args.end_user_id, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - llm_temperature=args.llm_temperature, - llm_max_tokens=args.llm_max_tokens, - search_type=args.search_type, - output_path=args.output_path, - skip_ingest_if_exists=args.skip_ingest_if_exists, - llm_timeout=args.llm_timeout, - llm_max_retries=args.llm_max_retries - )) - - print("\n" + "="*50) - print("📊 最终评测结果:") - print(f" 样本数量: {result['items']}") - print(f" F1: {result['metrics']['f1']:.3f}") - print(f" BLEU-1: {result['metrics']['b1']:.3f}") - print(f" Jaccard: {result['metrics']['j']:.3f}") - print(f" LoCoMo F1: {result['metrics']['loc_f1']:.3f}") - print(f" 平均上下文长度: {result['context']['avg_chars']:.0f} 字符") - print(f" 平均检索延迟: {result['latency']['search']['mean']:.1f}ms") - print(f" 平均LLM延迟: {result['latency']['llm']['mean']:.1f}ms") - - if result['by_category']: - print("\n📈 按类别细分:") - for cat, metrics in result['by_category'].items(): - print(f" {cat}:") - print(f" 样本数: {metrics['count']}") - print(f" F1: {metrics['f1']:.3f}") - print(f" LoCoMo F1: {metrics['loc_f1']:.3f}") - print(f" Jaccard: {metrics['j']:.3f} (±{metrics['j_std']:.3f}, IQR={metrics['j_iqr']:.3f})") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py b/api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py deleted file mode 100644 index aaf46e35..00000000 --- a/api/app/core/memory/evaluation/longmemeval/longmemeval_benchmark.py +++ /dev/null @@ -1,1339 +0,0 @@ -import argparse -import asyncio -import json -import os -import time -import re -import statistics -from datetime import datetime, timedelta -from typing import List, Dict, Any -from pathlib import Path - -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.models.base import RedBearModelConfig -from app.core.memory.utils.config.config_utils import get_embedder_config -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME -from app.core.memory.evaluation.common.metrics import f1_score as common_f1, jaccard, latency_stats, avg_context_tokens -from app.core.memory.evaluation.common.metrics import exact_match - - -def load_dataset_any(path: str) -> List[Dict[str, Any]]: - """健壮地加载数据集，支持三种格式： - 1. 标准 JSON 数组: [{...}, {...}] - 2. 单个 JSON 对象: {...} - 3. JSONL 格式（每行一个 JSON）: {...}\n{...}\n{...} - """ - with open(path, "r", encoding="utf-8") as f: - content = f.read().strip() - - # 尝试标准 JSON 解析 - try: - data = json.loads(content) - if isinstance(data, list): - return [item for item in data if isinstance(item, dict)] - elif isinstance(data, dict): - return [data] - except json.JSONDecodeError: - pass - - # 尝试 JSONL 格式（每行一个 JSON 对象） - items = [] - for line in content.splitlines(): - line = line.strip() - if not line: - continue - try: - obj = json.loads(line) - if isinstance(obj, dict): - items.append(obj) - elif isinstance(obj, list): - items.extend(item for item in obj if isinstance(item, dict)) - except json.JSONDecodeError: - continue - - return items - - -def is_chinese_text(s: str) -> bool: - return bool(re.search(r"[\u4e00-\u9fff]", s or "")) - - -def build_context_from_sessions(item: Dict[str, Any]) -> List[str]: - """从数据项的 haystack_sessions 构建上下文片段。 - - 优先返回包含 has_answer 的消息 - - 其次返回拼接后的整段会话 - """ - contexts: List[str] = [] - sessions = item.get("haystack_sessions", []) or item.get("sessions", []) - for session in sessions: - parts: List[str] = [] - if isinstance(session, list): - for msg in session: - role = msg.get("role", "") - content = msg.get("content", "") or msg.get("text", "") - if content: - parts.append(f"{role}: {content}" if role else str(content)) - if msg.get("has_answer", False): - contexts.append(f"{role}: {content}" if role else str(content)) - elif isinstance(session, dict): - role = session.get("role", "") - content = session.get("content", "") or session.get("text", "") - if content: - parts.append(f"{role}: {content}" if role else str(content)) - if session.get("has_answer", False): - contexts.append(f"{role}: {content}" if role else str(content)) - if parts: - contexts.append("\n".join(parts)) - # 兜底：存在单字段上下文 - if not contexts: - single_ctx = item.get("context") or item.get("dialogue") or item.get("conversation") - if isinstance(single_ctx, str) and single_ctx.strip(): - contexts.append(single_ctx.strip()) - return contexts - - -def extract_candidate_options(question: str) -> List[str]: - """从问题中提取候选选项（A-or-B 类问题）。""" - q = (question or "").strip() - options: List[str] = [] - - # 1) 引号包裹的片段 - for pat in [r"'([^']+)'", r'\"([^\"]+)\"', r'“([^”]+)”', r'‘([^’]+)’']: - for m in re.findall(pat, q): - val = (m or "").strip() - if val: - options.append(val) - - # 2) or/还是/或者 连接词 - if len(options) < 2: - pats = [ - r"([^,;，；]+?)\s+or\s+([^,;，；\?\.!.。！]+)", - r"([^,;，；]+?)\s+还是\s+([^,;，；\?\.!.。！]+)", - r"([^,;，；]+?)\s+或者\s+([^,;，；\?\.!.。！]+)", - ] - for pat in pats: - matches = list(re.finditer(pat, q, flags=re.IGNORECASE)) - if matches: - m = matches[-1] - cand1 = m.group(1).strip().strip("?？.,，;； ") - cand2 = m.group(2).strip().strip("?？.,，;； ") - options.extend([cand1, cand2]) - break - - # 去重 - seen = set() - uniq: List[str] = [] - for o in options: - o2 = o.strip() - key = o2.lower() if not is_chinese_text(o2) else o2 - if o2 and key not in seen: - uniq.append(o2) - seen.add(key) - return uniq - - -def extract_time_entities(text: str) -> List[Dict[str, Any]]: - """增强时间实体提取，专门用于时间推理问题""" - time_entities = [] - - # 日期模式 - date_patterns = [ - (r'\b(\d{4})-(\d{1,2})-(\d{1,2})\b', 'date'), # YYYY-MM-DD - (r'\b(\d{1,2})月(\d{1,2})日\b', 'date'), # 中文日期 - (r'\b(January|February|March|April|May|June|July|August|September|October|November|December)\s+(\d{1,2}),?\s+(\d{4})?', 'date'), # 英文月份 - (r'\b(Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)\s+(\d{1,2}),?\s+(\d{4})?', 'date'), # 英文月份缩写 - ] - - # 时间间隔模式 - duration_patterns = [ - (r'(\d+)\s*天', 'days'), - (r'(\d+)\s*周', 'weeks'), - (r'(\d+)\s*个月', 'months'), - (r'(\d+)\s*年', 'years'), - (r'(\d+)\s*days?', 'days'), - (r'(\d+)\s*weeks?', 'weeks'), - (r'(\d+)\s*months?', 'months'), - (r'(\d+)\s*years?', 'years'), - ] - - # 事件时间关系模式 - temporal_relation_patterns = [ - (r'(之前|以前|前)\s*(\d+)\s*天', 'days_before'), - (r'(之后|以后|后)\s*(\d+)\s*天', 'days_after'), - (r'(\d+)\s*天\s*(之前|以前|前)', 'days_before'), - (r'(\d+)\s*天\s*(之后|以后|后)', 'days_after'), - (r'(\d+)\s*days?\s*(before|ago)', 'days_before'), - (r'(\d+)\s*days?\s*(after|later)', 'days_after'), - ] - - # 提取日期 - for pattern, entity_type in date_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'start': match.start(), - 'end': match.end() - }) - - # 提取时间间隔 - for pattern, entity_type in duration_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'value': int(match.group(1)), - 'start': match.start(), - 'end': match.end() - }) - - # 提取时间关系 - for pattern, entity_type in temporal_relation_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'value': int(match.group(2)) if match.groups() >= 2 else int(match.group(1)), - 'start': match.start(), - 'end': match.end() - }) - - return time_entities - - -def calculate_time_difference(date1: str, date2: str) -> int: - """计算两个日期之间的天数差""" - try: - # 解析日期格式 - def parse_date(date_str: str) -> datetime: - # 尝试多种日期格式 - formats = [ - '%Y-%m-%d', - '%m月%d日', - '%B %d, %Y', - '%b %d, %Y', - '%Y年%m月%d日' - ] - - for fmt in formats: - try: - return datetime.strptime(date_str, fmt) - except ValueError: - continue - - # 如果都无法解析，返回当前日期 - return datetime.now() - - d1 = parse_date(date1) - d2 = parse_date(date2) - - # 计算天数差（绝对值） - return abs((d2 - d1).days) - except Exception: - return -1 # 表示计算失败 - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: - """增强版上下文选择：特别优化时间推理问题的处理""" - if not contexts: - return "" - - # 检测是否为时间推理问题 - is_temporal_question = any(keyword in question.lower() for keyword in - ['days', 'day', 'before', 'after', 'first', '先后', '顺序', '间隔', '多久', '多少天']) - - # 提取时间实体从问题中 - question_time_entities = extract_time_entities(question) - - # 英文关键词（去停用词） - question_lower = question.lower() - stop_words = { - 'what','when','where','who','why','how','did','do','does','is','are','was','were', - 'the','a','an','and','or','but','many','which','first' - } - eng_words = [w for w in set(re.findall(r'\b\w+\b', question_lower)) - if w not in stop_words and len(w) > 2] - - # 中文片段与候选选项 - cn_tokens = generate_query_keywords_cn(question) - options = extract_candidate_options(question) - - # 时间推理问题的特殊处理 - if is_temporal_question: - # 为时间问题添加时间相关关键词 - time_keywords = ['天', '日', '月', '年', 'before', 'after', 'days', 'first', '先后'] - eng_words = [w for w in eng_words if w not in ['days', 'first']] # 避免重复 - cn_tokens.extend([kw for kw in time_keywords if kw not in cn_tokens]) - - # 限制关键词数量，优先时间相关 - tokens = time_keywords[:2] + cn_tokens[:2] + eng_words[:1] + options[:1] - else: - # 常规问题处理 - tokens = cn_tokens[:3] + options[:2] + eng_words[:1] - - # 去重 - seen = set() - final_tokens: List[str] = [] - for t in tokens: - t2 = t.strip() - if t2 and t2 not in seen: - final_tokens.append(t2) - seen.add(t2) - - scored_contexts: List[tuple[float, str]] = [] - - # 时间推理问题的权重映射 - temporal_weight_map = { - "天": 2.0, "日": 2.0, "月": 1.8, "年": 1.8, "days": 2.0, - "before": 1.5, "after": 1.5, "first": 1.5, "先后": 1.5 - } - - # 常规问题的权重映射 - normal_weight_map = { - "问题": 2.0, "故障": 2.0, "异常": 1.8, "不正常": 1.8, "坏了": 1.8, - "系统": 1.3, "GPS": 1.5, "保养": 1.4, "设备": 1.2, "模块": 1.2, "功能": 1.1 - } - - weight_map = temporal_weight_map if is_temporal_question else normal_weight_map - - for i, context in enumerate(contexts): - context_str = str(context) - lines = re.split(r'[\r\n]+', context_str) - hit_lines: List[str] = [] - kw_hits: float = 0.0 - time_entity_count = 0 - - for line in lines: - ln = line.strip() - if not ln: - continue - - has_keyword = False - # 关键词匹配 - for tok in final_tokens: - if tok and tok in ln: - w = weight_map.get(tok, 1.0) - kw_hits += ln.count(tok) * w - has_keyword = True - - # 时间实体检测（特别针对时间推理问题） - if is_temporal_question: - time_entities = extract_time_entities(ln) - time_entity_count += len(time_entities) - if time_entities: - has_keyword = True - - if has_keyword: - # 对于时间推理问题，保留包含时间信息的完整行 - hit_lines.append(ln) - - snippet = "\n".join(hit_lines) if hit_lines else context_str.strip() - - # 限制单段长度，但对时间推理问题稍微放宽限制 - max_snippet_len = 600 if is_temporal_question else 500 - if len(snippet) > max_snippet_len: - snippet = snippet[:max_snippet_len] - - # 评分逻辑 - has_number = 1 if re.search(r'\d', snippet) else 0 - has_date = 1 if (re.search(r'\b\d{4}-\d{1,2}-\d{1,2}\b', snippet) or - re.search(r'\d{1,2}月\d{1,2}日', snippet)) else 0 - - # 时间推理问题的特殊评分 - if is_temporal_question: - time_bonus = time_entity_count * 2.0 # 时间实体奖励 - temporal_coherence = 3 if (has_date and time_entity_count >= 2) else 0 - else: - time_bonus = 0 - temporal_coherence = 0 - - length_bonus = 5 if 50 < len(snippet) < 1000 else (2 if len(snippet) >= 1000 else 0) - pos_bonus = 3 if i < 3 else 0 - - score = (kw_hits * 0.8 + (has_number + has_date) * 1.5 + - length_bonus + pos_bonus + time_bonus + temporal_coherence) - - scored_contexts.append((score, snippet)) - - # 选择累计至总字符预算 - scored_contexts.sort(key=lambda x: x[0], reverse=True) - selected: List[str] = [] - total_chars = 0 - - for score, snippet in scored_contexts: - if total_chars + len(snippet) <= max_chars: - selected.append(snippet) - total_chars += len(snippet) - else: - if not selected and len(snippet) > max_chars: - selected.append(snippet[:max_chars]) - break - - final_context = "\n\n".join(selected) - - # 对于时间推理问题，添加时间计算提示 - if is_temporal_question and question_time_entities: - time_prompt = "\n\n[时间推理提示：请仔细分析上述上下文中的日期和时间关系，计算时间间隔或确定事件顺序]" - if total_chars + len(time_prompt) <= max_chars: - final_context += time_prompt - - return final_context - - -# 中文关键词提取（短语级，含数词/日期/常见领域词） -def _extract_cn_tokens(text: str) -> List[str]: - if not text: - return [] - t = str(text) - # 去掉常见功能词（粗略，不依赖分词库） - stop_words = [ - "我","我们","你","他","她","它","这","那","哪","一个","一次","一些","什么","怎么","是否","吗","呢", - "很","更","最","已经","正在","将要","马上","尽快","最近","关于","有关","以及","并且","或者","还是", - "因为","所以","如果","但是","而且","然后","之后","之前","同时","另外","并","但","却","被","把","让","给", - "和","与","跟","及","还有","就","都","在","对","对于","的","了","着","过","到","于","从","以","为","向","至","是" - ] - for sw in stop_words: - t = t.replace(sw, " ") - # 去标点 - t = re.sub(r"[，。！？、；：,.!?;:\"'（）()［］\[\]\-—…·]", " ", t) - # 基础中文片段（>=2） - base = re.findall(r"[\u4e00-\u9fff]{2,}", t) - # 特殊组合：第X次XXXX - specials = re.findall(r"第[一二三四五六七八九十]+次[\u4e00-\u9fff]{2,6}", text) - # 领域词（简单词典） - # 日期与数字 - dates = re.findall(r"\d{4}年\d{1,2}月\d{1,2}日|\d{1,2}月\d{1,2}日|\d{4}-\d{1,2}-\d{1,2}", text) - numbers = re.findall(r"\b\d+\b", text) - - tokens: List[str] = specials + base + dates + numbers - - generic = {"建议","推荐","帮助","提升","技能","有效","团队","参与度","喜欢","开始"} - tokens: List[str] = specials + base + dates + numbers - uniq: List[str] = [] - seen = set() - for tok in tokens: - tok2 = tok.strip() - if len(tok2) < 2 or len(tok2) > 6: - continue - if tok2 in generic: - continue - if tok2 not in seen: - uniq.append(tok2) - seen.add(tok2) - # 排除常见疑问型短语 - blacklist_exact = {"是什么","多少","多少天","哪个","哪些","之间","先","后","之前","之后"} - uniq2: List[str] = [u for u in uniq if u not in blacklist_exact] - return uniq2[:12] - - -# 面向检索的中文关键词生成：强调"短语、核心名词、问题/故障" -def generate_query_keywords_cn(question: str) -> List[str]: - if not question: - return [] - raw = _extract_cn_tokens(question) - core: List[str] = [] - seen = set() - - def push(x: str): - x2 = x.strip() - if not x2: - return - if 2 <= len(x2) <= 6 and x2 not in seen: - core.append(x2) - seen.add(x2) - - # 检测时间推理问题 - is_temporal = any(keyword in question for keyword in ['天', '日', 'before', 'after', 'first', '先后', '间隔']) - if is_temporal: - push("天") - push("日") - push("先后") - - # 明确优先的核心词 - if "新车" in question: - push("新车") - # 第X次保养/维修 - specials = re.findall(r"第[一二三四五六七八九十]+次[\u4e00-\u9fff]{2,6}", question) - for s in specials: - if "保养" in s or "维修" in s: - push(s) - if "保养" in question: - push("保养") - # 问题/故障类词，如题含"问题"则扩展同义词 - if "问题" in question: - for w in ["问题","故障","异常","不正常"]: - push(w) - - # 补充：从原始片段筛更短的名词短语（过滤疑问型词） - blacklist = {"是什么","多少","哪个","还是","或者","之间","先","后","之前","之后"} - for tok in raw: - if tok in blacklist: - continue - push(tok) - - # 限制数量，避免过长列表影响检索稳定性 - return core[:4] # 稍微增加限制 - - -# 通过别名匹配进行实体关键词检索（多token合并） -async def _search_entities_by_aliases(connector: Neo4jConnector, tokens: List[str], end_user_id: str | None, limit: int) -> List[Dict[str, Any]]: - results: List[Dict[str, Any]] = [] - try: - for tok in tokens: - rows = await connector.execute_query(SEARCH_ENTITIES_BY_NAME, q=tok, end_user_id=end_user_id, limit=limit) - if rows: - results.extend(rows) - except Exception: - pass - - # 按 name 去重 - deduped: List[Dict[str, Any]] = [] - seen = set() - for r in results: - k = str(r.get("name", "")) - if k and k not in seen: - deduped.append(r) - seen.add(k) - return deduped - - -# 通过对话/陈述中的entity_ids反查实体名称 -_FETCH_ENTITIES_BY_IDS = """ -MATCH (e:ExtractedEntity) -WHERE e.id IN $ids AND ($end_user_id IS NULL OR e.end_user_id = $end_user_id) -RETURN e.id AS id, e.name AS name, e.end_user_id AS end_user_id, e.entity_type AS entity_type -""" - -async def _fetch_entities_by_ids(connector: Neo4jConnector, ids: List[str], end_user_id: str | None) -> List[Dict[str, Any]]: - if not ids: - return [] - try: - rows = await connector.execute_query(_FETCH_ENTITIES_BY_IDS, ids=list({i for i in ids if i}), end_user_id=end_user_id) - return rows or [] - except Exception: - return [] - - -# 增强的时间实体检索 -_TIME_ENTITY_SEARCH = """ -MATCH (e:ExtractedEntity) -WHERE e.entity_type CONTAINS "TIME" OR e.entity_type CONTAINS "DATE" OR e.name =~ $date_pattern -AND ($end_user_id IS NULL OR e.end_user_id = $end_user_id) -RETURN e.id AS id, e.name AS name, e.end_user_id AS end_user_id, e.entity_type AS entity_type -LIMIT $limit -""" - -async def _search_time_entities(connector: Neo4jConnector, end_user_id: str | None, limit: int = 5) -> List[Dict[str, Any]]: - """专门搜索时间相关的实体""" - try: - date_pattern = r".*\d{4}.*|.*\d{1,2}月\d{1,2}日.*" - rows = await connector.execute_query(_TIME_ENTITY_SEARCH, - date_pattern=date_pattern, - end_user_id=end_user_id, - limit=limit) - return rows or [] - except Exception: - return [] - - -# 中英相对时间解析：today/昨天/上周/3天后 等简单归一化为日期 -def _resolve_relative_times_cn_en(text: str, anchor: datetime) -> str: - t = str(text) if text is not None else "" - # 英文 today/yesterday/tomorrow - t = re.sub(r"\btoday\b", anchor.date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\byesterday\b", (anchor - timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\btomorrow\b", (anchor + timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - - # 英文 X days ago / in X days - def _ago_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor - timedelta(days=n)).date().isoformat() - def _in_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor + timedelta(days=n)).date().isoformat() - t = re.sub(r"\b(\d+)\s+days\s+ago\b", _ago_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\bin\s+(\d+)\s+days\b", _in_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\blast\s+week\b", (anchor - timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\bnext\s+week\b", (anchor + timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - - # 中文 今天/昨天/明天 - t = re.sub(r"今天", anchor.date().isoformat(), t) - t = re.sub(r"昨日|昨天", (anchor - timedelta(days=1)).date().isoformat(), t) - t = re.sub(r"明天", (anchor + timedelta(days=1)).date().isoformat(), t) - # 中文 X天前 / X天后 - t = re.sub(r"(\d+)天前", lambda m: (anchor - timedelta(days=int(m.group(1)))).date().isoformat(), t) - t = re.sub(r"(\d+)天后", lambda m: (anchor + timedelta(days=int(m.group(1)))).date().isoformat(), t) - # 中文 上周 / 下周（近似7天） - t = re.sub(r"上周", (anchor - timedelta(days=7)).date().isoformat(), t) - t = re.sub(r"下周", (anchor + timedelta(days=7)).date().isoformat(), t) - # 中文 月日（无年份）补全年份 - def _md_repl(m: re.Match[str]) -> str: - mon = int(m.group(1)); day = int(m.group(2)) - return f"{anchor.year}-{mon:02d}-{day:02d}" - t = re.sub(r"(\d{1,2})月(\d{1,2})日", _md_repl, t) - return t - - -async def run_longmemeval_test( - sample_size: int = 3, - end_user_id: str | None = None, - search_limit: int = 8, - context_char_budget: int = 4000, - llm_temperature: float = 0.0, - llm_max_tokens: int = 16, - search_type: str = "hybrid", - data_path: str | None = None, - start_index: int = 0, - max_contexts_per_item: int = 2, - save_chunk_output: bool = True, - save_chunk_output_path: str | None = None, - reset_group_before_ingest: bool = False, - skip_ingest: bool = False, -) -> Dict[str, Any]: - """LongMemEval 评估测试：增强时间推理能力""" - - # Use environment variable with fallback chain - if end_user_id is None: - end_user_id = os.getenv("LONGMEMEVAL_END_USER_ID") or os.getenv("EVAL_END_USER_ID", "longmemeval_zh_bak_3") - - # 数据路径 - if not data_path: - # 固定使用中文数据集：dataset/longmemeval_oracle_zh.json - dataset_dir = Path(__file__).resolve().parent.parent / "dataset" - data_path = str(dataset_dir / "longmemeval_oracle_zh.json") - - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 longmemeval_oracle_zh.json 放置在: {dataset_dir}" - ) - - qa_list: List[Dict[str, Any]] = load_dataset_any(data_path) - # 支持评估全部样本：当 sample_size <= 0 时，取从 start_index 到末尾 - if sample_size is None or sample_size <= 0: - items = qa_list[start_index:] - else: - items = qa_list[start_index:start_index + sample_size] - - # 可选：摄入上下文（默认启用） - if not skip_ingest: - # 选择上下文并限量 - contexts: List[str] = [] - for it in items: - built = build_context_from_sessions(it) - full_transcripts = [c for c in built if "\n" in c] - evidence_msgs = [c for c in built if "\n" not in c] - selected: List[str] = [] - take_e = min(len(evidence_msgs), max_contexts_per_item) - selected.extend(evidence_msgs[:take_e]) - remain = max_contexts_per_item - len(selected) - if remain > 0 and full_transcripts: - selected.extend(full_transcripts[:remain]) - if not selected and built: - selected.append(built[0]) - contexts.extend(selected) - - print(f"📥 摄入 {len(contexts)} 个上下文到数据库") - if reset_group_before_ingest and end_user_id: - try: - _tmp_conn = Neo4jConnector() - await _tmp_conn.delete_group(end_user_id) - print(f"🧹 已清空组 {end_user_id} 的历史图数据") - except Exception as _e: - print(f"⚠️ 清空组数据失败（忽略继续）: {end_user_id} - {_e}") - finally: - try: - await _tmp_conn.close() - except Exception: - pass - _ingest_fn = ingest_contexts_via_full_pipeline - if _ingest_fn is None: - print("⚠️ 摄入函数不可用，已跳过摄入。请确认 PYTHONPATH 包含 'src' 或从项目根运行。") - else: - await _ingest_fn( - contexts, - end_user_id, - save_chunk_output=save_chunk_output, - save_chunk_output_path=save_chunk_output_path, - ) - - # 初始化组件（摄入后再初始化连接器）- 使用异步LLM客户端 - from app.db import get_db - - db = next(get_db()) - try: - llm_client = get_llm_client(os.getenv("EVAL_LLM_ID"), db) - cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID"), db) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - finally: - db.close() - - connector = Neo4jConnector() - - # 指标收集 - latencies_llm: List[float] = [] - latencies_search: List[float] = [] - per_query_context_counts: List[int] = [] - per_query_context_avg_tokens: List[float] = [] - per_query_context_chars: List[int] = [] - - type_correct: Dict[str, List[float]] = {} - type_f1: Dict[str, List[float]] = {} - type_jacc: Dict[str, List[float]] = {} - - samples: List[Dict[str, Any]] = [] - # 统计重复的上下文预览（跨样本），便于诊断"相同上下文"问题 - preview_counter: Dict[str, int] = {} - - try: - for item in items: - question = item.get("question", "") - reference = item.get("answer", "") - qtype = item.get("question_type") or item.get("type", "unknown") - - print(f"\n=== 处理问题: {question} ===") - - # 检测问题类型 - is_temporal = any(keyword in question.lower() for keyword in - ['days', 'day', 'before', 'after', 'first', '先后', '顺序', '间隔', '多久', '多少天']) - - # 检索 - t0 = time.time() - contexts_all: List[str] = [] - dialogs, statements, entities = [], [], [] - - try: - if search_type == "embedding": - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], - ) - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - - # 实体摘要（最多3个） - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - elif search_type == "keyword": - search_results = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit, - ) - chunks = search_results.get("chunks", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - summaries = search_results.get("summaries", []) - - for c in chunks: - content = str(c.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - for sm in summaries: - summary_text = str(sm.get("summary", "")).strip() - if summary_text: - contexts_all.append(summary_text) - if entities: - entity_names = [str(e.get("name", "")).strip() for e in entities[:5] if e.get("name")] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - else: # hybrid（增强版：特别优化时间推理问题） - emb_chunks, emb_statements, emb_entities, emb_summaries, emb_dialogs = [], [], [], [], [] - kw_dialogs, kw_statements, kw_entities = [], [], [] - - # 1) 嵌入检索 - try: - emb_res = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], - ) - if isinstance(emb_res, dict): - emb_chunks = emb_res.get("chunks", []) or [] - emb_statements = emb_res.get("statements", []) or [] - emb_entities = emb_res.get("entities", []) or [] - emb_summaries = emb_res.get("summaries", []) or [] - emb_dialogs = emb_res.get("dialogues", []) or [] - except Exception as e: - print(f"⚠️ 嵌入检索失败，将继续进行关键词检索: {e}") - - # 2) 关键词检索（增强版） - try: - kw_res = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit, - ) - if isinstance(kw_res, dict): - kw_dialogs = kw_res.get("dialogues", []) or [] - kw_statements = kw_res.get("statements", []) or [] - kw_entities = kw_res.get("entities", []) or [] - - # 时间推理问题的特殊处理 - if is_temporal: - # 专门搜索时间实体 - time_entities = await _search_time_entities(connector, end_user_id, search_limit//2) - if time_entities: - kw_entities.extend(time_entities) - # 添加时间相关关键词检索 - time_keywords = ['天', '日', '月', '年', 'before', 'after', 'first'] - for tk in time_keywords: - try: - time_res = await search_graph( - connector=connector, - q=tk, - end_user_id=end_user_id, - limit=2, - ) - if isinstance(time_res, dict): - kw_dialogs.extend(time_res.get("dialogues", []) or []) - kw_statements.extend(time_res.get("statements", []) or []) - except Exception: - pass - - # 中文关键词拆分后做别名匹配 - cn_tokens = _extract_cn_tokens(question) - alias_entities = await _search_entities_by_aliases(connector, cn_tokens, end_user_id, search_limit) - if alias_entities: - kw_entities.extend(alias_entities) - - # 从对话/陈述中的 entity_ids 反查实体 - ids = [] - try: - for d in kw_dialogs: - ids.extend(d.get("entity_ids", []) or []) - for s in kw_statements: - ids.extend(s.get("entity_ids", []) or []) - except Exception: - pass - if ids: - id_entities = await _fetch_entities_by_ids(connector, ids, end_user_id) - if id_entities: - kw_entities.extend(id_entities) - - # 多关键词检索 - try: - eng_words = [w for w in set(re.findall(r"\b\w+\b", question.lower())) if len(w) > 2] - kw_list = generate_query_keywords_cn(question)[:3] + eng_words[:1] - for kw in kw_list: - if not kw: - continue - sub_res = await search_graph( - connector=connector, - q=str(kw), - end_user_id=end_user_id, - limit=max(3, search_limit // 2), - ) - if isinstance(sub_res, dict): - kw_dialogs.extend(sub_res.get("dialogues", []) or []) - kw_statements.extend(sub_res.get("statements", []) or []) - kw_entities.extend(sub_res.get("entities", []) or []) - except Exception: - pass - - # 选项参与关键词检索 - try: - opt_list = extract_candidate_options(question)[:2] - for opt in opt_list: - if not opt: - continue - opt_res = await search_graph( - connector=connector, - q=str(opt), - end_user_id=end_user_id, - limit=max(3, search_limit // 2), - ) - if isinstance(opt_res, dict): - kw_dialogs.extend(opt_res.get("dialogues", []) or []) - kw_statements.extend(opt_res.get("statements", []) or []) - kw_entities.extend(opt_res.get("entities", []) or []) - except Exception: - pass - except Exception as e: - print(f"❌ 关键词检索失败: {e}") - - # 3) 合并、排序并去重 - all_dialogs = emb_dialogs + kw_dialogs - all_statements = emb_statements + kw_statements - all_entities = emb_entities + kw_entities - - def dedup(items: List[Dict[str, Any]], key_field: str = "uuid") -> List[Dict[str, Any]]: - seen = set() - out = [] - for it in items: - key = str(it.get(key_field, "")) + str(it.get("content", "") + str(it.get("statement", ""))) - if key not in seen: - out.append(it) - seen.add(key) - return out - - # 时间推理问题优先排序包含时间信息的文档 - if is_temporal: - def temporal_score(item: Dict[str, Any]) -> float: - base_score = float(item.get("score", 0.0)) - content = str(item.get("content", "") + str(item.get("statement", ""))) - time_entities = extract_time_entities(content) - time_bonus = len(time_entities) * 0.5 - return base_score + time_bonus - - dialogs = dedup(sorted(all_dialogs, key=temporal_score, reverse=True)) - statements = dedup(sorted(all_statements, key=temporal_score, reverse=True)) - else: - dialogs = dedup(sorted(all_dialogs, key=lambda d: float(d.get("score", 0.0)), reverse=True)) - statements = dedup(sorted(all_statements, key=lambda s: float(s.get("score", 0.0)), reverse=True)) - - entities = dedup(all_entities, key_field="name") - - # 4) 构建上下文 - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - # 实体摘要 - try: - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - except Exception: - pass - - # 全局回退 - if not contexts_all and search_type in ("embedding", "hybrid"): - try: - print("🔁 检索为空，回退到关键词检索...") - kw_fallback = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=max(search_limit, 5), - ) - fb_dialogs = kw_fallback.get("dialogues", []) or [] - fb_statements = kw_fallback.get("statements", []) or [] - fb_entities = kw_fallback.get("entities", []) or [] - - for d in fb_dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in fb_statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - if fb_entities: - entity_names = [str(e.get("name", "")).strip() for e in fb_entities[:5] if e.get("name")] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - dialogs = fb_dialogs if fb_dialogs else dialogs - statements = fb_statements if fb_statements else statements - entities = fb_entities if fb_entities else entities - print(f"↩️ 回退到关键词检索: {len(fb_dialogs)} 对话, {len(fb_statements)} 条陈述, {len(fb_entities)} 个实体") - except Exception as fe: - print(f"❌ 关键词回退失败: {fe}") - - ent_count = len(entities) if isinstance(entities, list) else 0 - print(f"✅ {search_type}检索成功: {len(dialogs)} 对话, {len(statements)} 条陈述, {ent_count} 个实体") - if is_temporal: - print("⏰ 检测为时间推理问题，已启用时间优化检索") - - except Exception as e: - print(f"❌ {search_type}检索失败: {e}") - contexts_all = [] - - t1 = time.time() - latencies_search.append((t1 - t0) * 1000) - - # 智能上下文选择 - context_text = "" - if contexts_all: - context_text = smart_context_selection(contexts_all, question, max_chars=context_char_budget) - # 相对时间解析 - try: - context_text = _resolve_relative_times_cn_en(context_text, anchor=datetime.now()) - except Exception: - pass - # 诊断信息 - try: - cn_diag = generate_query_keywords_cn(question)[:3] - opts = extract_candidate_options(question)[:2] - qlw = [w for w in set(re.findall(r'\b\w+\b', question.lower())) if len(w) > 2][:1] - diag_tokens: List[str] = [] - for t in cn_diag + opts + qlw: - if t and t not in diag_tokens: - diag_tokens.append(t) - print(f"🔍 关键词/选项: {', '.join(diag_tokens)}") - preview = context_text[:200].replace('\n', ' ') - print(f"🔎 上下文预览: {preview}...") - key_preview = preview.strip() - if key_preview: - preview_counter[key_preview] = preview_counter.get(key_preview, 0) + 1 - except Exception: - pass - else: - print("❌ 没有检索到有效上下文") - context_text = "No relevant context found." - - # 记录上下文诊断信息 - per_query_context_counts.append(len(contexts_all)) - per_query_context_avg_tokens.append(avg_context_tokens([context_text])) - per_query_context_chars.append(len(context_text)) - - # LLM 推理（增强时间推理提示） - options = extract_candidate_options(question) - if len(options) >= 2: - opt_lines = "\n".join(f"- {o}" for o in options) - # 时间推理问题的特殊提示 - if is_temporal: - system_prompt = ( - "You are a QA assistant specializing in temporal reasoning. Analyze the dates and time relationships in the context carefully. " - "Return ONLY one string: exactly one option from the provided candidates. If the context is insufficient, respond with 'Unknown'. " - "Pay special attention to date sequences and time intervals." - ) - else: - system_prompt = ( - "You are a QA assistant. Respond in the same language as the question. Return ONLY one string: exactly one option from the provided candidates. " - "If the context is insufficient, respond with 'Unknown'. If the context expresses a synonym or paraphrase of a candidate, return the closest candidate. " - "Do not include explanations." - ) - - messages = [ - {"role": "system", "content": system_prompt}, - { - "role": "user", - "content": ( - f"Question: {question}\n\nCandidates:\n{opt_lines}\n\nContext:\n{context_text}\n\nReturn EXACTLY one candidate string (or 'Unknown')." - ), - }, - ] - else: - # 时间推理问题的特殊提示 - if is_temporal: - system_prompt = ( - "You are a QA assistant specializing in temporal reasoning. Analyze the dates and time relationships in the context carefully. " - "If the context contains the answer, return a concise answer phrase focusing on temporal information. " - "If the answer cannot be determined from the context, respond with 'Unknown'. Return ONLY the final answer string, no explanations." - ) - else: - system_prompt = ( - "You are a QA assistant. Respond in the same language as the question. If the context contains the answer, return a concise answer phrase. " - "If the answer cannot be determined from the context, respond with 'Unknown'. Return ONLY the final answer string, no explanations." - ) - - messages = [ - {"role": "system", "content": system_prompt}, - { - "role": "user", - "content": f"Question: {question}\n\nContext:\n{context_text}\n\nReturn ONLY the answer (or 'Unknown').", - }, - ] - - t2 = time.time() - # 使用异步调用 - resp = await llm_client.chat(messages=messages) - t3 = time.time() - latencies_llm.append((t3 - t2) * 1000) - - # 兼容不同的响应格式 - pred_raw = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else "Unknown") - - # 选项题输出规范化 - pred = pred_raw - if len(options) >= 2 and not pred_raw.lower().startswith("unknown"): - def _basic_norm(s: str) -> str: - s = s.lower().strip() - return re.sub(r"[^\w\s]", " ", s) - def _jaccard(a: str, b: str) -> float: - ta = set(t for t in _basic_norm(a).split() if t) - tb = set(t for t in _basic_norm(b).split() if t) - if not ta and not tb: - return 1.0 - if not ta or not tb: - return 0.0 - return len(ta & tb) / len(ta | tb) - best = None - best_score = -1.0 - for o in options: - score = _jaccard(pred_raw, o) - if score > best_score: - best = o - best_score = score - if best is not None and best_score > 0.0: - pred = best - - # 指标 - flag = exact_match(pred, reference) - f1_val = common_f1(str(pred), str(reference)) - j_val = jaccard(str(pred), str(reference)) - - type_correct.setdefault(qtype, []).append(flag) - type_f1.setdefault(qtype, []).append(f1_val) - type_jacc.setdefault(qtype, []).append(j_val) - - samples.append({ - "question": question, - "prediction": pred, - "answer": reference, - "question_type": qtype, - "is_temporal": is_temporal, - "question_id": item.get("question_id"), - "options": options, - "context_count": len(contexts_all), - "context_chars": len(context_text), - "retrieved_dialogue_count": len(dialogs), - "retrieved_statement_count": len(statements), - "metrics": { - "exact_match": bool(flag), - "f1": f1_val, - "jaccard": j_val - }, - "timing": { - "search_ms": (t1 - t0) * 1000, - "llm_ms": (t3 - t2) * 1000 - } - }) - - print(f"🤖 LLM 回答: {pred}") - print(f"✅ 正确答案: {reference}") - print(f"📈 当前指标 - Exact Match: {flag}, F1: {f1_val:.3f}, Jaccard: {j_val:.3f}") - - # 聚合结果 - type_acc = {t: (sum(v) / max(len(v), 1)) for t, v in type_correct.items()} - f1_by_type = {t: (sum(v) / max(len(v), 1)) for t, v in type_f1.items()} - jacc_by_type = {t: (sum(v) / max(len(v), 1)) for t, v in type_jacc.items()} - - result = { - "dataset": "longmemeval", - "items": len(items), - "accuracy_by_type": type_acc, - "f1_by_type": f1_by_type, - "jaccard_by_type": jacc_by_type, - "samples": samples, - "latency": { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm), - }, - "context": { - "avg_tokens": statistics.mean(per_query_context_avg_tokens) if per_query_context_avg_tokens else 0.0, - "avg_chars": statistics.mean(per_query_context_chars) if per_query_context_chars else 0.0, - "count_avg": statistics.mean(per_query_context_counts) if per_query_context_counts else 0.0, - }, - "params": { - "end_user_id": end_user_id, - "search_limit": search_limit, - "context_char_budget": context_char_budget, - "search_type": search_type, - "llm_id": os.getenv("EVAL_LLM_ID"), - "embedding_id": os.getenv("EVAL_EMBEDDING_ID"), - "sample_size": sample_size, - "start_index": start_index, - }, - "timestamp": datetime.now().isoformat() - } - - # 计算汇总指标 - try: - total_items = max(len(samples), 1) - correct_count = sum(1 for s in samples if s.get("metrics", {}).get("exact_match")) - score_accuracy = (correct_count / total_items) * 100.0 - - total_latencies_ms = [] - for s in samples: - t = s.get("timing", {}) - total_latencies_ms.append(float(t.get("search_ms", 0.0)) + float(t.get("llm_ms", 0.0))) - total_lat_stats = latency_stats(total_latencies_ms) if total_latencies_ms else {"p50": 0.0, "iqr": 0.0} - latency_median_s = total_lat_stats.get("p50", 0.0) / 1000.0 - latency_iqr_s = total_lat_stats.get("iqr", 0.0) / 1000.0 - - avg_ctx_tokens = statistics.mean(per_query_context_avg_tokens) if per_query_context_avg_tokens else 0.0 - avg_ctx_tokens_k = avg_ctx_tokens / 1000.0 - - result["metric_summary"] = { - "score_accuracy": score_accuracy, - "latency_median_s": latency_median_s, - "latency_iqr_s": latency_iqr_s, - "avg_context_tokens_k": avg_ctx_tokens_k, - } - except Exception: - result["metric_summary"] = { - "score_accuracy": 0.0, - "latency_median_s": 0.0, - "latency_iqr_s": 0.0, - "avg_context_tokens_k": 0.0, - } - - # 诊断信息 - try: - dups = sorted([(k, c) for k, c in preview_counter.items() if c > 1], key=lambda x: -x[1])[:5] - result["diagnostics"] = { - "duplicate_previews_top": [{"count": c, "preview": k[:120]} for k, c in dups], - "unique_preview_count": len(preview_counter), - } - except Exception: - pass - - return result - - finally: - await connector.close() - -def main(): - load_dotenv() - parser = argparse.ArgumentParser(description="LongMemEval 评估测试脚本（增强时间推理版）") - parser.add_argument("--sample-size", type=int, default=3, help="样本数量（<=0 表示全部）") - parser.add_argument("--all", action="store_true", help="评估全部样本（覆盖 --sample-size）") - parser.add_argument("--start-index", type=int, default=0, help="起始样本索引") - parser.add_argument("--end-user-id", type=str, default=None, help="图数据库 End User ID，默认使用环境变量") - parser.add_argument("--search-limit", type=int, default=8, help="检索条数上限") - parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") - parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") - parser.add_argument("--llm-max-tokens", type=int, default=16, help="LLM 最大输出 token") - parser.add_argument("--search-type", type=str, default="hybrid", choices=["embedding","keyword","hybrid"], help="检索类型") - parser.add_argument("--data-path", type=str, default=None, help="数据集路径") - parser.add_argument("--max-contexts-per-item", type=int, default=2, help="每条样本最多摄入的上下文段数") - parser.add_argument("--no-save-chunk-output", action="store_true", help="不保存分块结果（默认保存）") - parser.add_argument("--save-chunk-output-path", type=str, default=None, help="自定义分块输出路径") - parser.add_argument("--reset-group-before-ingest", action="store_true", help="摄入前清空该 Group 在图数据库中的历史数据") - parser.add_argument("--skip-ingest", action="store_true", help="跳过摄入，仅检索评估") - args = parser.parse_args() - - sample_size = 0 if args.all else args.sample_size - - result = asyncio.run( - run_longmemeval_test( - sample_size=sample_size, - end_user_id=args.end_user_id, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - llm_temperature=args.llm_temperature, - llm_max_tokens=args.llm_max_tokens, - search_type=args.search_type, - data_path=args.data_path, - start_index=args.start_index, - max_contexts_per_item=args.max_contexts_per_item, - save_chunk_output=(not args.no_save_chunk_output), - save_chunk_output_path=args.save_chunk_output_path, - reset_group_before_ingest=args.reset_group_before_ingest, - skip_ingest=args.skip_ingest, - ) - ) - - # 打印结果 - print("\n" + "="*50) - print("📊 LongMemEval 测试结果:") - print(f" 样本数量: {result['items']}") - - if result['accuracy_by_type']: - print("\n📈 按问题类型细分:") - for qtype, acc in result['accuracy_by_type'].items(): - print(f" {qtype}:") - print(f" Score (Accuracy): {acc:.3f}") - - print(f"\n📊 指标总览:") - ms = result.get('metric_summary', {}) - print(f" Score (Accuracy): {ms.get('score_accuracy', 0.0):.1f}%") - print(f" Latency (s): median {ms.get('latency_median_s', 0.0):.3f}s") - print(f" Latency IQR (s): {ms.get('latency_iqr_s', 0.0):.3f}s") - print(f" Avg Context Tokens (k): {ms.get('avg_context_tokens_k', 0.0):.3f}k") - - print(f"\n⏱️ 细分性能指标:") - print(f" 检索延迟(均值): {result['latency']['search']['mean']:.1f}ms") - print(f" LLM延迟(均值): {result['latency']['llm']['mean']:.1f}ms") - print(f" 上下文长度(均值): {result['context']['avg_chars']:.0f} 字符") - - - # 保存结果到文件 - try: - # 使用相对路径而不是 PROJECT_ROOT - out_dir = Path(__file__).resolve().parent / "results" - os.makedirs(out_dir, exist_ok=True) - ts = datetime.now().strftime("%Y%m%d_%H%M%S") - out_path = os.path.join(out_dir, f"longmemeval_{result['params']['search_type']}_{ts}.json") - with open(out_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n💾 结果已保存: {out_path}") - except Exception as e: - print(f"⚠️ 结果保存失败: {e}") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/longmemeval/test_eval.py b/api/app/core/memory/evaluation/longmemeval/test_eval.py deleted file mode 100644 index 08daa890..00000000 --- a/api/app/core/memory/evaluation/longmemeval/test_eval.py +++ /dev/null @@ -1,1312 +0,0 @@ -import argparse -import asyncio -import json -import os -import time -import re -import statistics -from datetime import datetime, timedelta -from typing import List, Dict, Any -from pathlib import Path - -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -# 与现有评估脚本保持一致的导入方式 -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.models.base import RedBearModelConfig -from app.core.memory.utils.config.config_utils import get_embedder_config -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.dialogue_queries import SEARCH_ENTITIES_BY_NAME -from app.core.memory.evaluation.common.metrics import f1_score as common_f1, jaccard, latency_stats, avg_context_tokens -from app.core.memory.evaluation.common.metrics import exact_match - - -def load_dataset_any(path: str) -> List[Dict[str, Any]]: - """健壮地加载数据集，支持三种格式： - 1. 标准 JSON 数组: [{...}, {...}] - 2. 单个 JSON 对象: {...} - 3. JSONL 格式（每行一个 JSON）: {...}\n{...}\n{...} - """ - with open(path, "r", encoding="utf-8") as f: - content = f.read().strip() - - # 尝试标准 JSON 解析 - try: - data = json.loads(content) - if isinstance(data, list): - return [item for item in data if isinstance(item, dict)] - elif isinstance(data, dict): - return [data] - except json.JSONDecodeError: - pass - - # 尝试 JSONL 格式（每行一个 JSON 对象） - items = [] - for line in content.splitlines(): - line = line.strip() - if not line: - continue - try: - obj = json.loads(line) - if isinstance(obj, dict): - items.append(obj) - elif isinstance(obj, list): - items.extend(item for item in obj if isinstance(item, dict)) - except json.JSONDecodeError: - continue - - return items - - -def is_chinese_text(s: str) -> bool: - return bool(re.search(r"[\u4e00-\u9fff]", s or "")) - - -def extract_candidate_options(question: str) -> List[str]: - """从问题中提取候选选项（A-or-B 类问题）。""" - q = (question or "").strip() - options: List[str] = [] - - # 1) 引号包裹的片段 - for pat in [r"'([^']+)'", r'\"([^\"]+)\"', r'“([^”]+)”', r'‘([^’]+)’']: - for m in re.findall(pat, q): - val = (m or "").strip() - if val: - options.append(val) - - # 2) or/还是/或者 连接词 - if len(options) < 2: - pats = [ - r"([^,;，；]+?)\s+or\s+([^,;，；\?\.!.。！]+)", - r"([^,;，；]+?)\s+还是\s+([^,;，；\?\.!.。！]+)", - r"([^,;，；]+?)\s+或者\s+([^,;，；\?\.!.。！]+)", - ] - for pat in pats: - matches = list(re.finditer(pat, q, flags=re.IGNORECASE)) - if matches: - m = matches[-1] - cand1 = m.group(1).strip().strip("?？.,，;； ") - cand2 = m.group(2).strip().strip("?？.,，;； ") - options.extend([cand1, cand2]) - break - - # 去重 - seen = set() - uniq: List[str] = [] - for o in options: - o2 = o.strip() - key = o2.lower() if not is_chinese_text(o2) else o2 - if o2 and key not in seen: - uniq.append(o2) - seen.add(key) - return uniq - - -def extract_time_entities(text: str) -> List[Dict[str, Any]]: - """增强时间实体提取，专门用于时间推理问题""" - time_entities = [] - - # 日期模式 - date_patterns = [ - (r'\b(\d{4})-(\d{1,2})-(\d{1,2})\b', 'date'), # YYYY-MM-DD - (r'\b(\d{1,2})月(\d{1,2})日\b', 'date'), # 中文日期 - (r'\b(January|February|March|April|May|June|July|August|September|October|November|December)\s+(\d{1,2}),?\s+(\d{4})?', 'date'), # 英文月份 - (r'\b(Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)\s+(\d{1,2}),?\s+(\d{4})?', 'date'), # 英文月份缩写 - ] - - # 时间间隔模式 - duration_patterns = [ - (r'(\d+)\s*天', 'days'), - (r'(\d+)\s*周', 'weeks'), - (r'(\d+)\s*个月', 'months'), - (r'(\d+)\s*年', 'years'), - (r'(\d+)\s*days?', 'days'), - (r'(\d+)\s*weeks?', 'weeks'), - (r'(\d+)\s*months?', 'months'), - (r'(\d+)\s*years?', 'years'), - ] - - # 事件时间关系模式 - temporal_relation_patterns = [ - (r'(之前|以前|前)\s*(\d+)\s*天', 'days_before'), - (r'(之后|以后|后)\s*(\d+)\s*天', 'days_after'), - (r'(\d+)\s*天\s*(之前|以前|前)', 'days_before'), - (r'(\d+)\s*天\s*(之后|以后|后)', 'days_after'), - (r'(\d+)\s*days?\s*(before|ago)', 'days_before'), - (r'(\d+)\s*days?\s*(after|later)', 'days_after'), - ] - - # 提取日期 - for pattern, entity_type in date_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'start': match.start(), - 'end': match.end() - }) - - # 提取时间间隔 - for pattern, entity_type in duration_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'value': int(match.group(1)), - 'start': match.start(), - 'end': match.end() - }) - - # 提取时间关系 - for pattern, entity_type in temporal_relation_patterns: - matches = re.finditer(pattern, text, re.IGNORECASE) - for match in matches: - time_entities.append({ - 'text': match.group(), - 'type': entity_type, - 'value': int(match.group(2)) if match.groups() >= 2 else int(match.group(1)), - 'start': match.start(), - 'end': match.end() - }) - - return time_entities - - -def calculate_time_difference(date1: str, date2: str) -> int: - """计算两个日期之间的天数差""" - try: - # 解析日期格式 - def parse_date(date_str: str) -> datetime: - # 尝试多种日期格式 - formats = [ - '%Y-%m-%d', - '%m月%d日', - '%B %d, %Y', - '%b %d, %Y', - '%Y年%m月%d日' - ] - - for fmt in formats: - try: - return datetime.strptime(date_str, fmt) - except ValueError: - continue - - # 如果都无法解析，返回当前日期 - return datetime.now() - - d1 = parse_date(date1) - d2 = parse_date(date2) - - # 计算天数差（绝对值） - return abs((d2 - d1).days) - except Exception: - return -1 # 表示计算失败 - - -def _extract_cn_tokens(text: str) -> List[str]: - """中文关键词提取（短语级，含数词/日期/常见领域词）""" - if not text: - return [] - t = str(text) - # 去掉常见功能词（粗略，不依赖分词库） - stop_words = [ - "我","我们","你","他","她","它","这","那","哪","一个","一次","一些","什么","怎么","是否","吗","呢", - "很","更","最","已经","正在","将要","马上","尽快","最近","关于","有关","以及","并且","或者","还是", - "因为","所以","如果","但是","而且","然后","之后","之前","同时","另外","并","但","却","被","把","让","给", - "和","与","跟","及","还有","就","都","在","对","对于","的","了","着","过","到","于","从","以","为","向","至","是" - ] - for sw in stop_words: - t = t.replace(sw, " ") - # 去标点 - t = re.sub(r"[，。！？、；：,.!?;:\"'（）()［］\[\]\-—…·]", " ", t) - # 基础中文片段（>=2） - base = re.findall(r"[\u4e00-\u9fff]{2,}", t) - # 特殊组合：第X次XXXX - specials = re.findall(r"第[一二三四五六七八九十]+次[\u4e00-\u9fff]{2,6}", text) - # 日期与数字 - dates = re.findall(r"\d{4}年\d{1,2}月\d{1,2}日|\d{1,2}月\d{1,2}日|\d{4}-\d{1,2}-\d{1,2}", text) - numbers = re.findall(r"\b\d+\b", text) - - generic = {"建议","推荐","帮助","提升","技能","有效","团队","参与度","喜欢","开始"} - tokens: List[str] = specials + base + dates + numbers - uniq: List[str] = [] - seen = set() - for tok in tokens: - tok2 = tok.strip() - if len(tok2) < 2 or len(tok2) > 6: - continue - if tok2 in generic: - continue - if tok2 not in seen: - uniq.append(tok2) - seen.add(tok2) - # 排除常见疑问型短语 - blacklist_exact = {"是什么","多少","多少天","哪个","哪些","之间","先","后","之前","之后"} - uniq2: List[str] = [u for u in uniq if u not in blacklist_exact] - return uniq2[:12] - - -def generate_query_keywords_cn(question: str) -> List[str]: - """增强版关键词提取，特别关注技术术语和专有名词""" - if not question: - return [] - - # 提取专有名词（带引号的内容） - quoted_terms = re.findall(r'["""]([^"""]+)["""]', question) - - # 提取技术术语（中英文混合） - tech_terms = re.findall(r'[A-Z][a-zA-Z]+\s+[A-Z][a-zA-Z]+|[A-Za-z]+[\u4e00-\u9fff]+|[\u4e00-\u9fff]+[A-Za-z]+', question) - - # 提取核心名词短语 - core_nouns = re.findall(r'[\u4e00-\u9fff]{2,5}系统|[\u4e00-\u9fff]{2,5}管理|[\u4e00-\u9fff]{2,5}分析|[\u4e00-\u9fff]{2,5}工作坊|[\u4e00-\u9fff]{2,5}研讨会', question) - - # 基础中文片段 - base_tokens = _extract_cn_tokens(question) - - # 特定领域关键词增强 - domain_keywords = [] - # GPS相关 - if any(term in question for term in ["GPS", "导航", "定位系统", "系统运行"]): - domain_keywords.extend(["GPS", "导航系统", "定位", "系统故障", "功能异常"]) - # 活动相关 - if any(term in question for term in ["工作坊", "研讨会", "网络研讨会", "活动"]): - domain_keywords.extend(["工作坊", "研讨会", "参加", "参与", "活动"]) - # 时间顺序相关 - if any(term in question for term in ["先", "后", "第一个", "之前", "首先"]): - domain_keywords.extend(["先", "后", "之前", "之后", "第一次", "首先"]) - # 设备相关 - if any(term in question for term in ["设备", "手机", "电脑", "笔记本电脑"]): - domain_keywords.extend(["设备", "手机", "电脑", "笔记本电脑", "购买"]) - - # 合并并去重 - all_tokens = quoted_terms + tech_terms + core_nouns + base_tokens + domain_keywords - seen = set() - final_tokens = [] - - for token in all_tokens: - token = token.strip() - if len(token) >= 2 and token not in seen: - final_tokens.append(token) - seen.add(token) - - return final_tokens[:8] - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: - """增强版上下文选择：特别优化技术术语和精确匹配""" - if not contexts: - return "" - - # 检测是否为时间推理问题 - is_temporal_question = any(keyword in question.lower() for keyword in - ['days', 'day', 'before', 'after', 'first', '先后', '顺序', '间隔', '多久', '多少天']) - - # 提取时间实体从问题中 - question_time_entities = extract_time_entities(question) - - # 提取关键技术实体 - key_entities = [] - # GPS相关 - if any(term in question for term in ["GPS", "导航", "定位系统", "系统运行"]): - key_entities.extend(["GPS", "导航", "定位", "系统", "功能", "问题", "故障"]) - # 活动相关 - if any(term in question for term in ["工作坊", "研讨会", "网络研讨会", "活动"]): - key_entities.extend(["工作坊", "研讨会", "参加", "参与", "活动", "时间"]) - # 时间顺序相关 - if any(term in question for term in ["先", "后", "第一个", "之前", "首先"]): - key_entities.extend(["先", "后", "之前", "之后", "第一次", "首先"]) - - # 英文关键词（去停用词） - question_lower = question.lower() - stop_words = { - 'what','when','where','who','why','how','did','do','does','is','are','was','were', - 'the','a','an','and','or','but','many','which','first' - } - eng_words = [w for w in set(re.findall(r'\b\w+\b', question_lower)) - if w not in stop_words and len(w) > 2] - - # 中文片段与候选选项 - cn_tokens = generate_query_keywords_cn(question) - options = extract_candidate_options(question) - - # 时间推理问题的特殊处理 - if is_temporal_question: - # 为时间问题添加时间相关关键词 - time_keywords = ['天', '日', '月', '年', 'before', 'after', 'days', 'first', '先后'] - eng_words = [w for w in eng_words if w not in ['days', 'first']] # 避免重复 - cn_tokens.extend([kw for kw in time_keywords if kw not in cn_tokens]) - - # 限制关键词数量，优先时间相关 - tokens = time_keywords[:2] + key_entities[:3] + cn_tokens[:2] + eng_words[:1] + options[:1] - else: - # 常规问题处理，优先关键技术实体 - tokens = key_entities[:4] + cn_tokens[:3] + options[:2] + eng_words[:1] - - # 去重 - seen = set() - final_tokens: List[str] = [] - for t in tokens: - t2 = t.strip() - if t2 and t2 not in seen: - final_tokens.append(t2) - seen.add(t2) - - scored_contexts: List[tuple[float, str]] = [] - - # 关键技术实体权重映射 - key_entity_weights = { - "GPS": 3.0, "导航": 2.5, "系统": 2.0, "功能": 2.0, "问题": 2.0, "故障": 2.5, - "工作坊": 2.5, "研讨会": 2.5, "参加": 2.0, "参与": 2.0, - "先": 2.0, "后": 2.0, "之前": 2.0, "之后": 2.0, "第一次": 2.5 - } - - # 时间推理问题的权重映射 - temporal_weight_map = { - "天": 2.0, "日": 2.0, "月": 1.8, "年": 1.8, "days": 2.0, - "before": 1.5, "after": 1.5, "first": 1.5, "先后": 1.5 - } - - # 常规问题的权重映射 - normal_weight_map = { - "问题": 2.0, "故障": 2.0, "异常": 1.8, "不正常": 1.8, "坏了": 1.8, - "系统": 1.3, "GPS": 1.5, "保养": 1.4, "设备": 1.2, "模块": 1.2, "功能": 1.1 - } - - # 合并权重映射 - weight_map = {**normal_weight_map, **temporal_weight_map, **key_entity_weights} - - for i, context in enumerate(contexts): - context_str = str(context) - lines = re.split(r'[\r\n]+', context_str) - hit_lines: List[str] = [] - kw_hits: float = 0.0 - time_entity_count = 0 - key_entity_hits = 0 - - for line in lines: - ln = line.strip() - if not ln: - continue - - has_keyword = False - # 关键词匹配 - for tok in final_tokens: - if tok and tok in ln: - w = weight_map.get(tok, 1.0) - hit_count = ln.count(tok) - kw_hits += hit_count * w - # 关键技术实体额外奖励 - if tok in key_entity_weights: - key_entity_hits += hit_count - has_keyword = True - - # 时间实体检测（特别针对时间推理问题） - if is_temporal_question: - time_entities = extract_time_entities(ln) - time_entity_count += len(time_entities) - if time_entities: - has_keyword = True - - # 精确匹配奖励（完整问题关键词出现在上下文中） - for q_word in question.split(): - if len(q_word) > 3 and q_word in ln: - kw_hits += 0.5 # 精确匹配奖励 - - if has_keyword: - # 对于包含关键信息的行，保留完整行 - hit_lines.append(ln) - - snippet = "\n".join(hit_lines) if hit_lines else context_str.strip() - - # 限制单段长度，但对包含关键信息的上下文稍微放宽限制 - max_snippet_len = 600 if (key_entity_hits > 0 or time_entity_count > 0) else 500 - if len(snippet) > max_snippet_len: - snippet = snippet[:max_snippet_len] - - # 评分逻辑 - has_number = 1 if re.search(r'\d', snippet) else 0 - has_date = 1 if (re.search(r'\b\d{4}-\d{1,2}-\d{1,2}\b', snippet) or - re.search(r'\d{1,2}月\d{1,2}日', snippet)) else 0 - - # 关键技术实体奖励 - key_entity_bonus = key_entity_hits * 1.0 - - # 时间推理问题的特殊评分 - if is_temporal_question: - time_bonus = time_entity_count * 2.0 # 时间实体奖励 - temporal_coherence = 3 if (has_date and time_entity_count >= 2) else 0 - else: - time_bonus = 0 - temporal_coherence = 0 - - length_bonus = 5 if 50 < len(snippet) < 1000 else (2 if len(snippet) >= 1000 else 0) - pos_bonus = 3 if i < 3 else 0 - - score = (kw_hits * 0.8 + (has_number + has_date) * 1.5 + - length_bonus + pos_bonus + time_bonus + temporal_coherence + key_entity_bonus) - - scored_contexts.append((score, snippet)) - - # 选择累计至总字符预算 - scored_contexts.sort(key=lambda x: x[0], reverse=True) - selected: List[str] = [] - total_chars = 0 - - for score, snippet in scored_contexts: - if total_chars + len(snippet) <= max_chars: - selected.append(snippet) - total_chars += len(snippet) - else: - if not selected and len(snippet) > max_chars: - selected.append(snippet[:max_chars]) - break - - final_context = "\n\n".join(selected) - - # 对于时间推理问题，添加时间计算提示 - if is_temporal_question and question_time_entities: - time_prompt = "\n\n[时间推理提示：请仔细分析上述上下文中的日期和时间关系，计算时间间隔或确定事件顺序]" - if total_chars + len(time_prompt) <= max_chars: - final_context += time_prompt - - return final_context - - -# 通过别名匹配进行实体关键词检索（多token合并） -async def _search_entities_by_aliases(connector: Neo4jConnector, tokens: List[str], end_user_id: str | None, limit: int) -> List[Dict[str, Any]]: - results: List[Dict[str, Any]] = [] - try: - for tok in tokens: - rows = await connector.execute_query(SEARCH_ENTITIES_BY_NAME, q=tok, end_user_id=end_user_id, limit=limit) - if rows: - results.extend(rows) - except Exception: - pass - - # 按 name 去重 - deduped: List[Dict[str, Any]] = [] - seen = set() - for r in results: - k = str(r.get("name", "")) - if k and k not in seen: - deduped.append(r) - seen.add(k) - return deduped - - -# 通过对话/陈述中的entity_ids反查实体名称 -_FETCH_ENTITIES_BY_IDS = """ -MATCH (e:ExtractedEntity) -WHERE e.id IN $ids AND ($end_user_id IS NULL OR e.end_user_id = $end_user_id) -RETURN e.id AS id, e.name AS name, e.end_user_id AS end_user_id, e.entity_type AS entity_type -""" - -async def _fetch_entities_by_ids(connector: Neo4jConnector, ids: List[str], end_user_id: str | None) -> List[Dict[str, Any]]: - if not ids: - return [] - try: - rows = await connector.execute_query(_FETCH_ENTITIES_BY_IDS, ids=list({i for i in ids if i}), end_user_id=end_user_id) - return rows or [] - except Exception: - return [] - - -# 增强的时间实体检索 -_TIME_ENTITY_SEARCH = """ -MATCH (e:ExtractedEntity) -WHERE e.entity_type CONTAINS "TIME" OR e.entity_type CONTAINS "DATE" OR e.name =~ $date_pattern -AND ($end_user_id IS NULL OR e.end_user_id = $end_user_id) -RETURN e.id AS id, e.name AS name, e.end_user_id AS end_user_id, e.entity_type AS entity_type -LIMIT $limit -""" - -async def _search_time_entities(connector: Neo4jConnector, end_user_id: str | None, limit: int = 5) -> List[Dict[str, Any]]: - """专门搜索时间相关的实体""" - try: - date_pattern = r".*\d{4}.*|.*\d{1,2}月\d{1,2}日.*" - rows = await connector.execute_query(_TIME_ENTITY_SEARCH, - date_pattern=date_pattern, - end_user_id=end_user_id, - limit=limit) - return rows or [] - except Exception: - return [] - - -# 技术术语专门检索 -async def _search_tech_terms(connector: Neo4jConnector, question: str, end_user_id: str | None, limit: int = 3) -> List[Dict[str, Any]]: - """专门搜索技术术语相关的实体""" - tech_entities = [] - try: - # GPS相关 - if any(term in question for term in ["GPS", "导航", "定位系统"]): - gps_rows = await connector.execute_query(SEARCH_ENTITIES_BY_NAME, q="GPS", end_user_id=end_user_id, limit=limit) - if gps_rows: - tech_entities.extend(gps_rows) - - # 活动相关 - if any(term in question for term in ["工作坊", "研讨会", "网络研讨会"]): - workshop_rows = await connector.execute_query(SEARCH_ENTITIES_BY_NAME, q="工作坊", end_user_id=end_user_id, limit=limit) - if workshop_rows: - tech_entities.extend(workshop_rows) - - # 时间顺序相关 - if any(term in question for term in ["先", "后", "第一个"]): - time_rows = await connector.execute_query(SEARCH_ENTITIES_BY_NAME, q="第一次", end_user_id=end_user_id, limit=limit) - if time_rows: - tech_entities.extend(time_rows) - - except Exception: - pass - - return tech_entities - - -# 中英相对时间解析：today/昨天/上周/3天后 等简单归一化为日期 -def _resolve_relative_times_cn_en(text: str, anchor: datetime) -> str: - t = str(text) if text is not None else "" - # 英文 today/yesterday/tomorrow - t = re.sub(r"\btoday\b", anchor.date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\byesterday\b", (anchor - timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\btomorrow\b", (anchor + timedelta(days=1)).date().isoformat(), t, flags=re.IGNORECASE) - - # 英文 X days ago / in X days - def _ago_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor - timedelta(days=n)).date().isoformat() - def _in_repl(m: re.Match[str]) -> str: - n = int(m.group(1)) - return (anchor + timedelta(days=n)).date().isoformat() - t = re.sub(r"\b(\d+)\s+days\s+ago\b", _ago_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\bin\s+(\d+)\s+days\b", _in_repl, t, flags=re.IGNORECASE) - t = re.sub(r"\blast\s+week\b", (anchor - timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - t = re.sub(r"\bnext\s+week\b", (anchor + timedelta(days=7)).date().isoformat(), t, flags=re.IGNORECASE) - - # 中文 今天/昨天/明天 - t = re.sub(r"今天", anchor.date().isoformat(), t) - t = re.sub(r"昨日|昨天", (anchor - timedelta(days=1)).date().isoformat(), t) - t = re.sub(r"明天", (anchor + timedelta(days=1)).date().isoformat(), t) - # 中文 X天前 / X天后 - t = re.sub(r"(\d+)天前", lambda m: (anchor - timedelta(days=int(m.group(1)))).date().isoformat(), t) - t = re.sub(r"(\d+)天后", lambda m: (anchor + timedelta(days=int(m.group(1)))).date().isoformat(), t) - # 中文 上周 / 下周（近似7天） - t = re.sub(r"上周", (anchor - timedelta(days=7)).date().isoformat(), t) - t = re.sub(r"下周", (anchor + timedelta(days=7)).date().isoformat(), t) - # 中文 月日（无年份）补全年份 - def _md_repl(m: re.Match[str]) -> str: - mon = int(m.group(1)); day = int(m.group(2)) - return f"{anchor.year}-{mon:02d}-{day:02d}" - t = re.sub(r"(\d{1,2})月(\d{1,2})日", _md_repl, t) - return t - - -async def run_longmemeval_test( - sample_size: int = 3, - end_user_id: str = "longmemeval_zh_bak_2", - search_limit: int = 8, - context_char_budget: int = 4000, - llm_temperature: float = 0.0, - llm_max_tokens: int = 16, - search_type: str = "hybrid", - data_path: str | None = None, - start_index: int = 0, -) -> Dict[str, Any]: - """LongMemEval 评估测试：增强技术术语检索能力""" - - # 数据路径 - if not data_path: - # 固定使用中文数据集：dataset/longmemeval_oracle_zh.json - dataset_dir = Path(__file__).resolve().parent.parent / "dataset" - data_path = str(dataset_dir / "longmemeval_oracle_zh.json") - - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 longmemeval_oracle_zh.json 放置在: {dataset_dir}" - ) - - qa_list: List[Dict[str, Any]] = load_dataset_any(data_path) - # 支持评估全部样本：当 sample_size <= 0 时，取从 start_index 到末尾 - if sample_size is None or sample_size <= 0: - items = qa_list[start_index:] - else: - items = qa_list[start_index:start_index + sample_size] - - # 初始化组件 - 使用异步LLM客户端 - llm_client = get_llm_client(os.getenv("EVAL_LLM_ID")) - connector = Neo4jConnector() - cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID")) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - - # 指标收集 - latencies_llm: List[float] = [] - latencies_search: List[float] = [] - per_query_context_counts: List[int] = [] - per_query_context_avg_tokens: List[float] = [] - per_query_context_chars: List[int] = [] - - type_correct: Dict[str, List[float]] = {} - type_f1: Dict[str, List[float]] = {} - type_jacc: Dict[str, List[float]] = {} - - samples: List[Dict[str, Any]] = [] - # 统计重复的上下文预览（跨样本），便于诊断"相同上下文"问题 - preview_counter: Dict[str, int] = {} - - try: - for item in items: - question = item.get("question", "") - reference = item.get("answer", "") - qtype = item.get("question_type") or item.get("type", "unknown") - - print(f"\n=== 处理问题: {question} ===") - - # 检测问题类型 - is_temporal = any(keyword in question.lower() for keyword in - ['days', 'day', 'before', 'after', 'first', '先后', '顺序', '间隔', '多久', '多少天']) - - # 检索 - t0 = time.time() - contexts_all: List[str] = [] - dialogs, statements, entities = [], [], [] - - try: - if search_type == "embedding": - search_results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["dialogues", "statements", "entities"], - ) - dialogs = search_results.get("dialogues", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - # 实体摘要（最多3个） - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - elif search_type == "keyword": - search_results = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit, - ) - dialogs = search_results.get("dialogues", []) - statements = search_results.get("statements", []) - entities = search_results.get("entities", []) - - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - if entities: - entity_names = [str(e.get("name", "")).strip() for e in entities[:5] if e.get("name")] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - else: # hybrid（增强版：特别优化技术术语检索） - emb_dialogs, emb_statements, emb_entities = [], [], [] - kw_dialogs, kw_statements, kw_entities = [], [], [] - - # 1) 嵌入检索 - try: - emb_res = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["dialogues", "statements", "entities"], - ) - if isinstance(emb_res, dict): - emb_dialogs = emb_res.get("dialogues", []) or [] - emb_statements = emb_res.get("statements", []) or [] - emb_entities = emb_res.get("entities", []) or [] - except Exception as e: - print(f"⚠️ 嵌入检索失败，将继续进行关键词检索: {e}") - - # 2) 关键词检索（增强版） - try: - kw_res = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit, - ) - if isinstance(kw_res, dict): - kw_dialogs = kw_res.get("dialogues", []) or [] - kw_statements = kw_res.get("statements", []) or [] - kw_entities = kw_res.get("entities", []) or [] - - # 技术术语专门检索 - tech_entities = await _search_tech_terms(connector, question, end_user_id, search_limit//2) - if tech_entities: - kw_entities.extend(tech_entities) - - # 时间推理问题的特殊处理 - if is_temporal: - # 专门搜索时间实体 - time_entities = await _search_time_entities(connector, end_user_id, search_limit//2) - if time_entities: - kw_entities.extend(time_entities) - # 添加时间相关关键词检索 - time_keywords = ['天', '日', '月', '年', 'before', 'after', 'first'] - for tk in time_keywords: - try: - time_res = await search_graph( - connector=connector, - q=tk, - end_user_id=end_user_id, - limit=2, - ) - if isinstance(time_res, dict): - kw_dialogs.extend(time_res.get("dialogues", []) or []) - kw_statements.extend(time_res.get("statements", []) or []) - except Exception: - pass - - # 中文关键词拆分后做别名匹配 - cn_tokens = generate_query_keywords_cn(question) # 使用增强版关键词提取 - alias_entities = await _search_entities_by_aliases(connector, cn_tokens, end_user_id, search_limit) - if alias_entities: - kw_entities.extend(alias_entities) - - # 从对话/陈述中的 entity_ids 反查实体 - ids = [] - try: - for d in kw_dialogs: - ids.extend(d.get("entity_ids", []) or []) - for s in kw_statements: - ids.extend(s.get("entity_ids", []) or []) - except Exception: - pass - if ids: - id_entities = await _fetch_entities_by_ids(connector, ids, end_user_id) - if id_entities: - kw_entities.extend(id_entities) - - # 多关键词检索（使用增强版关键词） - try: - eng_words = [w for w in set(re.findall(r"\b\w+\b", question.lower())) if len(w) > 2] - kw_list = generate_query_keywords_cn(question)[:4] # 使用更多关键词 - for kw in kw_list: - if not kw: - continue - sub_res = await search_graph( - connector=connector, - q=str(kw), - end_user_id=end_user_id, - limit=max(3, search_limit // 2), - ) - if isinstance(sub_res, dict): - kw_dialogs.extend(sub_res.get("dialogues", []) or []) - kw_statements.extend(sub_res.get("statements", []) or []) - kw_entities.extend(sub_res.get("entities", []) or []) - except Exception: - pass - - # 选项参与关键词检索 - try: - opt_list = extract_candidate_options(question)[:2] - for opt in opt_list: - if not opt: - continue - opt_res = await search_graph( - connector=connector, - q=str(opt), - end_user_id=end_user_id, - limit=max(3, search_limit // 2), - ) - if isinstance(opt_res, dict): - kw_dialogs.extend(opt_res.get("dialogues", []) or []) - kw_statements.extend(opt_res.get("statements", []) or []) - kw_entities.extend(opt_res.get("entities", []) or []) - except Exception: - pass - except Exception as e: - print(f"❌ 关键词检索失败: {e}") - - # 3) 合并、排序并去重 - all_dialogs = emb_dialogs + kw_dialogs - all_statements = emb_statements + kw_statements - all_entities = emb_entities + kw_entities - - def dedup(items: List[Dict[str, Any]], key_field: str = "uuid") -> List[Dict[str, Any]]: - seen = set() - out = [] - for it in items: - key = str(it.get(key_field, "")) + str(it.get("content", "") + str(it.get("statement", ""))) - if key not in seen: - out.append(it) - seen.add(key) - return out - - # 关键技术实体优先排序 - def enhanced_score(item: Dict[str, Any]) -> float: - score_val = item.get("score", 0.0) - base_score = float(score_val) if score_val is not None else 0.0 - content = str(item.get("content", "") + str(item.get("statement", ""))) - - # 关键技术实体奖励 - key_entities = [] - if any(term in question for term in ["GPS", "导航", "系统"]): - key_entities.extend(["GPS", "导航", "系统", "功能"]) - if any(term in question for term in ["工作坊", "研讨会", "活动"]): - key_entities.extend(["工作坊", "研讨会", "参加"]) - - key_bonus = 0 - for key_ent in key_entities: - if key_ent in content: - key_bonus += 1.0 - - # 时间实体奖励 - time_bonus = 0 - if is_temporal: - time_entities = extract_time_entities(content) - time_bonus = len(time_entities) * 0.5 - - return base_score + key_bonus + time_bonus - - dialogs = dedup(sorted(all_dialogs, key=enhanced_score, reverse=True)) - statements = dedup(sorted(all_statements, key=enhanced_score, reverse=True)) - entities = dedup(all_entities, key_field="name") - - # 4) 构建上下文 - for d in dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - # 实体摘要 - try: - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - except Exception: - pass - - # 全局回退 - if not contexts_all and search_type in ("embedding", "hybrid"): - try: - print("🔁 检索为空，回退到关键词检索...") - kw_fallback = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=max(search_limit, 5), - ) - fb_dialogs = kw_fallback.get("dialogues", []) or [] - fb_statements = kw_fallback.get("statements", []) or [] - fb_entities = kw_fallback.get("entities", []) or [] - - for d in fb_dialogs: - content = str(d.get("content", "")).strip() - if content: - contexts_all.append(content) - for s in fb_statements: - stmt_text = str(s.get("statement", "")).strip() - if stmt_text: - contexts_all.append(stmt_text) - if fb_entities: - entity_names = [str(e.get("name", "")).strip() for e in fb_entities[:5] if e.get("name")] - if entity_names: - contexts_all.append(f"EntitySummary: {', '.join(entity_names)}") - - dialogs = fb_dialogs if fb_dialogs else dialogs - statements = fb_statements if fb_statements else statements - entities = fb_entities if fb_entities else entities - print(f"↩️ 回退到关键词检索: {len(fb_dialogs)} 对话, {len(fb_statements)} 条陈述, {len(fb_entities)} 个实体") - except Exception as fe: - print(f"❌ 关键词回退失败: {fe}") - - ent_count = len(entities) if isinstance(entities, list) else 0 - print(f"✅ {search_type}检索成功: {len(dialogs)} 对话, {len(statements)} 条陈述, {ent_count} 个实体") - if is_temporal: - print("⏰ 检测为时间推理问题，已启用时间优化检索") - - except Exception as e: - print(f"❌ {search_type}检索失败: {e}") - contexts_all = [] - - t1 = time.time() - latencies_search.append((t1 - t0) * 1000) - - # 智能上下文选择 - context_text = "" - if contexts_all: - context_text = smart_context_selection(contexts_all, question, max_chars=context_char_budget) - # 相对时间解析 - try: - context_text = _resolve_relative_times_cn_en(context_text, anchor=datetime.now()) - except Exception: - pass - # 诊断信息 - try: - cn_diag = generate_query_keywords_cn(question)[:4] # 显示更多关键词 - opts = extract_candidate_options(question)[:2] - qlw = [w for w in set(re.findall(r'\b\w+\b', question.lower())) if len(w) > 2][:1] - diag_tokens: List[str] = [] - for t in cn_diag + opts + qlw: - if t and t not in diag_tokens: - diag_tokens.append(t) - print(f"🔍 关键词/选项: {', '.join(diag_tokens)}") - preview = context_text[:200].replace('\n', ' ') - print(f"🔎 上下文预览: {preview}...") - key_preview = preview.strip() - if key_preview: - preview_counter[key_preview] = preview_counter.get(key_preview, 0) + 1 - except Exception: - pass - else: - print("❌ 没有检索到有效上下文") - context_text = "No relevant context found." - - # 记录上下文诊断信息 - per_query_context_counts.append(len(contexts_all)) - per_query_context_avg_tokens.append(avg_context_tokens([context_text])) - per_query_context_chars.append(len(context_text)) - - # LLM 推理（增强技术术语提示） - options = extract_candidate_options(question) - if len(options) >= 2: - opt_lines = "\n".join(f"- {o}" for o in options) - # 技术术语问题的特殊提示 - if any(term in question for term in ["GPS", "系统", "功能", "工作坊", "研讨会"]): - system_prompt = ( - "You are a QA assistant specializing in technical and activity-related questions. " - "Pay special attention to technical terms like GPS, systems, functions, workshops, and seminars. " - "Return ONLY one string: exactly one option from the provided candidates. If the context is insufficient, respond with 'Unknown'. " - "Focus on matching technical details and activity sequences accurately." - ) - elif is_temporal: - system_prompt = ( - "You are a QA assistant specializing in temporal reasoning. Analyze the dates and time relationships in the context carefully. " - "Return ONLY one string: exactly one option from the provided candidates. If the context is insufficient, respond with 'Unknown'. " - "Pay special attention to date sequences and time intervals." - ) - else: - system_prompt = ( - "You are a QA assistant. Respond in the same language as the question. Return ONLY one string: exactly one option from the provided candidates. " - "If the context is insufficient, respond with 'Unknown'. If the context expresses a synonym or paraphrase of a candidate, return the closest candidate. " - "Do not include explanations." - ) - - messages = [ - {"role": "system", "content": system_prompt}, - { - "role": "user", - "content": ( - f"Question: {question}\n\nCandidates:\n{opt_lines}\n\nContext:\n{context_text}\n\nReturn EXACTLY one candidate string (or 'Unknown')." - ), - }, - ] - else: - # 技术术语问题的特殊提示 - if any(term in question for term in ["GPS", "系统", "功能", "工作坊", "研讨会"]): - system_prompt = ( - "You are a QA assistant specializing in technical and activity-related questions. " - "Pay special attention to technical terms like GPS, systems, functions, workshops, and seminars. " - "If the context contains the answer, return a concise answer phrase focusing on technical details. " - "If the answer cannot be determined from the context, respond with 'Unknown'. Return ONLY the final answer string, no explanations." - ) - elif is_temporal: - system_prompt = ( - "You are a QA assistant specializing in temporal reasoning. Analyze the dates and time relationships in the context carefully. " - "If the context contains the answer, return a concise answer phrase focusing on temporal information. " - "If the answer cannot be determined from the context, respond with 'Unknown'. Return ONLY the final answer string, no explanations." - ) - else: - system_prompt = ( - "You are a QA assistant. Respond in the same language as the question. If the context contains the answer, return a concise answer phrase. " - "If the answer cannot be determined from the context, respond with 'Unknown'. Return ONLY the final answer string, no explanations." - ) - - messages = [ - {"role": "system", "content": system_prompt}, - { - "role": "user", - "content": f"Question: {question}\n\nContext:\n{context_text}\n\nReturn ONLY the answer (or 'Unknown').", - }, - ] - - t2 = time.time() - # 使用异步调用 - resp = await llm_client.chat(messages=messages) - t3 = time.time() - latencies_llm.append((t3 - t2) * 1000) - - # 兼容不同的响应格式 - pred_raw = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else "Unknown") - - # 选项题输出规范化 - pred = pred_raw - if len(options) >= 2 and not pred_raw.lower().startswith("unknown"): - def _basic_norm(s: str) -> str: - s = s.lower().strip() - return re.sub(r"[^\w\s]", " ", s) - def _jaccard(a: str, b: str) -> float: - ta = set(t for t in _basic_norm(a).split() if t) - tb = set(t for t in _basic_norm(b).split() if t) - if not ta and not tb: - return 1.0 - if not ta or not tb: - return 0.0 - return len(ta & tb) / len(ta | tb) - best = None - best_score = -1.0 - for o in options: - score = _jaccard(pred_raw, o) - if score > best_score: - best = o - best_score = score - if best is not None and best_score > 0.0: - pred = best - - # 指标 - flag = exact_match(pred, reference) - f1_val = common_f1(str(pred), str(reference)) - j_val = jaccard(str(pred), str(reference)) - - type_correct.setdefault(qtype, []).append(flag) - type_f1.setdefault(qtype, []).append(f1_val) - type_jacc.setdefault(qtype, []).append(j_val) - - samples.append({ - "question": question, - "prediction": pred, - "answer": reference, - "question_type": qtype, - "is_temporal": is_temporal, - "question_id": item.get("question_id"), - "options": options, - "context_count": len(contexts_all), - "context_chars": len(context_text), - "retrieved_dialogue_count": len(dialogs), - "retrieved_statement_count": len(statements), - "metrics": { - "exact_match": bool(flag), - "f1": f1_val, - "jaccard": j_val - }, - "timing": { - "search_ms": (t1 - t0) * 1000, - "llm_ms": (t3 - t2) * 1000 - } - }) - - print(f"🤖 LLM 回答: {pred}") - print(f"✅ 正确答案: {reference}") - print(f"📈 当前指标 - Exact Match: {flag}, F1: {f1_val:.3f}, Jaccard: {j_val:.3f}") - - # 聚合结果 - type_acc = {t: (sum(v) / max(len(v), 1)) for t, v in type_correct.items()} - f1_by_type = {t: (sum(v) / max(len(v), 1)) for t, v in type_f1.items()} - jacc_by_type = {t: (sum(v) / max(len(v), 1)) for t, v in type_jacc.items()} - - result = { - "dataset": "longmemeval", - "items": len(items), - "accuracy_by_type": type_acc, - "f1_by_type": f1_by_type, - "jaccard_by_type": jacc_by_type, - "samples": samples, - "latency": { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm), - }, - "context": { - "avg_tokens": statistics.mean(per_query_context_avg_tokens) if per_query_context_avg_tokens else 0.0, - "avg_chars": statistics.mean(per_query_context_chars) if per_query_context_chars else 0.0, - "count_avg": statistics.mean(per_query_context_counts) if per_query_context_counts else 0.0, - }, - "params": { - "end_user_id": end_user_id, - "search_limit": search_limit, - "context_char_budget": context_char_budget, - "search_type": search_type, - "llm_id": os.getenv("EVAL_LLM_ID"), - "embedding_id": os.getenv("EVAL_EMBEDDING_ID"), - "sample_size": sample_size, - "start_index": start_index, - }, - "timestamp": datetime.now().isoformat() - } - - # 计算汇总指标 - try: - total_items = max(len(samples), 1) - correct_count = sum(1 for s in samples if s.get("metrics", {}).get("exact_match")) - score_accuracy = (correct_count / total_items) * 100.0 - - total_latencies_ms = [] - for s in samples: - t = s.get("timing", {}) - total_latencies_ms.append(float(t.get("search_ms", 0.0)) + float(t.get("llm_ms", 0.0))) - total_lat_stats = latency_stats(total_latencies_ms) if total_latencies_ms else {"p50": 0.0, "iqr": 0.0} - latency_median_s = total_lat_stats.get("p50", 0.0) / 1000.0 - latency_iqr_s = total_lat_stats.get("iqr", 0.0) / 1000.0 - - avg_ctx_tokens = statistics.mean(per_query_context_avg_tokens) if per_query_context_avg_tokens else 0.0 - avg_ctx_tokens_k = avg_ctx_tokens / 1000.0 - - result["metric_summary"] = { - "score_accuracy": score_accuracy, - "latency_median_s": latency_median_s, - "latency_iqr_s": latency_iqr_s, - "avg_context_tokens_k": avg_ctx_tokens_k, - } - except Exception: - result["metric_summary"] = { - "score_accuracy": 0.0, - "latency_median_s": 0.0, - "latency_iqr_s": 0.0, - "avg_context_tokens_k": 0.0, - } - - # 诊断信息 - try: - dups = sorted([(k, c) for k, c in preview_counter.items() if c > 1], key=lambda x: -x[1])[:5] - result["diagnostics"] = { - "duplicate_previews_top": [{"count": c, "preview": k[:120]} for k, c in dups], - "unique_preview_count": len(preview_counter), - } - except Exception: - pass - - return result - - finally: - await connector.close() - - -def main(): - load_dotenv() - parser = argparse.ArgumentParser(description="LongMemEval 评估测试脚本（增强技术术语检索版）") - parser.add_argument("--sample-size", type=int, default=3, help="样本数量（<=0 表示全部）") - parser.add_argument("--all", action="store_true", help="评估全部样本（覆盖 --sample-size）") - parser.add_argument("--start-index", type=int, default=0, help="起始样本索引") - parser.add_argument("--group-id", type=str, default="longmemeval_zh_bak_3", help="图数据库 Group ID") - parser.add_argument("--search-limit", type=int, default=8, help="检索条数上限") - parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") - parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") - parser.add_argument("--llm-max-tokens", type=int, default=16, help="LLM 最大输出 token") - parser.add_argument("--search-type", type=str, default="hybrid", choices=["embedding","keyword","hybrid"], help="检索类型") - parser.add_argument("--data-path", type=str, default=None, help="数据集路径") - args = parser.parse_args() - - sample_size = 0 if args.all else args.sample_size - - result = asyncio.run( - run_longmemeval_test( - sample_size=sample_size, - end_user_id=args.end_user_id, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - llm_temperature=args.llm_temperature, - llm_max_tokens=args.llm_max_tokens, - search_type=args.search_type, - data_path=args.data_path, - start_index=args.start_index, - ) - ) - - # 打印结果 - print("\n" + "="*50) - print("📊 LongMemEval 测试结果:") - print(f" 样本数量: {result['items']}") - - if result['accuracy_by_type']: - print("\n📈 按问题类型细分:") - for qtype, acc in result['accuracy_by_type'].items(): - print(f" {qtype}:") - print(f" Score (Accuracy): {acc:.3f}") - - print(f"\n📊 指标总览:") - ms = result.get('metric_summary', {}) - print(f" Score (Accuracy): {ms.get('score_accuracy', 0.0):.1f}%") - print(f" Latency (s): median {ms.get('latency_median_s', 0.0):.3f}s") - print(f" Latency IQR (s): {ms.get('latency_iqr_s', 0.0):.3f}s") - print(f" Avg Context Tokens (k): {ms.get('avg_context_tokens_k', 0.0):.3f}k") - - print(f"\n⏱️ 细分性能指标:") - print(f" 检索延迟(均值): {result['latency']['search']['mean']:.1f}ms") - print(f" LLM延迟(均值): {result['latency']['llm']['mean']:.1f}ms") - print(f" 上下文长度(均值): {result['context']['avg_chars']:.0f} 字符") - - - # 保存结果到文件 - try: - out_dir = os.path.join(PROJECT_ROOT, "evaluation", "longmemeval", "results") - os.makedirs(out_dir, exist_ok=True) - ts = datetime.now().strftime("%Y%m%d_%H%M%S") - out_path = os.path.join(out_dir, f"longmemeval_{result['params']['search_type']}_{ts}.json") - with open(out_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n💾 结果已保存: {out_path}") - except Exception as e: - print(f"⚠️ 结果保存失败: {e}") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py b/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py deleted file mode 100644 index e07b0cab..00000000 --- a/api/app/core/memory/evaluation/memsciqa/memsciqa-test.py +++ /dev/null @@ -1,559 +0,0 @@ -import argparse -import asyncio -import json -import os -import time -from datetime import datetime -from typing import List, Dict, Any -import re -from pathlib import Path - -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - print(f"✅ 加载评估配置: {eval_config_path}") - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.src.search import run_hybrid_search # 使用与 evaluate_qa.py 相同的检索函数 -from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient -from app.core.models.base import RedBearModelConfig -from app.core.memory.utils.config.config_utils import get_embedder_config - -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.common.metrics import exact_match, latency_stats, avg_context_tokens - -from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: - """基于问题关键词对上下文进行评分选择，并在预算内拼接文本。 - - 参考 evaluation/memsciqa/evaluate_qa.py 的实现，避免路径导入带来的不稳定。 - """ - if not contexts: - return "" - question_lower = (question or "").lower() - stop_words = { - 'what','when','where','who','why','how','did','do','does','is','are','was','were', - 'the','a','an','and','or','but' - } - question_words = set(re.findall(r"\b\w+\b", question_lower)) - question_words = {w for w in question_words if w not in stop_words and len(w) > 2} - - scored = [] - for i, ctx in enumerate(contexts): - ctx_lower = (ctx or "").lower() - score = 0 - matches = 0 - for w in question_words: - if w in ctx_lower: - matches += 1 - score += ctx_lower.count(w) * 2 - length = len(ctx) - if 100 < length < 2000: - score += 5 - elif length >= 2000: - score += 2 - if i < 3: - score += 3 - scored.append((score, ctx, matches)) - - scored.sort(key=lambda x: x[0], reverse=True) - - selected: List[str] = [] - total = 0 - for score, ctx, _ in scored: - if total + len(ctx) <= max_chars: - selected.append(ctx) - total += len(ctx) - else: - if score > 10 and total < max_chars - 200: - remaining = max_chars - total - lines = ctx.split('\n') - rel_lines: List[str] = [] - cur = 0 - for line in lines: - l = line.lower() - if any(w in l for w in question_words) and cur < remaining - 50: - rel_lines.append(line) - cur += len(line) - if rel_lines: - truncated = '\n'.join(rel_lines) - if len(truncated) > 50: - selected.append(truncated + "\n[相关内容截断...]") - total += len(truncated) - break - return "\n\n".join(selected) - - -def extract_question_keywords(question: str, max_keywords: int = 8) -> List[str]: - """提取问题中的关键词（简单英文分词，去停用词，长度>=3）。""" - ql = (question or "").lower() - stop_words = { - 'what','when','where','who','why','how','did','do','does','is','are','was','were', - 'the','a','an','and','or','but','of','to','in','on','for','with','from','that','this' - } - words = re.findall(r"\b[\w-]+\b", ql) - kws = [w for w in words if w not in stop_words and len(w) >= 3] - # 去重保序 - seen = set() - uniq = [] - for w in kws: - if w not in seen: - uniq.append(w) - seen.add(w) - if len(uniq) >= max_keywords: - break - return uniq - - -def analyze_contexts_simple(contexts: List[str], keywords: List[str], top_n: int = 5) -> List[Dict[str, int | float]]: - """对上下文进行简单相关性打分，仅用于控制台可视化。 - - 评分: score = match_count*200 + min(len(text), 100000)/100 - """ - results = [] - for ctx in contexts: - tl = (ctx or "").lower() - match_count = sum(1 for k in keywords if k in tl) - length = len(ctx) - score = match_count * 200 + min(length, 100000) / 100.0 - results.append({"score": float(f"{score:.0f}"), "match": match_count, "length": length}) - results.sort(key=lambda x: (x["score"], x["match"], x["length"]), reverse=True) - return results[:max(top_n, 0)] - - -# 纯测试脚本不进行摄入；若需摄入请使用 evaluate_qa.py - - -def load_dataset_memsciqa(data_path: str) -> List[Dict[str, Any]]: - if not os.path.exists(data_path): - raise FileNotFoundError(f"未找到数据集: {data_path}") - items: List[Dict[str, Any]] = [] - with open(data_path, "r", encoding="utf-8") as f: - for line in f: - line = line.strip() - if not line: - continue - try: - items.append(json.loads(line)) - except Exception: - # 跳过坏行但不中断 - continue - return items - - -async def run_memsciqa_test( - sample_size: int = 3, - end_user_id: str | None = None, - search_limit: int = 8, - context_char_budget: int = 4000, - llm_temperature: float = 0.0, - llm_max_tokens: int = 64, - search_type: str = "embedding", - data_path: str | None = None, - start_index: int = 0, - verbose: bool = True, -) -> Dict[str, Any]: - """memsciqa 增强测试脚本：结合 evaluate_qa 的三路检索与智能上下文选择。 - - - 支持从指定索引开始与评估全部样本（sample_size<=0） - - 支持在摄入前重置组（清空图）与跳过摄入 - - 支持 keyword / embedding / hybrid 三种检索 - """ - - # 默认使用指定的 memsci 组 ID - end_user_id = end_user_id or "group_memsci" - - # 数据路径解析 - if not data_path: - dataset_dir = Path(__file__).resolve().parent.parent / "dataset" - data_path = str(dataset_dir / "msc_self_instruct.jsonl") - - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 msc_self_instruct.jsonl 放置在: {dataset_dir}" - ) - - # 加载数据 - all_items = load_dataset_memsciqa(data_path) - if sample_size is None or sample_size <= 0: - items = all_items[start_index:] - else: - items = all_items[start_index:start_index + sample_size] - - # 初始化 LLM（纯测试：不进行摄入） - llm = get_llm_client(os.getenv("EVAL_LLM_ID")) - - # 初始化 Neo4j 连接与向量检索 Embedder（对齐 locomo_test） - connector = Neo4jConnector() - embedder = None - if search_type in ("embedding", "hybrid"): - cfg_dict = get_embedder_config(os.getenv("EVAL_EMBEDDING_ID")) - embedder = OpenAIEmbedderClient( - model_config=RedBearModelConfig.model_validate(cfg_dict) - ) - - # 评估循环 - latencies_llm: List[float] = [] - latencies_search: List[float] = [] - # 存储完整上下文文本用于统计 - contexts_used: List[str] = [] - per_query_context_chars: List[int] = [] - per_query_context_counts: List[int] = [] - correct_flags: List[float] = [] - f1s: List[float] = [] - b1s: List[float] = [] - jss: List[float] = [] - samples: List[Dict[str, Any]] = [] - - total_items = len(items) - for idx, item in enumerate(items): - if verbose: - print(f"\n🧪 评估样本: {idx+1}/{total_items}") - question = item.get("self_instruct", {}).get("B", "") or item.get("question", "") - reference = item.get("self_instruct", {}).get("A", "") or item.get("answer", "") - - # 检索：使用与 evaluate_qa.py 相同的 run_hybrid_search - t0 = time.time() - results = None - try: - if search_type in ("embedding", "hybrid"): - # 使用嵌入检索（与 qwen_search_eval 对齐） - results = await search_graph_by_embedding( - connector=connector, - embedder_client=embedder, - query_text=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], # 使用 chunks 而不是 dialogues - ) - elif search_type == "keyword": - # 关键词检索（直接调用 graph_search） - results = await search_graph( - connector=connector, - q=question, - end_user_id=end_user_id, - limit=search_limit, - include=["chunks", "statements", "entities", "summaries"], # 使用 chunks 而不是 dialogues - ) - except Exception: - results = None - t1 = time.time() - search_ms = (t1 - t0) * 1000 - latencies_search.append(search_ms) - - # 构建上下文：与 evaluate_qa.py 完全一致的逻辑 - contexts_all: List[str] = [] - retrieved_counts: Dict[str, int] = {} - if results: - # 处理 hybrid 搜索结果 - if search_type == "hybrid": - emb = results.get("embedding_search", {}) if isinstance(results.get("embedding_search"), dict) else {} - kw = results.get("keyword_search", {}) if isinstance(results.get("keyword_search"), dict) else {} - emb_dialogs = emb.get("dialogues", []) - emb_statements = emb.get("statements", []) - emb_entities = emb.get("entities", []) - kw_dialogs = kw.get("dialogues", []) - kw_statements = kw.get("statements", []) - kw_entities = kw.get("entities", []) - all_dialogs = emb_dialogs + kw_dialogs - all_statements = emb_statements + kw_statements - all_entities = emb_entities + kw_entities - - # 简单去重 - seen_dialog = set() - dialogues = [] - for d in all_dialogs: - key = (str(d.get("uuid", "")), str(d.get("content", ""))) - if key not in seen_dialog: - dialogues.append(d) - seen_dialog.add(key) - - seen_stmt = set() - statements = [] - for s in all_statements: - key = str(s.get("statement", "")) - if key not in seen_stmt: - statements.append(s) - seen_stmt.add(key) - - seen_ent = set() - entities = [] - for e in all_entities: - key = str(e.get("name", "")) - if key not in seen_ent: - entities.append(e) - seen_ent.add(key) - else: - # embedding 或 keyword 单独搜索 - dialogues = results.get("dialogues", []) - statements = results.get("statements", []) - entities = results.get("entities", []) - - retrieved_counts = { - "dialogues": len(dialogues), - "statements": len(statements), - "entities": len(entities), - } - - # 构建上下文文本 - for d in dialogues: - text = str(d.get("content", "")).strip() - if text: - contexts_all.append(text) - - for s in statements: - text = str(s.get("statement", "")).strip() - if text: - contexts_all.append(text) - - # 实体摘要 - if entities: - scored = [e for e in entities if e.get("score") is not None] - top_entities = sorted(scored, key=lambda x: x.get("score", 0), reverse=True)[:3] if scored else entities[:3] - if top_entities: - summary_lines = [] - for e in top_entities: - name = str(e.get("name", "")).strip() - etype = str(e.get("entity_type", "")).strip() - score = e.get("score") - if name: - meta = [] - if etype: - meta.append(f"type={etype}") - if isinstance(score, (int, float)): - meta.append(f"score={score:.3f}") - summary_lines.append(f"EntitySummary: {name}{(' [' + '; '.join(meta) + ']') if meta else ''}") - if summary_lines: - contexts_all.append("\n".join(summary_lines)) - - if verbose: - if retrieved_counts: - print(f"✅ 检索成功: {retrieved_counts.get('dialogues',0)} dialogues, {retrieved_counts.get('statements',0)} 条陈述, {retrieved_counts.get('entities',0)} 个实体, {retrieved_counts.get('summaries',0)} 个摘要") - print(f"📊 有效上下文数量: {len(contexts_all)}") - q_keywords = extract_question_keywords(question, max_keywords=8) - if q_keywords: - print(f"🔍 问题关键词: {set(q_keywords)}") - if contexts_all: - analysis = analyze_contexts_simple(contexts_all, q_keywords, top_n=5) - if analysis: - print("📊 上下文相关性分析:") - for a in analysis: - print(f" - 得分: {int(a['score'])}, 关键词匹配: {a['match']}, 长度: {a['length']}") - # 打印检索到的上下文预览，便于定位为何为 Unknown - print("🔎 上下文预览（最多前10条，每条截断展示）:") - for i, ctx in enumerate(contexts_all[:10]): - preview = str(ctx).replace("\n", " ") - if len(preview) > 300: - preview = preview[:300] + "..." - print(f" [{i+1}] 长度: {len(ctx)} | 片段: {preview}") - # 标注参考答案是否出现在任一上下文中 - ref_lower = (str(reference) or "").lower() - if ref_lower: - hits = [] - for i, ctx in enumerate(contexts_all): - if ref_lower in str(ctx).lower(): - hits.append(i+1) - print(f"🔗 参考答案命中上下文条数: {len(hits)}" + (f" | 命中索引: {hits}" if hits else "")) - - context_text = smart_context_selection(contexts_all, question, max_chars=context_char_budget) if contexts_all else "" - if not context_text: - context_text = "No relevant context found." - contexts_used.append(context_text) - per_query_context_chars.append(len(context_text)) - per_query_context_counts.append(len(contexts_all)) - - if verbose: - selected_count = (context_text.count("\n\n") + 1) if context_text else 0 - print(f"✅ 智能选择: {selected_count}个上下文, 总长度: {len(context_text)}字符") - # 展示拼接后的上下文片段，便于核查是否包含答案 - concat_preview = context_text.replace("\n", " ") - if len(concat_preview) > 600: - concat_preview = concat_preview[:600] + "..." - print(f"🧵 拼接上下文预览: {concat_preview}") - - messages = [ - { - "role": "system", - "content": ( - "You are a QA assistant. Answer in English. Follow these guidelines:\n" - "1) If the context contains information to answer the question, provide a concise answer based on the context;\n" - "2) If the context does not contain enough information to answer the question, respond with 'Unknown';\n" - "3) Keep your answer brief and to the point;\n" - "4) Do not add explanations or additional text beyond the answer." - ), - }, - {"role": "user", "content": f"Question: {question}\n\nContext:\n{context_text}"}, - ] - - t2 = time.time() - try: - # 使用异步调用 - resp = await llm.chat(messages=messages) - # 更健壮的响应解析，处理不同的LLM响应格式 - if hasattr(resp, 'content'): - pred = resp.content.strip() - elif isinstance(resp, dict) and "choices" in resp and len(resp["choices"]) > 0: - pred = resp["choices"][0]["message"]["content"].strip() - elif isinstance(resp, dict) and "content" in resp: - pred = resp["content"].strip() - elif isinstance(resp, str): - pred = resp.strip() - else: - pred = "Unknown" - print(f"⚠️ LLM响应格式异常: {type(resp)} - {resp}") - - # 检查预测是否为"Unknown"或空，如果是则检查上下文是否真的没有答案 - if pred.lower() in ["unknown", ""]: - # 如果参考答案在上下文中存在，但LLM返回Unknown，可能是提示词问题 - ref_lower = (str(reference) or "").lower() - if ref_lower and any(ref_lower in ctx.lower() for ctx in contexts_all): - print("⚠️ 参考答案在上下文中存在但LLM返回Unknown，检查提示词") - except Exception as e: - # 更详细的错误处理 - pred = "Unknown" - print(f"⚠️ LLM调用异常: {e}") - t3 = time.time() - llm_ms = (t3 - t2) * 1000 - latencies_llm.append(llm_ms) - - exact = exact_match(pred, reference) - correct_flags.append(exact) - f1_val = f1_score(str(pred), str(reference)) - b1_val = bleu1(str(pred), str(reference)) - j_val = jaccard(str(pred), str(reference)) - f1s.append(f1_val) - b1s.append(b1_val) - jss.append(j_val) - - if verbose: - print(f"🤖 LLM 回答: {pred}") - print(f"✅ 正确答案: {reference}") - print(f"📈 当前指标 - F1: {f1_val:.3f}, BLEU-1: {b1_val:.3f}, Jaccard: {j_val:.3f}") - print(f"⏱️ 延迟 - 检索: {search_ms:.0f}ms, LLM: {llm_ms:.0f}ms") - - # 对齐 locomo/qwen_search_eval.py 的样本输出结构 - samples.append({ - "question": str(question), - "answer": str(reference), - "prediction": str(pred), - "metrics": { - "f1": f1_val, - "b1": b1_val, - "j": j_val - }, - "retrieval": { - "retrieved_documents": len(contexts_all), - "context_length": len(context_text), - "search_limit": search_limit, - "max_chars": context_char_budget - }, - "timing": { - "search_ms": search_ms, - "llm_ms": llm_ms - } - }) - - # 计算总体指标与聚合 - acc = sum(correct_flags) / max(len(correct_flags), 1) - ctx_avg_tokens = avg_context_tokens(contexts_used) - result = { - "dataset": "memsciqa", - "items": len(items), - "metrics": { - "f1": (sum(f1s) / max(len(f1s), 1)) if f1s else 0.0, - "b1": (sum(b1s) / max(len(b1s), 1)) if b1s else 0.0, - "j": (sum(jss) / max(len(jss), 1)) if jss else 0.0, - }, - "context": { - "avg_tokens": ctx_avg_tokens, - "avg_chars": (sum(per_query_context_chars) / max(len(per_query_context_chars), 1)) if per_query_context_chars else 0.0, - "count_avg": (sum(per_query_context_counts) / max(len(per_query_context_counts), 1)) if per_query_context_counts else 0.0, - "avg_memory_tokens": 0.0 - }, - "latency": { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm), - }, - "samples": samples, - "params": { - "end_user_id": end_user_id, - "search_limit": search_limit, - "context_char_budget": context_char_budget, - "llm_temperature": llm_temperature, - "llm_max_tokens": llm_max_tokens, - "search_type": search_type, - "start_index": start_index, - "llm_id": os.getenv("EVAL_LLM_ID"), - "retrieval_embedding_id": os.getenv("EVAL_EMBEDDING_ID") - }, - "timestamp": datetime.now().isoformat(), - } - try: - await connector.close() - except Exception: - pass - return result - - -def main(): - load_dotenv() - parser = argparse.ArgumentParser(description="memsciqa 测试脚本（三路检索 + 智能上下文选择）") - parser.add_argument("--sample-size", type=int, default=10, help="样本数量（<=0 表示全部）") - parser.add_argument("--all", action="store_true", help="评估全部样本（覆盖 --sample-size）") - parser.add_argument("--start-index", type=int, default=0, help="起始样本索引") - parser.add_argument("--group-id", type=str, default="group_memsci", help="图数据库 Group ID（默认 group_memsci）") - parser.add_argument("--search-limit", type=int, default=8, help="检索条数上限") - parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") - parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") - parser.add_argument("--llm-max-tokens", type=int, default=64, help="LLM 最大输出 token") - parser.add_argument("--search-type", type=str, default="embedding", choices=["embedding","keyword","hybrid"], help="检索类型（hybrid 等同于 embedding）") - parser.add_argument("--data-path", type=str, default=None, help="数据集路径（默认 data/msc_self_instruct.jsonl）") - parser.add_argument("--output", type=str, default=None, help="将评估结果保存到指定文件路径（JSON）") - parser.add_argument("--verbose", action="store_true", default=True, help="打印过程日志（默认开启）") - parser.add_argument("--quiet", action="store_true", help="关闭过程日志") - args = parser.parse_args() - - sample_size = 0 if args.all else args.sample_size - - verbose_flag = False if args.quiet else args.verbose - result = asyncio.run( - run_memsciqa_test( - sample_size=sample_size, - end_user_id=args.end_user_id, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - llm_temperature=args.llm_temperature, - llm_max_tokens=args.llm_max_tokens, - search_type=args.search_type, - data_path=args.data_path, - start_index=args.start_index, - verbose=verbose_flag, - ) - ) - - print(json.dumps(result, ensure_ascii=False, indent=2)) - - # 结果保存 - out_path = args.output - if not out_path: - eval_dir = os.path.dirname(os.path.abspath(__file__)) - dataset_results_dir = os.path.join(eval_dir, "results") - ts = datetime.now().strftime("%Y%m%d_%H%M%S") - out_path = os.path.join(dataset_results_dir, f"memsciqa_{result['params']['search_type']}_{ts}.json") - try: - os.makedirs(os.path.dirname(out_path), exist_ok=True) - with open(out_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n💾 结果已保存: {out_path}") - except Exception as e: - print(f"⚠️ 结果保存失败: {e}") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py b/api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py deleted file mode 100644 index 40684f4c..00000000 --- a/api/app/core/memory/evaluation/memsciqa/memsciqa_benchmark.py +++ /dev/null @@ -1,369 +0,0 @@ -import argparse -import asyncio -import json -import os -import time -from datetime import datetime -from typing import List, Dict, Any -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent.parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector -from app.core.memory.src.search import run_hybrid_search # 使用旧版本（重构前） -from app.core.memory.utils.llm.llm_utils import get_llm_client -from app.core.memory.evaluation.extraction_utils import ingest_contexts_via_full_pipeline -from app.core.memory.evaluation.common.metrics import exact_match, latency_stats, avg_context_tokens - - -def smart_context_selection(contexts: List[str], question: str, max_chars: int = 4000) -> str: - """基于问题关键词对上下文进行评分选择，并在预算内拼接文本。""" - if not contexts: - return "" - import re - # 提取问题关键词（移除停用词） - question_lower = (question or "").lower() - stop_words = { - 'what','when','where','who','why','how','did','do','does','is','are','was','were', - 'the','a','an','and','or','but' - } - question_words = set(re.findall(r"\b\w+\b", question_lower)) - question_words = {w for w in question_words if w not in stop_words and len(w) > 2} - - # 评分 - scored = [] - for i, ctx in enumerate(contexts): - ctx_lower = (ctx or "").lower() - score = 0 - matches = 0 - for w in question_words: - if w in ctx_lower: - matches += 1 - score += ctx_lower.count(w) * 2 - length = len(ctx) - if 100 < length < 2000: - score += 5 - elif length >= 2000: - score += 2 - if i < 3: - score += 3 - scored.append((score, ctx, matches)) - - scored.sort(key=lambda x: x[0], reverse=True) - - # 选择直到达到字符限制，必要时截断包含关键词的段落 - selected: List[str] = [] - total = 0 - for score, ctx, _ in scored: - if total + len(ctx) <= max_chars: - selected.append(ctx) - total += len(ctx) - else: - if score > 10 and total < max_chars - 200: - remaining = max_chars - total - lines = ctx.split('\n') - rel_lines: List[str] = [] - cur = 0 - for line in lines: - l = line.lower() - if any(w in l for w in question_words) and cur < remaining - 50: - rel_lines.append(line) - cur += len(line) - if rel_lines: - truncated = '\n'.join(rel_lines) - if len(truncated) > 50: - selected.append(truncated + "\n[相关内容截断...]") - total += len(truncated) - break - return "\n\n".join(selected) - - -def build_context_from_dialog(dialog_obj: Dict[str, Any]) -> str: - """Compose a text context from `dialog` list in msc_self_instruct item.""" - parts: List[str] = [] - for turn in dialog_obj.get("dialog", []): - speaker = turn.get("speaker", "") - text = turn.get("text", "") - if text: - parts.append(f"{speaker}: {text}") - return "\n".join(parts) - - -def _combine_dialogues_for_hybrid(results: Dict[str, Any]) -> List[Dict[str, Any]]: - """Combine dialogues from embedding and keyword searches (embedding first).""" - if results is None: - return [] - emb = [] - kw = [] - if isinstance(results.get("embedding_search"), dict): - emb = results.get("embedding_search", {}).get("dialogues", []) or [] - elif isinstance(results.get("dialogues"), list): - emb = results.get("dialogues", []) or [] - if isinstance(results.get("keyword_search"), dict): - kw = results.get("keyword_search", {}).get("dialogues", []) or [] - seen = set() - merged: List[Dict[str, Any]] = [] - for d in emb: - k = (str(d.get("uuid", "")), str(d.get("content", ""))) - if k not in seen: - merged.append(d) - seen.add(k) - for d in kw: - k = (str(d.get("uuid", "")), str(d.get("content", ""))) - if k not in seen: - merged.append(d) - seen.add(k) - return merged - - - -async def run_memsciqa_eval(sample_size: int = 1, end_user_id: str | None = None, search_limit: int = 8, context_char_budget: int = 4000, llm_temperature: float = 0.0, llm_max_tokens: int = 64, search_type: str = "hybrid", memory_config: "MemoryConfig" = None) -> Dict[str, Any]: - end_user_id = end_user_id or SELECTED_GROUP_ID - - # Load data - dataset_dir = Path(__file__).resolve().parent.parent / "dataset" - data_path = dataset_dir / "msc_self_instruct.jsonl" - - if not os.path.exists(data_path): - raise FileNotFoundError( - f"数据集文件不存在: {data_path}\n" - f"请将 msc_self_instruct.jsonl 放置在: {dataset_dir}" - ) - with open(data_path, "r", encoding="utf-8") as f: - lines = f.readlines() - items: List[Dict[str, Any]] = [json.loads(l) for l in lines[:sample_size]] - - - # 改为：每条样本仅摄入一个上下文（完整对话转录），避免多上下文摄入 - # 说明：memsciqa 数据集的每个样本天然只有一个对话，保持按样本一上下文的策略 - contexts: List[str] = [build_context_from_dialog(item) for item in items] - await ingest_contexts_via_full_pipeline(contexts, end_user_id) - - # LLM client (使用异步调用) - from app.db import get_db - - db = next(get_db()) - try: - llm_client = get_llm_client(os.getenv("EVAL_LLM_ID"), db) - finally: - db.close() - - # Evaluate each item - connector = Neo4jConnector() - latencies_llm: List[float] = [] - latencies_search: List[float] = [] - contexts_used: List[str] = [] - correct_flags: List[float] = [] - f1s: List[float] = [] - b1s: List[float] = [] - jss: List[float] = [] - try: - for item in items: - question = item.get("self_instruct", {}).get("B", "") or item.get("question", "") - reference = item.get("self_instruct", {}).get("A", "") or item.get("answer", "") - # 检索：对齐 locomo 的三路检索（dialogues/statements/entities） - t0 = time.time() - try: - results = await run_hybrid_search( - query_text=question, - search_type=search_type, - end_user_id=end_user_id, - limit=search_limit, - include=["dialogues", "statements", "entities"], - output_path=None, - ) - except Exception: - results = None - t1 = time.time() - latencies_search.append((t1 - t0) * 1000) - - # 构建上下文：包含对话、陈述和实体摘要，并智能选择 - contexts_all: List[str] = [] - if results: - if search_type == "hybrid": - emb = results.get("embedding_search", {}) if isinstance(results.get("embedding_search"), dict) else {} - kw = results.get("keyword_search", {}) if isinstance(results.get("keyword_search"), dict) else {} - emb_dialogs = emb.get("dialogues", []) - emb_statements = emb.get("statements", []) - emb_entities = emb.get("entities", []) - kw_dialogs = kw.get("dialogues", []) - kw_statements = kw.get("statements", []) - kw_entities = kw.get("entities", []) - all_dialogs = emb_dialogs + kw_dialogs - all_statements = emb_statements + kw_statements - all_entities = emb_entities + kw_entities - - # 简单去重与限制 - seen_texts = set() - for d in all_dialogs: - text = str(d.get("content", "")).strip() - if text and text not in seen_texts: - contexts_all.append(text) - seen_texts.add(text) - if len(contexts_all) >= search_limit: - break - for s in all_statements: - text = str(s.get("statement", "")).strip() - if text and text not in seen_texts: - contexts_all.append(text) - seen_texts.add(text) - if len(contexts_all) >= search_limit: - break - # 实体摘要（最多3个） - names = [] - merged_entities = all_entities[:] - for e in merged_entities: - name = str(e.get("name", "")).strip() - if name and name not in names: - names.append(name) - if len(names) >= 3: - break - if names: - contexts_all.append("EntitySummary: " + ", ".join(names)) - else: - dialogs = results.get("dialogues", []) - statements = results.get("statements", []) - entities = results.get("entities", []) - for d in dialogs: - text = str(d.get("content", "")).strip() - if text: - contexts_all.append(text) - for s in statements: - text = str(s.get("statement", "")).strip() - if text: - contexts_all.append(text) - names = [str(e.get("name", "")).strip() for e in entities[:3] if e.get("name")] - if names: - contexts_all.append("EntitySummary: " + ", ".join(names)) - - # 智能选择并截断到预算 - context_text = smart_context_selection(contexts_all, question, max_chars=context_char_budget) if contexts_all else "" - if not context_text: - context_text = "No relevant context found." - contexts_used.append(context_text[:200]) - - # Call LLM (使用异步调用) - messages = [ - {"role": "system", "content": "You are a QA assistant. Answer in English. Strictly follow: 1) If the context contains the answer, copy the shortest exact span from the context as the answer; 2) If the answer cannot be determined from the context, respond with 'Unknown'; 3) Return ONLY the answer text, no explanations."}, - {"role": "user", "content": f"Question: {question}\n\nContext:\n{context_text}"}, - ] - t2 = time.time() - resp = await llm_client.chat(messages=messages) - t3 = time.time() - latencies_llm.append((t3 - t2) * 1000) - pred = resp.content.strip() if hasattr(resp, 'content') else (resp["choices"][0]["message"]["content"].strip() if isinstance(resp, dict) else str(resp).strip()) - # Metrics: F1, BLEU-1, Jaccard; keep exact match for reference - correct_flags.append(exact_match(pred, reference)) - from app.core.memory.evaluation.common.metrics import f1_score, bleu1, jaccard - f1s.append(f1_score(str(pred), str(reference))) - b1s.append(bleu1(str(pred), str(reference))) - jss.append(jaccard(str(pred), str(reference))) - - # Aggregate metrics - acc = sum(correct_flags) / max(len(correct_flags), 1) - ctx_avg_tokens = avg_context_tokens(contexts_used) - result = { - "dataset": "memsciqa", - "items": len(items), - "metrics": { - "accuracy": acc, - # Placeholders for extensibility - "f1": (sum(f1s) / max(len(f1s), 1)) if f1s else 0.0, - "bleu1": (sum(b1s) / max(len(b1s), 1)) if b1s else 0.0, - "jaccard": (sum(jss) / max(len(jss), 1)) if jss else 0.0, - }, - "latency": { - "search": latency_stats(latencies_search), - "llm": latency_stats(latencies_llm), - }, - "avg_context_tokens": ctx_avg_tokens, - } - return result - finally: - await connector.close() - - -def main(): - # Load environment variables first - load_dotenv() - - # Get defaults from environment variables - env_sample_size = os.getenv("MEMSCIQA_SAMPLE_SIZE") - env_search_limit = os.getenv("MEMSCIQA_SEARCH_LIMIT") - env_context_budget = os.getenv("MEMSCIQA_CONTEXT_CHAR_BUDGET") - env_llm_max_tokens = os.getenv("MEMSCIQA_LLM_MAX_TOKENS") - env_skip_ingest = os.getenv("MEMSCIQA_SKIP_INGEST", "false").lower() in ("true", "1", "yes") - env_output_dir = os.getenv("MEMSCIQA_OUTPUT_DIR") - - # Convert to appropriate types with fallback to code defaults - default_sample_size = int(env_sample_size) if env_sample_size else 1 - default_search_limit = int(env_search_limit) if env_search_limit else 8 - default_context_budget = int(env_context_budget) if env_context_budget else 4000 - default_llm_max_tokens = int(env_llm_max_tokens) if env_llm_max_tokens else 64 - default_output_dir = env_output_dir if env_output_dir else None - - parser = argparse.ArgumentParser(description="Evaluate DMR (memsciqa) with graph search and Qwen") - - parser.add_argument("--sample-size", type=int, default=1, help="评测样本数量") - parser.add_argument("--end-user-id", type=str, default=None, help="可选 end_user_id，默认使用环境变量") - parser.add_argument("--search-limit", type=int, default=8, help="每类检索最大返回数") - parser.add_argument("--context-char-budget", type=int, default=4000, help="上下文字符预算") - - parser.add_argument("--llm-temperature", type=float, default=0.0, help="LLM 温度") - parser.add_argument("--llm-max-tokens", type=int, default=default_llm_max_tokens, - help=f"LLM 最大生成长度 (env: MEMSCIQA_LLM_MAX_TOKENS={env_llm_max_tokens or 'not set'})") - parser.add_argument("--search-type", type=str, choices=["keyword","embedding","hybrid"], default="hybrid", help="检索类型") - parser.add_argument("--skip-ingest", action="store_true", default=env_skip_ingest, - help=f"跳过数据摄入，使用 Neo4j 中的现有数据 (env: MEMSCIQA_SKIP_INGEST={os.getenv('MEMSCIQA_SKIP_INGEST', 'false')})") - parser.add_argument("--output-dir", type=str, default=default_output_dir, - help=f"结果保存目录 (env: MEMSCIQA_OUTPUT_DIR={env_output_dir or 'not set'})") - args = parser.parse_args() - - result = asyncio.run( - run_memsciqa_eval( - sample_size=args.sample_size, - end_user_id=args.end_user_id, - search_limit=args.search_limit, - context_char_budget=args.context_char_budget, - llm_temperature=args.llm_temperature, - llm_max_tokens=args.llm_max_tokens, - search_type=args.search_type, - skip_ingest=args.skip_ingest, - ) - ) - - # Print results to console - print(json.dumps(result, ensure_ascii=False, indent=2)) - - # Save results to file - output_dir = args.output_dir - if output_dir is None: - # Use absolute path to ensure results are saved in the correct location - script_dir = Path(__file__).resolve().parent - output_dir = script_dir / "results" - elif not Path(output_dir).is_absolute(): - # If relative path, make it relative to this script's directory - script_dir = Path(__file__).resolve().parent - output_dir = script_dir / output_dir - else: - output_dir = Path(output_dir) - - output_dir.mkdir(parents=True, exist_ok=True) - - timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S") - output_path = output_dir / f"memsciqa_{timestamp_str}.json" - - try: - with open(output_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n✅ 结果已保存到: {output_path}") - except Exception as e: - print(f"\n❌ 保存结果失败: {e}") - - -if __name__ == "__main__": - main() diff --git a/api/app/core/memory/evaluation/run_eval.py b/api/app/core/memory/evaluation/run_eval.py deleted file mode 100644 index 56b2e790..00000000 --- a/api/app/core/memory/evaluation/run_eval.py +++ /dev/null @@ -1,147 +0,0 @@ -import argparse -import asyncio -import json -import os -from typing import Any, Dict -from pathlib import Path -from dotenv import load_dotenv - -# Load evaluation config -eval_config_path = Path(__file__).resolve().parent / ".env.evaluation" -if eval_config_path.exists(): - load_dotenv(eval_config_path, override=True) - -from app.repositories.neo4j.neo4j_connector import Neo4jConnector - -from app.core.memory.evaluation.memsciqa.evaluate_qa import run_memsciqa_eval -from app.core.memory.evaluation.longmemeval.qwen_search_eval import run_longmemeval_test -from app.core.memory.evaluation.locomo.qwen_search_eval import run_locomo_eval - - -async def run( - dataset: str, - sample_size: int, - reset_group: bool, - end_user_id: str | None, - judge_model: str | None = None, - search_limit: int | None = None, - context_char_budget: int | None = None, - llm_temperature: float | None = None, - llm_max_tokens: int | None = None, - search_type: str | None = None, - start_index: int | None = None, - max_contexts_per_item: int | None = None, -) -> Dict[str, Any]: - # Use environment variable with fallback chain if not provided - if end_user_id is None: - end_user_id = os.getenv("EVAL_END_USER_ID", "benchmark_default") - - if reset_group: - connector = Neo4jConnector() - try: - await connector.delete_group(end_user_id) - finally: - await connector.close() - - if dataset == "locomo": - kwargs: Dict[str, Any] = {"sample_size": sample_size, "end_user_id": end_user_id} - if search_limit is not None: - kwargs["search_limit"] = search_limit - if context_char_budget is not None: - kwargs["context_char_budget"] = context_char_budget - if llm_temperature is not None: - kwargs["llm_temperature"] = llm_temperature - if llm_max_tokens is not None: - kwargs["llm_max_tokens"] = llm_max_tokens - if search_type is not None: - kwargs["search_type"] = search_type - return await run_locomo_eval(**kwargs) - - if dataset == "memsciqa": - kwargs: Dict[str, Any] = {"sample_size": sample_size, "end_user_id": end_user_id} - if search_limit is not None: - kwargs["search_limit"] = search_limit - if context_char_budget is not None: - kwargs["context_char_budget"] = context_char_budget - if llm_temperature is not None: - kwargs["llm_temperature"] = llm_temperature - if llm_max_tokens is not None: - kwargs["llm_max_tokens"] = llm_max_tokens - if search_type is not None: - kwargs["search_type"] = search_type - return await run_memsciqa_eval(**kwargs) - - if dataset == "longmemeval": - kwargs: Dict[str, Any] = {"sample_size": sample_size, "end_user_id": end_user_id} - if search_limit is not None: - kwargs["search_limit"] = search_limit - if context_char_budget is not None: - kwargs["context_char_budget"] = context_char_budget - if llm_temperature is not None: - kwargs["llm_temperature"] = llm_temperature - if llm_max_tokens is not None: - kwargs["llm_max_tokens"] = llm_max_tokens - if search_type is not None: - kwargs["search_type"] = search_type - if start_index is not None: - kwargs["start_index"] = start_index - if max_contexts_per_item is not None: - kwargs["max_contexts_per_item"] = max_contexts_per_item - return await run_longmemeval_test(**kwargs) - raise ValueError(f"未知数据集: {dataset}") - - -def main(): - load_dotenv() - parser = argparse.ArgumentParser(description="统一评估入口：memsciqa / longmemeval / locomo") - parser.add_argument("--dataset", choices=["memsciqa", "longmemeval", "locomo"], required=True) - parser.add_argument("--sample-size", type=int, default=1, help="先用一条数据跑通") - parser.add_argument("--reset-group", action="store_true", help="运行前清空当前 end_user_id 的图数据") - parser.add_argument("--group-id", type=str, default=None, help="可选 end_user_id，默认取 runtime.json") - parser.add_argument("--judge-model", type=str, default=None, help="可选：longmemeval 判别式评测模型名") - parser.add_argument("--search-limit", type=int, default=None, help="检索返回的对话节点数量上限（不提供则使用各脚本默认）") - parser.add_argument("--context-char-budget", type=int, default=None, help="上下文字符预算（不提供则使用各脚本默认）") - parser.add_argument("--llm-temperature", type=float, default=None, help="生成温度（不提供则使用各脚本默认）") - parser.add_argument("--llm-max-tokens", type=int, default=None, help="最大生成 tokens（不提供则使用各脚本默认）") - parser.add_argument("--search-type", type=str, default=None, choices=["keyword", "embedding", "hybrid"], help="检索类型（不提供则使用各脚本默认）") - # 仅透传到 longmemeval；其他数据集忽略 - parser.add_argument("--start-index", type=int, default=None, help="仅 longmemeval：起始样本索引（不提供则用脚本默认）") - parser.add_argument("--max-contexts-per-item", type=int, default=None, help="仅 longmemeval：每条样本摄入的上下文数量上限（不提供则用脚本默认）") - parser.add_argument("--output", type=str, default=None, help="可选：将评估结果保存到指定文件路径（JSON）；不提供时默认保存到 evaluation//results 目录") - args = parser.parse_args() - - result = asyncio.run(run( - args.dataset, - args.sample_size, - args.reset_group, - args.end_user_id, - args.judge_model, - args.search_limit, - args.context_char_budget, - args.llm_temperature, - args.llm_max_tokens, - args.search_type, - args.start_index, - args.max_contexts_per_item, - )) - print(json.dumps(result, ensure_ascii=False, indent=2)) - - # 结果输出逻辑保持不变 - if args.output: - out_path = args.output - else: - eval_dir = os.path.dirname(os.path.abspath(__file__)) - dataset_results_dir = os.path.join(eval_dir, args.dataset, "results") - out_filename = f"{args.dataset}_{args.sample_size}.json" - out_path = os.path.join(dataset_results_dir, out_filename) - - out_dir = os.path.dirname(out_path) - if out_dir and not os.path.exists(out_dir): - os.makedirs(out_dir, exist_ok=True) - with open(out_path, "w", encoding="utf-8") as f: - json.dump(result, f, ensure_ascii=False, indent=2) - print(f"\n结果已保存到: {out_path}") - - -if __name__ == "__main__": - main() diff --git a/redbear-mem-benchmark b/redbear-mem-benchmark index d9a00be6..558c023d 160000 --- a/redbear-mem-benchmark +++ b/redbear-mem-benchmark @@ -1 +1 @@ -Subproject commit d9a00be62d974c0ad071c27e86f878b921c675b6 +Subproject commit 558c023dadb5327a05561b22d8fb363c6ee2be29