From 8466c8e0192c84641ec2bd8f088075604b814bb2 Mon Sep 17 00:00:00 2001 From: lanceyq <1982376970@qq.com> Date: Tue, 3 Mar 2026 23:30:54 +0800 Subject: [PATCH] [fix] Revising the judgment method for the interest analysis tags --- .../controllers/memory_agent_controller.py | 32 ++--- .../core/memory/analytics/hot_memory_tags.py | 112 ++++++++++++++++++ .../core/memory/utils/prompt/prompt_utils.py | 17 +++ .../prompt/prompts/interest_filter.jinja2 | 47 ++++++++ api/app/services/memory_agent_service.py | 32 ++--- 5 files changed, 210 insertions(+), 30 deletions(-) create mode 100644 api/app/core/memory/utils/prompt/prompts/interest_filter.jinja2 diff --git a/api/app/controllers/memory_agent_controller.py b/api/app/controllers/memory_agent_controller.py index b88e65ff..8f2e5c31 100644 --- a/api/app/controllers/memory_agent_controller.py +++ b/api/app/controllers/memory_agent_controller.py @@ -661,34 +661,38 @@ async def get_knowledge_type_stats_api( return fail(BizCode.INTERNAL_ERROR, "获取知识库类型统计失败", str(e)) -@router.get("/analytics/hot_memory_tags/by_user", response_model=ApiResponse) -async def get_hot_memory_tags_by_user_api( - end_user_id: Optional[str] = Query(None, description="用户ID（可选）"), - limit: int = Query(20, description="返回标签数量限制"), +@router.get("/analytics/interest_distribution/by_user", response_model=ApiResponse) +async def get_interest_distribution_by_user_api( + end_user_id: Optional[str] = Query(None, description="用户ID（必填）"), + limit: int = Query(5, le=5, description="返回兴趣标签数量限制，最多5个"), + language_type: str = Header(default=None, alias="X-Language-Type"), current_user: User = Depends(get_current_user), - db: Session=Depends(get_db), + db: Session = Depends(get_db), ): """ - 获取指定用户的热门记忆标签 + 获取指定用户的兴趣分布标签 - 注意：标签语言由写入时的 X-Language-Type 决定，查询时不进行翻译 + 与热门标签不同，此接口专注于识别用户的兴趣活动（运动、爱好、学习、创作等）， + 过滤掉纯物品、工具、地点等不代表用户主动参与活动的名词。 返回格式： [ - {"name": "标签名", "frequency": 频次}, + {"name": "兴趣活动名", "frequency": 频次}, ... ] """ - api_logger.info(f"Hot memory tags by user requested: end_user_id={end_user_id}") + language = get_language_from_header(language_type) + api_logger.info(f"Interest distribution by user requested: end_user_id={end_user_id}, language={language}") try: - result = await memory_agent_service.get_hot_memory_tags_by_user( + result = await memory_agent_service.get_interest_distribution_by_user( end_user_id=end_user_id, - limit=limit + limit=limit, + language=language ) - return success(data=result, msg="获取热门记忆标签成功") + return success(data=result, msg="获取兴趣分布标签成功") except Exception as e: - api_logger.error(f"Hot memory tags by user failed: {str(e)}") - return fail(BizCode.INTERNAL_ERROR, "获取热门记忆标签失败", str(e)) + api_logger.error(f"Interest distribution by user failed: {str(e)}") + return fail(BizCode.INTERNAL_ERROR, "获取兴趣分布标签失败", str(e)) @router.get("/analytics/user_profile", response_model=ApiResponse) diff --git a/api/app/core/memory/analytics/hot_memory_tags.py b/api/app/core/memory/analytics/hot_memory_tags.py index abb0f138..da08e88e 100644 --- a/api/app/core/memory/analytics/hot_memory_tags.py +++ b/api/app/core/memory/analytics/hot_memory_tags.py @@ -16,6 +16,10 @@ class FilteredTags(BaseModel): """用于接收LLM筛选后的核心标签列表的模型。""" meaningful_tags: List[str] = Field(..., description="从原始列表中筛选出的具有核心代表意义的名词列表。") +class InterestTags(BaseModel): + """用于接收LLM筛选后的兴趣活动标签列表的模型。""" + interest_tags: List[str] = Field(..., description="从原始列表中筛选出的代表用户兴趣活动的标签列表。") + async def filter_tags_with_llm(tags: List[str], end_user_id: str) -> List[str]: """ 使用LLM筛选标签列表，仅保留具有代表性的核心名词。 @@ -89,6 +93,70 @@ async def filter_tags_with_llm(tags: List[str], end_user_id: str) -> List[str]: # 在LLM失败时返回原始标签，确保流程继续 return tags +async def filter_interests_with_llm(tags: List[str], end_user_id: str, language: str = "zh") -> List[str]: + """ + 使用LLM从标签列表中筛选出代表用户兴趣活动的标签。 + + 与 filter_tags_with_llm 不同，此函数专注于识别"活动/行为"类兴趣， + 过滤掉纯物品、工具、地点等不代表用户主动参与活动的名词。 + + Args: + tags: 原始标签列表 + end_user_id: 用户ID，用于获取LLM配置 + + Returns: + 筛选后的兴趣活动标签列表 + """ + try: + with get_db_context() as db: + from app.services.memory_agent_service import ( + get_end_user_connected_config, + ) + connected_config = get_end_user_connected_config(end_user_id, db) + config_id = connected_config.get("memory_config_id") + workspace_id = connected_config.get("workspace_id") + + if not config_id and not workspace_id: + raise ValueError( + f"No memory_config_id found for end_user_id: {end_user_id}." + ) + + config_service = MemoryConfigService(db) + memory_config = config_service.load_memory_config( + config_id=config_id, + workspace_id=workspace_id + ) + + if not memory_config.llm_model_id: + raise ValueError( + f"No llm_model_id found in memory config {config_id}." + ) + + factory = MemoryClientFactory(db) + llm_client = factory.get_llm_client(memory_config.llm_model_id) + + tag_list_str = ", ".join(tags) + from app.core.memory.utils.prompt.prompt_utils import render_interest_filter_prompt + rendered_prompt = render_interest_filter_prompt(tag_list_str, language=language) + messages = [ + { + "role": "user", + "content": rendered_prompt + } + ] + + structured_response = await llm_client.response_structured( + messages=messages, + response_model=InterestTags + ) + + return structured_response.interest_tags + + except Exception as e: + print(f"兴趣标签LLM筛选过程中发生错误: {e}") + return tags + + async def get_raw_tags_from_db( connector: Neo4jConnector, end_user_id: str, @@ -183,3 +251,47 @@ async def get_hot_memory_tags(end_user_id: str, limit: int = 10, by_user: bool = finally: # 确保关闭连接 await connector.close() + +async def get_interest_distribution(end_user_id: str, limit: int = 10, by_user: bool = False, language: str = "zh") -> List[Tuple[str, int]]: + """ + 获取用户的兴趣分布标签。 + + 与 get_hot_memory_tags 不同，此函数使用专门针对"活动/行为"的LLM prompt， + 过滤掉纯物品、工具、地点等，只保留能代表用户兴趣爱好的活动类标签。 + + Args: + end_user_id: 必需参数。如果by_user=False，则为end_user_id；如果by_user=True，则为user_id + limit: 最终返回的标签数量限制（默认10） + by_user: 是否按user_id查询（默认False，按end_user_id查询） + + Raises: + ValueError: 如果end_user_id未提供或为空 + """ + if not end_user_id or not end_user_id.strip(): + raise ValueError( + "end_user_id is required. Please provide a valid end_user_id or user_id." + ) + + connector = Neo4jConnector() + try: + # 查询更多原始标签，给LLM提供充足上下文 + query_limit = 40 + raw_tags_with_freq = await get_raw_tags_from_db(connector, end_user_id, query_limit, by_user=by_user) + if not raw_tags_with_freq: + return [] + + raw_tag_names = [tag for tag, freq in raw_tags_with_freq] + + # 使用兴趣活动专用prompt进行筛选 + interest_tag_names = await filter_interests_with_llm(raw_tag_names, end_user_id, language=language) + + # 保留原始频率，按兴趣筛选结果过滤 + final_tags = [ + (tag, freq) + for tag, freq in raw_tags_with_freq + if tag in interest_tag_names + ] + + return final_tags[:limit] + finally: + await connector.close() diff --git a/api/app/core/memory/utils/prompt/prompt_utils.py b/api/app/core/memory/utils/prompt/prompt_utils.py index d88f50cf..0cea98f2 100644 --- a/api/app/core/memory/utils/prompt/prompt_utils.py +++ b/api/app/core/memory/utils/prompt/prompt_utils.py @@ -548,3 +548,20 @@ async def render_ontology_extraction_prompt( }) return rendered_prompt + + +def render_interest_filter_prompt(tag_list: str, language: str = "zh") -> str: + """ + Renders the interest filter prompt using the interest_filter.jinja2 template. + + Args: + tag_list: Comma-separated string of raw tags to filter + language: Output language ("zh" for Chinese, "en" for English) + + Returns: + Rendered prompt content as string + """ + template = prompt_env.get_template("interest_filter.jinja2") + rendered_prompt = template.render(tag_list=tag_list, language=language) + log_prompt_rendering('interest filter', rendered_prompt) + return rendered_prompt diff --git a/api/app/core/memory/utils/prompt/prompts/interest_filter.jinja2 b/api/app/core/memory/utils/prompt/prompts/interest_filter.jinja2 new file mode 100644 index 00000000..1e3aac55 --- /dev/null +++ b/api/app/core/memory/utils/prompt/prompts/interest_filter.jinja2 @@ -0,0 +1,47 @@ +{% if language == "zh" %} +You are a user interest analysis expert. Your task is to identify activity-based tags from a tag list that represent the user's hobbies and interests. Please output the results in Chinese. + +**Keep Rules** (keep if any condition is met): +- Tags representing sports or physical activities the user actively participates in (e.g., '攀岩', '篮球', '游泳', '跑步') +- Tags representing cultural or entertainment hobbies (e.g., '读书', '看电影', '听音乐', '摄影') +- Tags representing learning or creative activities (e.g., '编程', '绘画', '写作', '烹饪') +- Tags representing specific interest domains or hobby categories (e.g., '历史', '天文', '园艺') + +**Filter Rules** (remove if any condition is met): +- Pure object or tool names that do not represent an activity (e.g., '篮球鞋', '相机', '书桌') +- Pure location or venue names (e.g., '篮球场', '图书馆', '健身房') +- Abstract concepts or quality descriptions (e.g., '核心力量', '团队合作', '专注力') +- Person names, brand names, or proper nouns (e.g., '乔丹', 'Nike') + +**Merge Rules**: For semantically similar tags, keep only the most representative one. +For example: keep '篮球' over '打篮球'; keep '读书' over '阅读'. + +**Example**: +Input: ['攀岩', '篮球场', '篮球鞋', '篮球', '《三体》', '历史', '核心力量', '烹饪', '菜刀'] +Output: ['攀岩', '篮球', '历史', '烹饪'] + +Please filter the following tag list and return only the tags that represent user interest activities in Chinese: {{ tag_list }} +{% else %} +You are a user interest analysis expert. Your task is to identify activity-based tags from a tag list that represent the user's hobbies and interests. Please output the results in English. + +**Keep Rules** (keep if any condition is met): +- Tags representing sports or physical activities the user actively participates in (e.g., 'rock climbing', 'basketball', 'swimming', 'running') +- Tags representing cultural or entertainment hobbies (e.g., 'reading', 'watching movies', 'listening to music', 'photography') +- Tags representing learning or creative activities (e.g., 'programming', 'painting', 'writing', 'cooking') +- Tags representing specific interest domains or hobby categories (e.g., 'history', 'astronomy', 'gardening') + +**Filter Rules** (remove if any condition is met): +- Pure object or tool names that do not represent an activity (e.g., 'basketball shoes', 'camera', 'desk') +- Pure location or venue names (e.g., 'basketball court', 'library', 'gym') +- Abstract concepts or quality descriptions (e.g., 'core strength', 'teamwork', 'focus') +- Person names, brand names, or proper nouns (e.g., 'Jordan', 'Nike') + +**Merge Rules**: For semantically similar tags, keep only the most representative one. +For example: keep 'basketball' over 'playing basketball'; keep 'reading' over 'reading books'. + +**Example**: +Input: ['rock climbing', 'basketball court', 'basketball shoes', 'basketball', 'The Three-Body Problem', 'history', 'core strength', 'cooking', 'kitchen knife'] +Output: ['rock climbing', 'basketball', 'history', 'cooking'] + +Please filter the following tag list and return only the tags that represent user interest activities in English: {{ tag_list }} +{% endif %} diff --git a/api/app/services/memory_agent_service.py b/api/app/services/memory_agent_service.py index 1f3667a6..16aee283 100644 --- a/api/app/services/memory_agent_service.py +++ b/api/app/services/memory_agent_service.py @@ -36,7 +36,7 @@ from app.core.memory.agent.utils.messages_tools import ( ) from app.core.memory.agent.utils.type_classifier import status_typle from app.core.memory.agent.utils.write_tools import write # 新增：直接导入 write 函数 -from app.core.memory.analytics.hot_memory_tags import get_hot_memory_tags +from app.core.memory.analytics.hot_memory_tags import get_hot_memory_tags, get_interest_distribution from app.core.memory.utils.llm.llm_utils import MemoryClientFactory from app.db import get_db_context from app.models.knowledge_model import Knowledge, KnowledgeType @@ -890,36 +890,36 @@ class MemoryAgentService: return result - async def get_hot_memory_tags_by_user( + + async def get_interest_distribution_by_user( self, end_user_id: Optional[str] = None, - limit: int = 20 + limit: int = 5, + language: str = "zh" ) -> List[Dict[str, Any]]: """ - 获取指定用户的热门记忆标签 + 获取指定用户的兴趣分布标签。 + + 与热门标签不同，此接口专注于识别用户的兴趣活动（运动、爱好、学习等）， + 过滤掉纯物品、工具、地点等不代表用户主动参与活动的名词。 参数： - - end_user_id: 用户ID（可选），对应Neo4j中的end_user_id字段 + - end_user_id: 用户ID（必填） - limit: 返回标签数量限制 + - language: 输出语言（"zh" 中文, "en" 英文） 返回格式： [ - {"name": "标签名", "frequency": 频次}, + {"name": "兴趣活动名", "frequency": 频次}, ... ] - - 注意：标签语言由写入时的 X-Language-Type 决定，查询时不进行翻译 """ try: - # by_user=False 表示按 end_user_id 查询（在Neo4j中，end_user_id就是用户维度） - tags = await get_hot_memory_tags(end_user_id, limit=limit, by_user=False) - payload = [] - for tag, freq in tags: - payload.append({"name": tag, "frequency": freq}) - return payload + tags = await get_interest_distribution(end_user_id, limit=limit, by_user=False, language=language) + return [{"name": tag, "frequency": freq} for tag, freq in tags] except Exception as e: - logger.error(f"热门记忆标签查询失败: {e}") - raise Exception(f"热门记忆标签查询失败: {e}") + logger.error(f"兴趣分布标签查询失败: {e}") + raise Exception(f"兴趣分布标签查询失败: {e}") async def get_user_profile(