From a01525e239b00f3fbeed87c769eb8ada5b08863b Mon Sep 17 00:00:00 2001
From: Eternity <1533512157@qq.com>
Date: Thu, 16 Apr 2026 13:27:36 +0800
Subject: [PATCH] refactor(memory): consolidate memory search services and
 update model client handling

- Consolidate memory search services by removing separate content_search.py and perceptual_search.py
- Update model client handling in base_pipeline.py to use ModelApiKeyService for LLM client initialization
- Add new prompt files and modify existing services to support consolidated search architecture
- Refactor memory read pipeline and related services to use updated model client approach
---
 api/app/core/memory/memory_service.py         |  11 +-
 api/app/core/memory/models/service_models.py  |  17 +-
 .../core/memory/pipelines/base_pipeline.py    |  31 ++-
 api/app/core/memory/pipelines/memory_read.py  |  34 ++-
 api/app/core/memory/prompt/__init__.py        |  85 +++++++
 .../core/memory/prompt/problem_split.jinja2   | 212 ++++++++++++++++
 .../memory/read_services/content_search.py    |  46 ++--
 .../read_services/memory_search/__init__.py   |   0
 .../memory_search/content_search.py           |  14 --
 .../memory_search/perceptual_search.py        | 228 ------------------
 .../read_services/query_preprocessor.py       |  29 ++-
 .../memory/read_services/result_builder.py    |   2 +-
 .../memory/read_services/retrieval_summary.py |  11 +
 api/app/core/memory/utils/llm/llm_utils.py    |  48 +++-
 api/app/repositories/neo4j/graph_search.py    |  10 +-
 .../prompt/prompt_optimizer_system.jinja2     |   2 +-
 16 files changed, 471 insertions(+), 309 deletions(-)
 create mode 100644 api/app/core/memory/prompt/__init__.py
 create mode 100644 api/app/core/memory/prompt/problem_split.jinja2
 delete mode 100644 api/app/core/memory/read_services/memory_search/__init__.py
 delete mode 100644 api/app/core/memory/read_services/memory_search/content_search.py
 delete mode 100644 api/app/core/memory/read_services/memory_search/perceptual_search.py
 create mode 100644 api/app/core/memory/read_services/retrieval_summary.py

diff --git a/api/app/core/memory/memory_service.py b/api/app/core/memory/memory_service.py
index 67c814b1..15ea14b4 100644
--- a/api/app/core/memory/memory_service.py
+++ b/api/app/core/memory/memory_service.py
@@ -1,7 +1,7 @@
 from sqlalchemy.orm import Session
 
 from app.core.memory.enums import StorageType, SearchStrategy
-from app.core.memory.models.service_models import Memory, MemoryContext
+from app.core.memory.models.service_models import MemoryContext, MemorySearchResult
 from app.core.memory.pipelines.memory_read import ReadPipeLine
 from app.db import get_db_context
 from app.services.memory_config_service import MemoryConfigService
@@ -35,9 +35,14 @@ class MemoryService:
     async def write(self, messages: list[dict]) -> str:
         raise NotImplementedError
 
-    async def read(self, query: str, history: list, search_switch: SearchStrategy) -> list[Memory]:
+    async def read(
+            self,
+            query: str,
+            search_switch: SearchStrategy,
+            limit: int = 10,
+    ) -> MemorySearchResult:
         with get_db_context() as db:
-            return await ReadPipeLine(self.ctx, db).run(query, search_switch, limit=10)
+            return await ReadPipeLine(self.ctx, db).run(query, search_switch, limit)
 
     async def forget(self, max_batch: int = 100, min_days: int = 30) -> dict:
         raise NotImplementedError
diff --git a/api/app/core/memory/models/service_models.py b/api/app/core/memory/models/service_models.py
index 82a867c7..477c0ba8 100644
--- a/api/app/core/memory/models/service_models.py
+++ b/api/app/core/memory/models/service_models.py
@@ -1,6 +1,7 @@
-from pydantic import BaseModel, Field, field_serializer, ConfigDict
+from pydantic import BaseModel, Field, field_serializer, ConfigDict, model_validator, computed_field
 
 from app.core.memory.enums import Neo4jNodeType, StorageType
+from app.core.validators import file_validator
 from app.schemas.memory_config_schema import MemoryConfig
 
 
@@ -24,3 +25,17 @@ class Memory(BaseModel):
     @field_serializer("source")
     def serialize_source(self, v) -> str:
         return v.value
+
+
+class MemorySearchResult(BaseModel):
+    memories: list[Memory]
+
+    @computed_field
+    @property
+    def content(self) -> str:
+        return "\n".join([memory.content for memory in self.memories])
+
+    @computed_field
+    @property
+    def count(self) -> int:
+        return len(self.memories)
diff --git a/api/app/core/memory/pipelines/base_pipeline.py b/api/app/core/memory/pipelines/base_pipeline.py
index 322f6787..60c48b9d 100644
--- a/api/app/core/memory/pipelines/base_pipeline.py
+++ b/api/app/core/memory/pipelines/base_pipeline.py
@@ -4,22 +4,32 @@ from typing import Any
 
 from sqlalchemy.orm import Session
 
-from app.core.memory.llm_tools import OpenAIEmbedderClient
 from app.core.memory.models.service_models import MemoryContext
-from app.core.models import RedBearModelConfig
+from app.core.models import RedBearModelConfig, RedBearLLM, RedBearEmbeddings
 from app.services.memory_config_service import MemoryConfigService
+from app.services.model_service import ModelApiKeyService
 
 
 class ModelClientMixin(ABC):
     @staticmethod
-    def get_llm_client(db: Session, model_id: uuid.UUID):
-        pass
+    def get_llm_client(db: Session, model_id: uuid.UUID) -> RedBearLLM:
+        api_config = ModelApiKeyService.get_available_api_key(db, model_id)
+        return RedBearLLM(
+            RedBearModelConfig(
+                model_name=api_config.model_name,
+                provider=api_config.provider,
+                api_key=api_config.api_key,
+                base_url=api_config.api_base,
+                is_omni=api_config.is_omni,
+                support_thinking="thinking" in (api_config.capability or []),
+            )
+        )
 
     @staticmethod
-    def get_embedding_client(db: Session, model_id: uuid.UUID) -> OpenAIEmbedderClient:
+    def get_embedding_client(db: Session, model_id: uuid.UUID) -> RedBearEmbeddings:
         config_service = MemoryConfigService(db)
         embedder_client_config = config_service.get_embedder_config(str(model_id))
-        return OpenAIEmbedderClient(
+        return RedBearEmbeddings(
             RedBearModelConfig(
                 model_name=embedder_client_config["model_name"],
                 provider=embedder_client_config["provider"],
@@ -30,10 +40,15 @@ class ModelClientMixin(ABC):
 
 
 class BasePipeline(ABC):
-    def __init__(self, ctx: MemoryContext, db: Session):
+    def __init__(self, ctx: MemoryContext):
         self.ctx = ctx
-        self.db = db
 
     @abstractmethod
     async def run(self, *args, **kwargs) -> Any:
         pass
+
+
+class DBRequiredPipeline(BasePipeline, ABC):
+    def __init__(self, ctx: MemoryContext, db: Session):
+        super().__init__(ctx)
+        self.db = db
diff --git a/api/app/core/memory/pipelines/memory_read.py b/api/app/core/memory/pipelines/memory_read.py
index 5f5a1a1f..83662d90 100644
--- a/api/app/core/memory/pipelines/memory_read.py
+++ b/api/app/core/memory/pipelines/memory_read.py
@@ -1,31 +1,41 @@
-from app.core.memory.enums import SearchStrategy
-from app.core.memory.pipelines.base_pipeline import BasePipeline, ModelClientMixin
-from app.core.memory.read_services.content_search import Neo4jSearchService
+from app.core.memory.enums import SearchStrategy, StorageType
+from app.core.memory.models.service_models import MemorySearchResult
+from app.core.memory.pipelines.base_pipeline import ModelClientMixin, DBRequiredPipeline
+from app.core.memory.read_services.content_search import Neo4jSearchService, RAGSearchService
 from app.core.memory.read_services.query_preprocessor import QueryPreprocessor
 
 
-class ReadPipeLine(ModelClientMixin, BasePipeline):
-    async def run(self, query: str, search_switch: SearchStrategy, limit: int = 10):
+class ReadPipeLine(ModelClientMixin, DBRequiredPipeline):
+    async def run(self, query: str, search_switch: SearchStrategy, limit: int = 10, includes=None) -> MemorySearchResult:
         query = QueryPreprocessor.process(query)
+        if self.ctx.storage_type == StorageType.RAG:
+            return await self._rag_read(query, limit)
         match search_switch:
             case SearchStrategy.DEEP:
-                return await self._deep_read()
+                return await self._deep_read(query, limit, includes)
             case SearchStrategy.NORMAL:
-                return await self._normal_read(query)
+                return await self._normal_read(query, limit, includes)
             case SearchStrategy.QUICK:
-                return await self._quick_read(query, limit)
+                return await self._quick_read(query, limit, includes)
             case _:
                 raise RuntimeError("Unsupported search strategy")
 
-    async def _deep_read(self):
+    async def _rag_read(self, query: str, limit: int) -> MemorySearchResult:
+        service = RAGSearchService(
+            self.ctx
+        )
+        return await service.search()
+
+    async def _deep_read(self, query: str, limit: int, includes=None) -> MemorySearchResult:
         pass
 
-    async def _normal_read(self, query):
+    async def _normal_read(self, query: str, limit: int, includes=None) -> MemorySearchResult:
         pass
 
-    async def _quick_read(self, query, limit):
+    async def _quick_read(self, query: str, limit: int, includes=None) -> MemorySearchResult:
         search_service = Neo4jSearchService(
             self.ctx,
-            self.get_embedding_client(self.db, self.ctx.memory_config.embedding_model_id)
+            self.get_embedding_client(self.db, self.ctx.memory_config.embedding_model_id),
+            includes=includes,
         )
         return await search_service.search(query, limit)
diff --git a/api/app/core/memory/prompt/__init__.py b/api/app/core/memory/prompt/__init__.py
new file mode 100644
index 00000000..299470f8
--- /dev/null
+++ b/api/app/core/memory/prompt/__init__.py
@@ -0,0 +1,85 @@
+import logging
+import threading
+from pathlib import Path
+
+from jinja2 import Environment, FileSystemLoader, TemplateNotFound, TemplateSyntaxError
+
+logger = logging.getLogger(__name__)
+
+PROMPT_DIR = Path(__file__).parent
+
+
+class PromptRenderError(Exception):
+    def __init__(self, template_name: str, error: Exception):
+        self.template_name = template_name
+        self.error = error
+        super().__init__(f"Failed to render prompt '{template_name}': {error}")
+
+
+class PromptManager:
+    _instance = None
+    _lock = threading.Lock()
+
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            with cls._lock:
+                if cls._instance is None:
+                    cls._instance = super().__new__(cls)
+                    cls._instance._init_once()
+        return cls._instance
+
+    def _init_once(self):
+        self.env = Environment(
+            loader=FileSystemLoader(str(PROMPT_DIR)),
+            autoescape=False,
+            keep_trailing_newline=True,
+        )
+        logger.info(f"PromptManager initialized: template_dir={PROMPT_DIR}")
+
+    def __repr__(self):
+        templates = self.list_templates()
+        return f"<PromptManager: {len(templates)} prompts: {templates}>"
+
+    def list_templates(self) -> list[str]:
+        return [
+            Path(name).stem
+            for name in self.env.loader.list_templates()
+            if name.endswith('.jinja2')
+        ]
+
+    def get(self, name: str) -> str:
+        template_name = self._resolve_name(name)
+        try:
+            source, _, _ = self.env.loader.get_source(self.env, template_name)
+            return source
+        except TemplateNotFound:
+            raise FileNotFoundError(
+                f"Prompt '{name}' not found. "
+                f"Available: {self.list_templates()}"
+            )
+
+    def render(self, name: str, **kwargs) -> str:
+        template_name = self._resolve_name(name)
+        try:
+            template = self.env.get_template(template_name)
+            return template.render(**kwargs)
+        except TemplateNotFound:
+            raise FileNotFoundError(
+                f"Prompt '{name}' not found. "
+                f"Available: {self.list_templates()}"
+            )
+        except TemplateSyntaxError as e:
+            logger.error(f"Prompt syntax error in '{name}': {e}", exc_info=True)
+            raise PromptRenderError(name, e)
+        except Exception as e:
+            logger.error(f"Prompt render failed for '{name}': {e}", exc_info=True)
+            raise PromptRenderError(name, e)
+
+    @staticmethod
+    def _resolve_name(name: str) -> str:
+        if not name.endswith('.jinja2'):
+            return f"{name}.jinja2"
+        return name
+
+
+prompt_manager = PromptManager()
diff --git a/api/app/core/memory/prompt/problem_split.jinja2 b/api/app/core/memory/prompt/problem_split.jinja2
new file mode 100644
index 00000000..ff134ddb
--- /dev/null
+++ b/api/app/core/memory/prompt/problem_split.jinja2
@@ -0,0 +1,212 @@
+
+# 角色：{#InputSlot placeholder="角色名称" mode="input"#}{#/InputSlot#}
+你是一个智能数据拆分助手，请根据数据特性判断输入属于哪种类型：
+## 目标：
+你需要根据以下类型对输入数据进行分类，并生成相应的拆分策略和示例。
+---
+
+### 历史信息参考
+在生成扩展问题时，你可以参考以下历史数据（如果提供）：
+- 历史对话或任务的主题；
+- 历史中出现的关键实体（时间、人物、地点、研究主题等）；
+- 历史中已解答的问题（避免重复）；
+- 历史推理链（保持逻辑一致性）。
+
+> 如果没有提供历史信息，则仅根据当前输入问题进行分析。
+输入历史信息内容:{{history}}
+
+## User Input
+{{ sentence }}
+
+## 需求：
+1:首先判断类型（单跳、多跳、开放域、时间）。
+2:根据类型进行拆分。
+3:拆分后的内容需保证信息完整且可独立处理。
+4:对每个拆分条目，可附加示例或说明。
+5:拆分问题的时候可以考虑输入的历史内容，以保持逻辑连贯。
+    比如：输入历史信息内容:[{'Query': '4月27日，我和你推荐过一本书，书名是什么？', 'ANswer': '张曼玉推荐了《小王子》'}]
+    拆分问题：4月27日，我和你推荐过一本书，书名是什么？，可以拆分为：4月27日，张曼玉推荐过一本书，书名是什么？
+
+## 指代消歧规则（Coreference Resolution）：
+在拆分问题时，必须解析并替换所有指代词和抽象称呼，使问题具体化：
+
+1. **"用户"的消歧**：
+   - "用户是谁？" → 分析历史记录，找出对话发起者的姓名
+   - 如果历史中有"我叫X"、"我的名字是X"、或多次提到某个人物（如"老李"、"李建国"），则"用户"指的就是这个人
+   - 示例：历史中反复出现"老李/李建国/建国哥"，则"用户是谁？"应拆分为"李建国是谁？"或"老李（李建国）是谁？"
+
+2. **"我"的消歧**：
+   - "我喜欢什么？" → 从历史中找出对话发起者的姓名，替换为"X喜欢什么？"
+   - 示例：历史中有"张曼玉推荐了《小王子》"，则"我推荐的书是什么？"应拆分为"张曼玉推荐的书是什么？"
+
+3. **"他/她/它"的消歧**：
+   - 从上下文或历史中找出最近提到的同类实体
+   - 示例：历史中有"老李的同事叫他建国哥"，则"他的同事怎么称呼他？"应拆分为"老李的同事怎么称呼他？"
+
+4. **"那个人/这个人"的消歧**：
+   - 从历史中找出最近提到的人物
+   - 示例：历史中有"李建国"，则"那个人的原名是什么？"应拆分为"李建国的原名是什么？"
+
+5. **优先级**：
+   - 如果历史记录中反复出现某个人物（如"老李"、"李建国"、"建国哥"），则"用户"很可能指的就是这个人
+   - 如果无法从历史中确定指代对象，保留原问题，但在reason中说明"无法确定指代对象"
+
+## 指令：
+你是一个智能数据拆分助手，请根据数据特性判断输入属于哪种类型：
+单跳（Single-hop）
+    描述：问题或数据只需要通过一步即可得到答案或完成拆分，不依赖其他信息。
+    拆分策略：直接识别核心信息或关键字段，生成可独立处理的片段。
+    示例：
+    输入数据："请列出今年诺贝尔物理学奖的得主"
+    拆分结果：[
+          {
+            "id": "Q1",
+            "question": "今年诺贝尔物理学奖得主是谁",
+            "type": "单跳’",
+          }
+        ]
+    注意： 当遇到上下文依赖问题时，明确指出缺失的信息类型并且，question可填写输入问题
+多跳（Multi-hop）:
+    描述：问题或数据需要通过多步推理或跨多个信息源才能得到答案。
+    拆分策略：将问题拆解为多个子问题，每个子问题对应一个独立处理步骤，需要具备推理链条与逻辑连接数量。
+    示例：
+        输入数据："今年诺贝尔物理学奖得主的研究领域及代表性成果"
+        拆分结果：
+        [
+          {
+            "id": "Q1",
+            "question": 今年诺贝尔物理学奖得主是谁？",
+            "type": "多跳’",
+          },
+          {
+            "id": "Q2",
+            "question": "该得主的研究领域是什么？",
+            "type": "多跳’",
+          },
+          {
+            "id": "Q3",
+            "question": "该得主的代表性成果有哪些？",
+            "type": "多跳’"
+          }
+        ]
+开放域（Open-domain）:
+    描述：问题或数据不局限于特定知识库，需要从大范围信息中检索和生成答案，而不是从一个已知的小范围数据源中查找。。
+    拆分策略：根据主题或关键实体拆分，同时保留上下文以便检索外部知识，问题涉及一般性、常识性、跨学科内容，可能是开放式回答（描述性、推理性、综合性）
+            需要外部知识检索或推理才能确定，比如：“为什么人类需要睡眠？”、“量子计算与经典计算的主要区别是什么？”。
+    示例：
+        输入数据："介绍量子计算的最新研究进展"
+        拆分结果：
+         [
+          {
+            "id": "Q1",
+            "question": 量子计算的基本概念是什么？",
+            "type": "开放域’",
+          },
+          {
+            "id": "Q2",
+            "question": "当前量子计算的主要研究方向有哪些？",
+            "type": "开放域’",
+          },
+          {
+            "id": "Q3",
+            "question": "近期在量子计算领域有哪些重大进展？",
+            "type": "开放域’",
+          }
+        ]
+
+时间（Temporal）:
+    描述：问题或数据涉及时间维度，需要按时间顺序或时间点拆分。
+    拆分策略：根据事件时间或时间段拆分为独立条目或问题。
+    示例：
+        输入数据："列出苹果公司过去五年的重大事件"
+        拆分结果：
+        [
+          {
+            "id": "Q1",
+            "question": 苹果公司2019年的重大事件有哪些？",
+            "type": "时间’",
+          },
+          {
+            "id": "Q2",
+            "question": "苹果公司2020年的重大事件有哪些？",
+            "type": "时间’",
+          },
+          {
+            "id": "Q3",
+            "question": "苹果公司2021年的重大事件有哪些？",
+            "type": "时间’",
+          },
+          {
+            "id": "Q3",
+            "question": "苹果公司2022年的重大事件有哪些？",
+            "type": "时间’",
+          }
+          ,
+          {
+            "id": "Q4",
+            "question": "苹果公司2023年的重大事件有哪些？",
+            "type": "时间’",
+          }
+        ]
+
+输出要求：
+- 每个子问题包括：
+  - `id`: 子问题编号（Q1, Q2...）
+  - `question`: 子问题内容
+  - `type`: 类型（事实检索 / 澄清 / 定义 / 比较 / 行动建议等）
+  - `reason`: 拆分的理由（为什么要这样拆）
+- 格式案例:
+[
+          {
+            "id": "Q1",
+            "question": 量子计算的基本概念是什么？",
+            "type": "开放域’",
+          },
+          {
+            "id": "Q2",
+            "question": "当前量子计算的主要研究方向有哪些？",
+            "type": "开放域’",
+          },
+          {
+            "id": "Q3",
+            "question": "近期在量子计算领域有哪些重大进展？",
+            "type": "开放域’",
+          }
+]
+- 必须通过json.loads()的格式支持的形式输出
+- 必须通过json.loads()的格式支持的形式输出,响应必须是与此确切模式匹配的有效JSON对象。不要在JSON之前或之后包含任何文本。
+
+## 指代消歧示例（重要）：
+示例1 - "用户"的消歧：
+输入历史：[{'Query': '老李的原名叫什么？', 'Answer': '李建国'}, {'Query': '老李的同事叫他什么？', 'Answer': '建国哥'}]
+输入问题："用户是谁？"
+输出：
+[
+  {
+    "id": "Q1",
+    "question": "李建国是谁？",
+    "type": "单跳",
+    "reason": "历史中反复提到'老李/李建国/建国哥'，'用户'指的就是对话发起者李建国"
+  }
+]
+
+示例2 - "我"的消歧：
+输入历史：[{'Query': '张曼玉推荐了什么书？', 'Answer': '《小王子》'}]
+输入问题："我推荐的书是什么？"
+输出：
+[
+  {
+    "id": "Q1",
+    "question": "张曼玉推荐的书是什么？",
+    "type": "单跳",
+    "reason": "历史中提到张曼玉推荐了书，'我'指的就是张曼玉"
+  }
+]
+
+- 关键的JSON格式要求
+1.JSON结构仅使用标准ASCII双引号（“）-切勿使用中文引号（“”）或其他Unicode引号
+2.如果提取的语句文本包含引号，请使用反斜杠（\“）正确转义它们
+3.确保所有JSON字符串都正确关闭并以逗号分隔
+4.JSON字符串值中不包括换行符
+5.正确转义的例子：“statement”：“Zhang Xinhua said：\”我非常喜欢这本书\""
+6.不允许输出```json```相关符号，如```json```、``````、```python```、```javascript```、```html```、```css```、```sql```、```java```、```c```、```c++```、```c#```、```ruby```
diff --git a/api/app/core/memory/read_services/content_search.py b/api/app/core/memory/read_services/content_search.py
index 69ca6b11..58356e84 100644
--- a/api/app/core/memory/read_services/content_search.py
+++ b/api/app/core/memory/read_services/content_search.py
@@ -1,37 +1,28 @@
 import asyncio
 import logging
 import math
-import time
-
-from pydantic import BaseModel, Field
 
 from app.core.memory.enums import Neo4jNodeType
-from app.core.memory.llm_tools import OpenAIEmbedderClient
 from app.core.memory.memory_service import MemoryContext
-from app.core.memory.models.service_models import Memory
+from app.core.memory.models.service_models import Memory, MemorySearchResult
 from app.core.memory.read_services.result_builder import data_builder_factory
+from app.core.models import RedBearEmbeddings
 from app.repositories.neo4j.graph_search import search_graph, search_graph_by_embedding
 from app.repositories.neo4j.neo4j_connector import Neo4jConnector
 
 logger = logging.getLogger(__name__)
 
-
-class MemorySearchResult(BaseModel):
-    memories: dict[str, list[dict]] = Field(default_factory=dict)
-    content: str = Field(default="")
-    count: int = Field(default=0)
+DEFAULT_ALPHA = 0.7
+DEFAULT_FULLTEXT_SCORE_THRESHOLD = 1
+DEFAULT_COSINE_SCORE_THRESHOLD = 0.5
+DEFAULT_CONTENT_SCORE_THRESHOLD = 0.5
 
 
 class Neo4jSearchService:
-    DEFAULT_ALPHA = 0.6
-    DEFAULT_FULLTEXT_SCORE_THRESHOLD = 1
-    DEFAULT_COSINE_SCORE_THRESHOLD = 0.5
-    DEFAULT_CONTENT_SCORE_THRESHOLD = 0.5
-
     def __init__(
             self,
             ctx: MemoryContext,
-            embedder: OpenAIEmbedderClient,
+            embedder: RedBearEmbeddings,
             includes: list[Neo4jNodeType] | None = None,
             alpha: float = DEFAULT_ALPHA,
             fulltext_score_threshold: float = DEFAULT_FULLTEXT_SCORE_THRESHOLD,
@@ -44,7 +35,7 @@ class Neo4jSearchService:
         self.cosine_score_threshold = cosine_score_threshold
         self.content_score_threshold = content_score_threshold
 
-        self.embedder: OpenAIEmbedderClient = embedder
+        self.embedder: RedBearEmbeddings = embedder
         self.connector: Neo4jConnector | None = None
 
         self.includes = includes
@@ -121,9 +112,12 @@ class Neo4jSearchService:
             kw = float(combined[item_id].get("kw_score", 0) or 0)
             emb = float(combined[item_id].get("embedding_score", 0) or 0)
             base = self.alpha * emb + (1 - self.alpha) * kw
-            combined[item_id]["content_score"] = base + min(1 - base, 0.1 * kw * emb)
+            combined[item_id]["content_score"] = base + min(1 - base, kw * emb)
         results = sorted(combined.values(), key=lambda x: x["content_score"], reverse=True)
-        # results = [res for res in results if res["content_score"] > self.content_score_threshold]
+        # results = [
+        #     res for res in results
+        #     if res["content_score"] > self.content_score_threshold
+        # ]
         results = results[:limit]
 
         logger.info(
@@ -137,14 +131,14 @@ class Neo4jSearchService:
             return items
         scores = [float(it.get("score", 0) or 0) for it in items]
         for it, s in zip(items, scores):
-            it[f"normalized_kw_score"] = 1 / (1 + math.exp(-(s - self.fulltext_score_threshold) / 2))
+            it[f"normalized_kw_score"] = 1 / (1 + math.exp(-(s - self.fulltext_score_threshold) / 2)) if s else 0
         return items
 
     async def search(
             self,
             query: str,
             limit: int = 10,
-    ) -> list[Memory]:
+    ) -> MemorySearchResult:
         async with Neo4jConnector() as connector:
             self.connector = connector
             kw_task = self._keyword_search(query, limit)
@@ -175,4 +169,12 @@ class Neo4jSearchService:
                     query=query
                 ))
         memories.sort(key=lambda x: x.score, reverse=True)
-        return memories[:limit]
+        return MemorySearchResult(memories=memories[:limit])
+
+
+class RAGSearchService:
+    def __init__(self, ctx: MemoryContext):
+        pass
+
+    async def search(self) -> MemorySearchResult:
+        pass
diff --git a/api/app/core/memory/read_services/memory_search/__init__.py b/api/app/core/memory/read_services/memory_search/__init__.py
deleted file mode 100644
index e69de29b..00000000
diff --git a/api/app/core/memory/read_services/memory_search/content_search.py b/api/app/core/memory/read_services/memory_search/content_search.py
deleted file mode 100644
index f5e58696..00000000
--- a/api/app/core/memory/read_services/memory_search/content_search.py
+++ /dev/null
@@ -1,14 +0,0 @@
-# -*- coding: UTF-8 -*-
-# Author: Eternity
-# @Email: 1533512157@qq.com
-# @Time : 2026/4/9 16:48
-from app.core.memory.llm_tools import OpenAIEmbedderClient
-from app.core.memory.memory_service import MemoryContext
-
-
-class ContentSearch:
-    def __init__(self, ctx: MemoryContext):
-        self.ctx = ctx
-
-    async def search(self, query):
-        pass
\ No newline at end of file
diff --git a/api/app/core/memory/read_services/memory_search/perceptual_search.py b/api/app/core/memory/read_services/memory_search/perceptual_search.py
deleted file mode 100644
index db81e2f8..00000000
--- a/api/app/core/memory/read_services/memory_search/perceptual_search.py
+++ /dev/null
@@ -1,228 +0,0 @@
-import asyncio
-import logging
-from typing import Any
-
-from pydantic import BaseModel
-
-from app.core.memory.llm_tools import OpenAIEmbedderClient
-from app.core.memory.memory_service import MemoryContext
-from app.core.memory.utils.data import escape_lucene_query
-from app.repositories.neo4j.graph_search import search_perceptual, search_perceptual_by_embedding
-from app.repositories.neo4j.neo4j_connector import Neo4jConnector
-
-logger = logging.getLogger(__name__)
-
-
-class PerceptualResult(BaseModel):
-    memories: list[dict[str, Any]] = []
-    content: str = ""
-    keyword_raw: int = 0
-    embedding_raw: int = 0
-
-
-class PerceptualRetrieverService:
-    DEFAULT_ALPHA = 0.6
-    DEFAULT_FULLTEXT_SCORE_THRESHOLD = 0.5
-    DEFAULT_COSINE_SCORE_THRESHOLD = 0.7
-
-    def __init__(
-            self,
-            ctx: MemoryContext,
-            embedder: OpenAIEmbedderClient,
-            alpha: float = DEFAULT_ALPHA,
-            fulltext_score_threshold: float = DEFAULT_FULLTEXT_SCORE_THRESHOLD,
-            cosine_score_threshold: float = DEFAULT_COSINE_SCORE_THRESHOLD
-    ):
-        self.ctx = ctx
-        self.alpha = alpha
-        self.fulltext_score_threshold = fulltext_score_threshold
-        self.cosine_score_threshold = cosine_score_threshold
-
-        self.embedder: OpenAIEmbedderClient = embedder
-        self.connector = Neo4jConnector()
-
-    async def search(
-            self,
-            query: str,
-            keywords: list[str] | None = None,
-            limit: int = 10
-    ) -> PerceptualResult:
-        if keywords is None:
-            keywords = [query] if query else []
-
-        try:
-            kw_task = self._keyword_search(keywords, limit)
-            emb_task = self._embedding_search(query, limit)
-            kw_results, emb_results = await asyncio.gather(kw_task, emb_task, return_exceptions=True)
-            if isinstance(kw_results, Exception):
-                logger.warning(f"[PerceptualSearch] keyword search error: {kw_results}")
-                kw_results = []
-            if isinstance(emb_results, Exception):
-                logger.warning(f"[PerceptualSearch] embedding search error: {emb_results}")
-                emb_results = []
-
-            reranked = self._rerank(kw_results, emb_results, limit)
-
-            memories = []
-            content_parts = []
-            for record in reranked:
-                fmt = self._format_result(record)
-                fmt["score"] = round(record.get("content_score", 0), 4)
-                memories.append(fmt)
-                content_parts.append(self._build_content_text(fmt))
-
-            logger.info(
-                f"[PerceptualSearch] {len(memories)} results after rerank "
-                f"(keyword_raw={len(kw_results)}, embedding_raw={len(emb_results)})"
-            )
-            return PerceptualResult(
-                memories=memories,
-                content="\n\n".join(content_parts),
-                keyword_raw=len(kw_results),
-                embedding_raw=len(emb_results),
-            )
-        except Exception as e:
-            logger.error(f"[PerceptualSearch] search failed: {e}", exc_info=True)
-            return PerceptualResult()
-        finally:
-            await self.connector.close()
-
-    async def _keyword_search(
-            self,
-            keywords: list[str],
-            limit: int
-    ) -> list[dict]:
-        seen_ids: set = set()
-        all_results: list[dict] = []
-
-        async def _one(kw: str):
-            escaped = escape_lucene_query(kw)
-            if not escaped.strip():
-                return []
-            r = await search_perceptual(
-                connector=self.connector, q=escaped,
-                end_user_id=self.ctx.end_user_id, limit=limit
-            )
-            perceptuals = r.get("perceptuals", [])
-            return [perceptual for perceptual in perceptuals if perceptual["score"] > self.fulltext_score_threshold]
-
-        tasks = [_one(kw) for kw in keywords]
-        batch = await asyncio.gather(*tasks, return_exceptions=True)
-
-        for result in batch:
-            if isinstance(result, Exception):
-                logger.warning(f"[PerceptualSearch] keyword sub-query error: {result}")
-                continue
-            for rec in result:
-                rid = rec.get("id", "")
-                if rid and rid not in seen_ids:
-                    seen_ids.add(rid)
-                    all_results.append(rec)
-        all_results.sort(key=lambda x: float(x.get("score", 0)), reverse=True)
-        return all_results[:limit]
-
-    async def _embedding_search(
-            self,
-            query: str,
-            limit: int
-    ) -> list[dict]:
-        r = await search_perceptual_by_embedding(
-            connector=self.connector,
-            embedder_client=self.embedder,
-            query_text=query,
-            end_user_id=self.ctx.end_user_id,
-            limit=limit
-        )
-        perceptuals = r.get("perceptuals", [])
-        return [perceptual for perceptual in perceptuals if perceptual["score"] > self.cosine_score_threshold]
-
-    def _rerank(
-            self,
-            keyword_results: list[dict],
-            embedding_results: list[dict],
-            limit: int,
-    ) -> list[dict]:
-        keyword_results = self._normalize_scores(keyword_results)
-        embedding_results = self._normalize_scores(embedding_results)
-
-        kw_norm_map = {}
-        for item in keyword_results:
-            item_id = item["id"]
-            kw_norm_map[item_id] = float(item.get("normalized_score", 0))
-
-        emb_norm_map = {}
-        for item in embedding_results:
-            item_id = item["id"]
-            emb_norm_map[item_id] = float(item.get("normalized_score", 0))
-
-        combined = {}
-        for item in keyword_results:
-            item_id = item["id"]
-            combined[item_id] = item.copy()
-            combined[item_id]["kw_score"] = kw_norm_map.get(item_id, 0)
-            combined[item_id]["embedding_score"] = emb_norm_map.get(item_id, 0)
-
-        for item in embedding_results:
-            item_id = item["id"]
-            if item_id in combined:
-                combined[item_id]["embedding_score"] = emb_norm_map.get(item_id, 0)
-            else:
-                combined[item_id] = item.copy()
-                combined[item_id]["kw_score"] = kw_norm_map.get(item_id, 0)
-                combined[item_id]["embedding_score"] = emb_norm_map.get(item_id, 0)
-
-        for item in combined.values():
-            kw = float(item.get("kw_score", 0) or 0)
-            emb = float(item.get("embedding_score", 0) or 0)
-            item["content_score"] = self.alpha * emb + (1 - self.alpha) * kw
-
-        results = list(combined.values())
-        results.sort(key=lambda x: x["content_score"], reverse=True)
-        results = results[:limit]
-
-        logger.info(
-            f"[PerceptualSearch] rerank: merged={len(combined)}, after_threshold={len(results)} "
-            f"(alpha={self.alpha})"
-        )
-        return results
-
-    @staticmethod
-    def _normalize_scores(items: list[dict], field: str = "score") -> list[dict]:
-        """Min-max 归一化，将分数线性映射到 [0, 1]。"""
-        if not items:
-            return items
-        scores = [float(it.get(field, 0) or 0) for it in items]
-        min_s = min(scores)
-        max_s = max(scores)
-        diff = max_s - min_s
-        for it, s in zip(items, scores):
-            it[f"normalized_{field}"] = (s - min_s) / diff if diff > 0 else 1.0
-        return items
-
-    @staticmethod
-    def _format_result(record: dict) -> dict:
-        return {
-            "id": record.get("id", ""),
-            "perceptual_type": record.get("perceptual_type", ""),
-            "file_name": record.get("file_name", ""),
-            "file_path": record.get("file_path", ""),
-            "summary": record.get("summary", ""),
-            "topic": record.get("topic", ""),
-            "domain": record.get("domain", ""),
-            "keywords": record.get("keywords", []),
-            "created_at": str(record.get("created_at", "")),
-            "file_type": record.get("file_type", ""),
-            "score": record.get("score", 0),
-        }
-
-    @staticmethod
-    def _build_content_text(formatted: dict) -> str:
-        content_text = (f"<history-file-info>"
-                        f"<file-name>{formatted["file_name"]}</file-name>"
-                        f"<file-path>{formatted["file_path"]}</file-path>"
-                        f"<file-type>{formatted["file_type"]}</file-type>"
-                        f"<file-topic>{formatted["topic"]}</file-topic>"
-                        f"<file-domain>{formatted["keywords"]}</file-domain>"
-                        f"<file-summary>{formatted["summary"]}</file-summary>"
-                        f"</history-file-info>")
-        return content_text
diff --git a/api/app/core/memory/read_services/query_preprocessor.py b/api/app/core/memory/read_services/query_preprocessor.py
index 02d757c9..123cae40 100644
--- a/api/app/core/memory/read_services/query_preprocessor.py
+++ b/api/app/core/memory/read_services/query_preprocessor.py
@@ -1,11 +1,13 @@
-# -*- coding: UTF-8 -*-
-# Author: Eternity
-# @Email: 1533512157@qq.com
-# @Time : 2026/4/8 18:11
+import logging
 import re
 
+from app.core.memory.prompt import prompt_manager
+from app.core.memory.utils.llm.llm_utils import StructResponse
+from app.core.models import RedBearLLM
 from app.schemas.memory_agent_schema import AgentMemoryDataset
 
+logger = logging.getLogger(__name__)
+
 
 class QueryPreprocessor:
     @staticmethod
@@ -16,3 +18,22 @@ class QueryPreprocessor:
 
         text = re.sub(rf"{"|".join(AgentMemoryDataset.PRONOUN)}", AgentMemoryDataset.NAME, text)
         return text
+
+    @staticmethod
+    async def split(query: str, llm_client: RedBearLLM):
+        system_prompt = prompt_manager.render(
+            name="problem_split",
+            history=[],
+            sentence=query,
+        )
+        messages = [{"role": "system", "content": system_prompt}]
+        try:
+            sub_queries = await llm_client.ainvoke(messages) | StructResponse(mode='json')
+        except Exception as e:
+            logger.error(f"[QueryPreprocessor] Sub-question segmentation failed - {e}")
+            sub_queries = None
+        return sub_queries or query
+
+    @staticmethod
+    async def extension(query: str, llm_client: RedBearLLM):
+        pass
diff --git a/api/app/core/memory/read_services/result_builder.py b/api/app/core/memory/read_services/result_builder.py
index 10ff8c86..949ff3ed 100644
--- a/api/app/core/memory/read_services/result_builder.py
+++ b/api/app/core/memory/read_services/result_builder.py
@@ -114,7 +114,7 @@ class PerceptualBuilder(BaseBuilder):
                 f"<domain>{self.record.get('domain')}</domain>"
                 f"<keywords>{self.record.get('keywords')}</keywords>"
                 f"<file-type>{self.record.get('file_type')}</file-type>"
-                "</<history-file-info>")
+                "</history-file-info>")
 
 
 class CommunityBuilder(BaseBuilder):
diff --git a/api/app/core/memory/read_services/retrieval_summary.py b/api/app/core/memory/read_services/retrieval_summary.py
new file mode 100644
index 00000000..6b166cf2
--- /dev/null
+++ b/api/app/core/memory/read_services/retrieval_summary.py
@@ -0,0 +1,11 @@
+from app.core.models import RedBearLLM
+
+
+class RetrievalSummaryProcessor:
+    @staticmethod
+    def summary(content: str, llm_client: RedBearLLM):
+        return
+
+    @staticmethod
+    def verify(content: str, llm_client: RedBearLLM):
+        return
\ No newline at end of file
diff --git a/api/app/core/memory/utils/llm/llm_utils.py b/api/app/core/memory/utils/llm/llm_utils.py
index 19d76d68..c4eee82f 100644
--- a/api/app/core/memory/utils/llm/llm_utils.py
+++ b/api/app/core/memory/utils/llm/llm_utils.py
@@ -1,4 +1,7 @@
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Literal, Type
+
+from json_repair import json_repair
+from langchain_core.messages import AIMessage
 
 from app.core.memory.llm_tools.openai_client import OpenAIClient
 from app.core.models.base import RedBearModelConfig
@@ -13,6 +16,27 @@ async def handle_response(response: type[BaseModel]) -> dict:
     return response.model_dump()
 
 
+class StructResponse:
+    def __init__(self, mode: Literal["json", "pydantic"], model: Type[BaseModel] = None):
+        self.mode = mode
+        if mode == "pydantic" and model is None:
+            raise ValueError("Pydantic model is required")
+
+        self.model = model
+
+    def __ror__(self, other: AIMessage):
+        if not isinstance(other, AIMessage):
+            raise RuntimeError(f"Unsupported struct type {type(other)}")
+        text = ''
+        for block in other.content_blocks:
+            if block.get("type") == "text":
+                text += block.get("text", "")
+        fixed_json = json_repair.repair_json(text, return_objects=True)
+        if self.mode == "json":
+            return fixed_json
+        return self.model.model_validate(fixed_json)
+
+
 class MemoryClientFactory:
     """
     Factory for creating LLM, embedder, and reranker clients.
@@ -24,21 +48,21 @@ class MemoryClientFactory:
         >>> llm_client = factory.get_llm_client(model_id)
         >>> embedder_client = factory.get_embedder_client(embedding_id)
     """
-    
+
     def __init__(self, db: Session):
         from app.services.memory_config_service import MemoryConfigService
         self._config_service = MemoryConfigService(db)
-    
+
     def get_llm_client(self, llm_id: str) -> OpenAIClient:
         """Get LLM client by model ID."""
         if not llm_id:
             raise ValueError("LLM ID is required")
-        
+
         try:
             model_config = self._config_service.get_model_config(llm_id)
         except Exception as e:
             raise ValueError(f"Invalid LLM ID '{llm_id}': {str(e)}") from e
-        
+
         try:
             return OpenAIClient(
                 RedBearModelConfig(
@@ -52,19 +76,19 @@ class MemoryClientFactory:
         except Exception as e:
             model_name = model_config.get('model_name', 'unknown')
             raise ValueError(f"Failed to initialize LLM client for model '{model_name}': {str(e)}") from e
-    
+
     def get_embedder_client(self, embedding_id: str):
         """Get embedder client by model ID."""
         from app.core.memory.llm_tools.openai_embedder import OpenAIEmbedderClient
-        
+
         if not embedding_id:
             raise ValueError("Embedding ID is required")
-        
+
         try:
             embedder_config = self._config_service.get_embedder_config(embedding_id)
         except Exception as e:
             raise ValueError(f"Invalid embedding ID '{embedding_id}': {str(e)}") from e
-        
+
         try:
             return OpenAIEmbedderClient(
                 RedBearModelConfig(
@@ -77,17 +101,17 @@ class MemoryClientFactory:
         except Exception as e:
             model_name = embedder_config.get('model_name', 'unknown')
             raise ValueError(f"Failed to initialize embedder client for model '{model_name}': {str(e)}") from e
-    
+
     def get_reranker_client(self, rerank_id: str) -> OpenAIClient:
         """Get reranker client by model ID."""
         if not rerank_id:
             raise ValueError("Rerank ID is required")
-        
+
         try:
             model_config = self._config_service.get_model_config(rerank_id)
         except Exception as e:
             raise ValueError(f"Invalid rerank ID '{rerank_id}': {str(e)}") from e
-        
+
         try:
             return OpenAIClient(
                 RedBearModelConfig(
diff --git a/api/app/repositories/neo4j/graph_search.py b/api/app/repositories/neo4j/graph_search.py
index 336f4134..354c0e23 100644
--- a/api/app/repositories/neo4j/graph_search.py
+++ b/api/app/repositories/neo4j/graph_search.py
@@ -8,6 +8,7 @@ import numpy as np
 from app.core.memory.enums import Neo4jNodeType
 from app.core.memory.llm_tools import OpenAIEmbedderClient
 from app.core.memory.utils.data.text_utils import escape_lucene_query
+from app.core.models import RedBearEmbeddings
 from app.repositories.neo4j.cypher_queries import (
     EXPAND_COMMUNITY_STATEMENTS,
     SEARCH_CHUNK_BY_CHUNK_ID,
@@ -358,7 +359,7 @@ async def search_by_embedding(
             USER_ID_QUERY_CYPHER_MAPPING[node_type],
             end_user_id=end_user_id,
         )
-        records = [record for record in records if record if record["embedding"] is not None]
+        records = [record for record in records if record and record.get("embedding") is not None]
         ids = [item['id'] for item in records]
         vectors = [item['embedding'] for item in records]
         sim_res = cosine_similarity_search(query_embedding, vectors, limit=limit)
@@ -469,7 +470,7 @@ async def search_graph(
 
 async def search_graph_by_embedding(
         connector: Neo4jConnector,
-        embedder_client,
+        embedder_client: RedBearEmbeddings | OpenAIEmbedderClient,
         query_text: str,
         end_user_id: str,
         limit: int = 50,
@@ -495,7 +496,10 @@ async def search_graph_by_embedding(
             Neo4jNodeType.PERCEPTUAL
         ]
 
-    embeddings = await embedder_client.response([query_text])
+    if isinstance(embedder_client, RedBearEmbeddings):
+        embeddings = embedder_client.embed_documents([query_text])
+    else:
+        embeddings = await embedder_client.response([query_text])
     if not embeddings or not embeddings[0]:
         logger.warning(f"search_graph_by_embedding: embedding generation failed for '{query_text[:50]}'")
         return {search_key: [] for search_key in include}
diff --git a/api/app/services/prompt/prompt_optimizer_system.jinja2 b/api/app/services/prompt/prompt_optimizer_system.jinja2
index 39a4ba68..5611ae94 100644
--- a/api/app/services/prompt/prompt_optimizer_system.jinja2
+++ b/api/app/services/prompt/prompt_optimizer_system.jinja2
@@ -34,7 +34,7 @@ Readability Guideline: Ensure optimized prompts have good readability and logica
 Constraint Handling Guideline: Do not mention variable-related limitations under the [Constraints] label.{% endraw %}{% endif %}
 
 Constraints
-Output Constraint: Must output in JSON format including the fields "prompt" and "desc".
+Output Constraint: Must output in JSON format including the string fields "prompt" and "desc".
 Content Constraint: Must not include any explanations, analyses, or additional comments.
 Language Constraint: Must use clear and concise language.
 {% if skill != true %}Completeness Constraint: Must fully define all missing elements (input details, output format, constraints, etc.).{% endif %}