diff --git a/.gitignore b/.gitignore
index 9e8ca1d..4120caa 100644
--- a/.gitignore
+++ b/.gitignore
@@ -7,16 +7,14 @@
 /*
 
 # 2. 放行需要的文件夹及其内容
-!app/
-!app/**
+!backend/
+!backend/**
 !frontend/
 !frontend/**
 !scripts/
 !scripts/**
 !rag_indexer/
 !rag_indexer/**
-!rag_core/
-!rag_core/**
 !docker/
 !docker/**
 !.gitea/
diff --git a/backend/app/__init__.py b/backend/app/__init__.py
new file mode 100644
index 0000000..c32b8e6
--- /dev/null
+++ b/backend/app/__init__.py
@@ -0,0 +1,8 @@
+"""
+AI Agent 应用模块
+"""
+
+from ..agent import AIAgentService
+from ..graph.graph_tools import AVAILABLE_TOOLS, TOOLS_BY_NAME
+
+__all__ = ["AIAgentService", "AVAILABLE_TOOLS", "TOOLS_BY_NAME"]
diff --git a/backend/app/agent/__init__.py b/backend/app/agent/__init__.py
new file mode 100644
index 0000000..a5e2eba
--- /dev/null
+++ b/backend/app/agent/__init__.py
@@ -0,0 +1,7 @@
+"""
+Agent 子模块
+"""
+
+from .service import AIAgentService
+
+__all__ = ["AIAgentService"]
diff --git a/backend/app/agent/history.py b/backend/app/agent/history.py
new file mode 100644
index 0000000..e619da9
--- /dev/null
+++ b/backend/app/agent/history.py
@@ -0,0 +1,185 @@
+"""
+历史对话查询模块
+利用 LangGraph 的 checkpointer 获取对话历史和摘要
+"""
+
+from typing import List, Dict, Any
+from ..logger import error # 保持兼容，或者替换为 logger
+
+class ThreadHistoryService:
+    """线程历史查询服务"""
+    
+    def __init__(self, checkpointer):
+        self.checkpointer = checkpointer
+    
+    async def get_user_threads(self, user_id: str, limit: int = 50) -> List[Dict[str, Any]]:
+        """
+        获取指定用户的所有线程摘要信息
+        
+        Args:
+            user_id: 用户 ID
+            limit: 返回数量限制
+            
+        Returns:
+            线程列表，每个包含 thread_id, last_updated, summary, message_count
+        """
+        try:
+            # 查询 checkpoints 表获取用户的线程列表
+            async with self.checkpointer.conn.cursor() as cur:
+                # 在较新的 LangGraph 版本中，AsyncPostgresSaver 创建的 checkpoints 表
+                # 没有 created_at 列，而是使用 checkpoint_id 作为时间排序依据。
+                # 我们可以直接按 thread_id 去重，并用 checkpoint_id 排序。
+                # 另外，用户的 metadata 存储在 metadata JSONB 列中。
+                query = """
+                    SELECT 
+                        thread_id,
+                        MAX(checkpoint_id) as last_updated
+                    FROM checkpoints
+                    WHERE metadata->>'user_id' = %s
+                    GROUP BY thread_id
+                    ORDER BY last_updated DESC
+                    LIMIT %s
+                """
+                await cur.execute(query, (user_id, limit))
+                rows = await cur.fetchall()
+                
+                threads = []
+                for row in rows:
+                    thread_id = row['thread_id']
+                    
+                    # 获取该线程的状态
+                    state = await self.checkpointer.aget_tuple({"configurable": {"thread_id": thread_id}})
+                    
+                    if state and hasattr(state, 'checkpoint') and isinstance(state.checkpoint, dict):
+                        messages = state.checkpoint.get("channel_values", {}).get("messages", [])
+                        
+                        if messages:
+                            summary = self._extract_summary(messages)
+                            message_count = len([m for m in messages if hasattr(m, 'type') and m.type in ["human", "ai"]])
+                            
+                            threads.append({
+                                "thread_id": thread_id,
+                                # checkpoint_id 是一个类似于 uuid 的字符串，其中可能包含时间戳信息，也可以直接作为唯一标识
+                                "last_updated": row['last_updated'] if row['last_updated'] else "",
+                                "summary": summary,
+                                "message_count": message_count
+                            })
+                
+                return threads
+                
+        except Exception as e:
+            error(f"获取用户线程列表失败 (user_id={user_id}): {e}")
+            return []
+    
+    async def get_thread_messages(self, thread_id: str) -> List[Dict[str, str]]:
+        """
+        获取指定线程的完整消息历史
+        
+        Args:
+            thread_id: 线程 ID
+            
+        Returns:
+            消息列表，格式 [{"role": "user/assistant", "content": "..."}]
+        """
+        try:
+            state = await self.checkpointer.aget_tuple({"configurable": {"thread_id": thread_id}})
+            
+            if state is None:
+                return []
+            
+            messages = state.checkpoint.get("channel_values", {}).get("messages", []) if hasattr(state, 'checkpoint') and isinstance(state.checkpoint, dict) else []
+            
+            if not messages:
+                return []
+            
+            # 转换 LangChain 消息对象为字典
+            result = []
+            for msg in messages:
+                # 跳过 system 消息
+                if hasattr(msg, 'type') and msg.type == "system":
+                    continue
+                    
+                if hasattr(msg, 'type'):
+                    role = "user" if msg.type == "human" else "assistant" if msg.type == "ai" else msg.type
+                    result.append({
+                        "role": role,
+                        "content": msg.content
+                    })
+                elif isinstance(msg, dict):
+                    role = msg.get("role", msg.get("type", "unknown"))
+                    if role in ["human", "user"]:
+                        role = "user"
+                    elif role in ["ai", "assistant"]:
+                        role = "assistant"
+                    result.append({
+                        "role": role,
+                        "content": msg.get("content", "")
+                    })
+            
+            return result
+            
+        except Exception as e:
+            error(f"获取线程消息历史失败: {e}")
+            return []
+    
+    async def get_thread_summary(self, thread_id: str) -> Dict[str, Any]:
+        """
+        获取线程摘要（用于历史列表展示）
+        
+        Args:
+            thread_id: 线程 ID
+            
+        Returns:
+            包含摘要信息的字典
+        """
+        try:
+            state = await self.checkpointer.aget_tuple({"configurable": {"thread_id": thread_id}})
+            
+            if state is None or not state.values:
+                return {"thread_id": thread_id, "summary": "空对话", "message_count": 0}
+            
+            messages = state.values.get("messages", [])
+            summary = self._extract_summary(messages)
+            message_count = len([m for m in messages if hasattr(m, 'type') and m.type in ["human", "ai"]])
+            
+            # 获取最后更新时间
+            last_updated = ""
+            if state.metadata and "created_at" in state.metadata:
+                last_updated = state.metadata["created_at"].isoformat()
+            
+            return {
+                "thread_id": thread_id,
+                "summary": summary,
+                "message_count": message_count,
+                "last_updated": last_updated
+            }
+            
+        except Exception as e:
+            error(f"获取线程摘要失败: {e}")
+            return {"thread_id": thread_id, "summary": "加载失败", "message_count": 0}
+    
+    def _extract_summary(self, messages: List) -> str:
+        """
+        从消息列表中提取摘要
+        
+        策略：
+        1. 如果有 summarize 节点生成的 summary，优先使用
+        2. 否则使用第一条用户消息的前 50 字
+        """
+        # 查找是否有 summary 字段
+        for msg in messages:
+            if hasattr(msg, 'additional_kwargs') and msg.additional_kwargs.get('summary'):
+                return msg.additional_kwargs['summary']
+            elif isinstance(msg, dict) and msg.get('summary'):
+                return msg['summary']
+        
+        # 使用第一条用户消息作为摘要
+        for msg in messages:
+            if hasattr(msg, 'type') and msg.type == "human":
+                content = msg.content
+                return content[:50] + "..." if len(content) > 50 else content
+            elif isinstance(msg, dict) and msg.get("role") in ["user", "human"]:
+                content = msg.get("content", "")
+                return content[:50] + "..." if len(content) > 50 else content
+        
+        return "空对话"
\ No newline at end of file
diff --git a/backend/app/agent/llm_factory.py b/backend/app/agent/llm_factory.py
new file mode 100644
index 0000000..e0ca134
--- /dev/null
+++ b/backend/app/agent/llm_factory.py
@@ -0,0 +1,57 @@
+# app/llm_factory.py
+import os
+from ..config import ZHIPUAI_API_KEY, DEEPSEEK_API_KEY, VLLM_BASE_URL, LLAMACPP_API_KEY
+from langchain_community.chat_models import ChatZhipuAI
+from langchain_openai import ChatOpenAI
+from pydantic import SecretStr
+
+class LLMFactory:
+    @staticmethod
+    def create_zhipu():
+        api_key = ZHIPUAI_API_KEY
+        if not api_key:
+            raise ValueError("ZHIPUAI_API_KEY not set")
+        return ChatZhipuAI(
+            model="glm-4.7-flash",
+            api_key=api_key,
+            temperature=0.1,
+            max_tokens=4096,
+            timeout=120.0,
+            max_retries=3,
+            streaming=True,
+        )
+
+    @staticmethod
+    def create_deepseek():
+        api_key = DEEPSEEK_API_KEY
+        if not api_key:
+            raise ValueError("DEEPSEEK_API_KEY not set")
+        return ChatOpenAI(
+            base_url="https://api.deepseek.com",
+            api_key=SecretStr(api_key),
+            model="deepseek-reasoner",
+            temperature=0.1,
+            max_tokens=4096,
+            timeout=60.0,
+            max_retries=2,
+            streaming=True,
+        )
+
+    @staticmethod
+    def create_local():
+        base_url = VLLM_BASE_URL
+        return ChatOpenAI(
+            base_url=base_url,
+            api_key=SecretStr(LLAMACPP_API_KEY),
+            model="gemma-4-E4B-it",
+            timeout=60.0,
+            max_retries=2,
+            streaming=True,
+        )
+
+    # 模型创建器映射
+    CREATORS = {
+        "local": create_local,
+        "deepseek": create_deepseek,
+        "zhipu": create_zhipu,
+    }
\ No newline at end of file
diff --git a/backend/app/agent/prompts.py b/backend/app/agent/prompts.py
new file mode 100644
index 0000000..8b05050
--- /dev/null
+++ b/backend/app/agent/prompts.py
@@ -0,0 +1,37 @@
+# app/prompts.py
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+
+def create_system_prompt(tools: list = None) -> ChatPromptTemplate:
+    """
+    创建系统提示模板，可选择动态注入工具描述。
+    """
+    tools_section = ""
+    if tools:
+        tool_descs = []
+        for tool in tools:
+            # 提取工具名称和描述的第一行
+            name = getattr(tool, 'name', None) or getattr(tool, '__name__', 'unknown_tool')
+            desc = (tool.description or "").split('\n')[0]
+            tool_descs.append(f"- {name}: {desc}")
+        tools_section = "\n".join(tool_descs)
+
+    system_template = (
+        "你是一个个人生活助手和数据分析助手，请使用中文交流。\n\n"
+        "【用户背景信息】\n"
+        "以下是对当前用户的已知信息和长期记忆，你必须优先采纳并在回答中体现：\n"
+        "{memory_context}\n"
+        "若包含姓名、偏好等个人信息，请自然融入回应（例如称呼名字、提及偏好）。\n\n"
+        "【可用工具与使用规则】\n"
+        f"{tools_section}\n"
+        "工具调用时请直接返回所需参数，无需额外说明。\n\n"
+        "【回答要求（必须遵守）】\n"
+        "1. 回答必须简洁、直接。\n"
+        "2. 如果你认为该问题需要进行深入的推理或思考，请务必将你的思维链或推理过程用 `<think>` 和 `</think>` 标签包裹起来，放在回答的最前面。\n"
+        "3. 优先利用已知用户信息进行个性化回复。\n"
+        "4. 若无信息可依，礼貌询问或提供通用帮助。"
+    )
+    
+    return ChatPromptTemplate.from_messages([
+        ("system", system_template),
+        MessagesPlaceholder(variable_name="messages")
+    ])
\ No newline at end of file
diff --git a/backend/app/agent/rag_initializer.py b/backend/app/agent/rag_initializer.py
new file mode 100644
index 0000000..b637fc8
--- /dev/null
+++ b/backend/app/agent/rag_initializer.py
@@ -0,0 +1,23 @@
+# app/rag_initializer.py
+from ..rag.tools import create_rag_tool_sync
+from rag_core import create_parent_retriever
+from ..logger import info, warning
+
+async def init_rag_tool(local_llm_creator):
+    """初始化 RAG 工具，失败返回 None"""
+    try:
+        info("🔄 正在初始化 RAG 检索系统...")
+        retriever = create_parent_retriever(
+            collection_name="rag_documents",
+            search_k=5,
+        )
+        rewrite_llm = local_llm_creator()
+        rag_tool = create_rag_tool_sync(
+            retriever, rewrite_llm,
+            num_queries=3, rerank_top_n=5
+        )
+        info("✅ RAG 检索工具初始化成功")
+        return rag_tool
+    except Exception as e:
+        warning(f"⚠️ RAG 检索工具初始化失败: {e}")
+        return None
\ No newline at end of file
diff --git a/backend/app/agent/service.py b/backend/app/agent/service.py
new file mode 100644
index 0000000..2146daf
--- /dev/null
+++ b/backend/app/agent/service.py
@@ -0,0 +1,154 @@
+"""
+AI Agent 服务类 - 支持多模型动态切换
+接收外部传入的 checkpointer，不负责管理连接生命周期
+"""
+
+import json
+
+# 本地模块
+from ..graph.graph_builder import GraphBuilder, GraphContext
+from ..graph.graph_tools import AVAILABLE_TOOLS, TOOLS_BY_NAME
+from .llm_factory import LLMFactory
+from .rag_initializer import init_rag_tool
+from ..logger import info, warning
+
+class AIAgentService:
+    def __init__(self, checkpointer):
+        self.checkpointer = checkpointer
+        self.graphs = {}
+        self.tools = AVAILABLE_TOOLS.copy()
+        self.tools_by_name = TOOLS_BY_NAME.copy()
+
+    async def initialize(self):
+        # 1. 初始化 RAG 工具（如果需要）
+        rag_tool = await init_rag_tool(LLMFactory.create_local)
+        if rag_tool:
+            self.tools.append(rag_tool)
+            self.tools_by_name[rag_tool.name] = rag_tool
+
+        # 2. 构建各模型的 Graph
+        for name, creator in LLMFactory.CREATORS.items():
+            try:
+                info(f"🔄 初始化模型 '{name}'...")
+                llm = creator()
+                builder = GraphBuilder(llm, self.tools, self.tools_by_name).build()
+                graph = builder.compile(checkpointer=self.checkpointer)
+                self.graphs[name] = graph
+                info(f"✅ 模型 '{name}' 初始化成功")
+            except Exception as e:
+                warning(f"⚠️ 模型 '{name}' 初始化失败: {e}")
+
+        if not self.graphs:
+            raise RuntimeError("没有可用的模型")
+        return self
+
+    async def process_message(self, message: str, thread_id: str, model: str = "local", user_id: str = "default_user") -> dict:
+        """处理用户消息，返回包含回复、token统计和耗时的字典"""
+        if model not in self.graphs:
+            # 回退到第一个可用模型
+            available = list(self.graphs.keys())
+            if not available:
+                raise RuntimeError("没有可用的模型")
+            model = available[0]
+            warning(f"模型 '{model}' 不可用，已回退到 '{model}'")
+
+        graph = self.graphs[model]
+        config = {
+            "configurable": {"thread_id": thread_id},
+            "metadata": {"user_id": user_id}
+        }
+        input_state = {"messages": [{"role": "user", "content": message}]}
+        context = GraphContext(user_id=user_id)
+
+        result = await graph.ainvoke(input_state, config=config, context=context)
+
+        reply = result["messages"][-1].content
+        token_usage = result.get("last_token_usage", {})
+        elapsed_time = result.get("last_elapsed_time", 0.0)
+
+        return {
+            "reply": reply,
+            "token_usage": token_usage,
+            "elapsed_time": elapsed_time
+        }
+
+    def _serialize_value(self, value):
+        """递归将 LangChain 对象转换为可 JSON 序列化的格式"""
+        if hasattr(value, 'content'):
+            msg_type = getattr(value, 'type', 'message')
+            return {
+                "role": msg_type,
+                "content": getattr(value, 'content', ''),
+                "additional_kwargs": getattr(value, 'additional_kwargs', {}),
+                "tool_calls": getattr(value, 'tool_calls', [])
+            }
+        elif isinstance(value, dict):
+            return {k: self._serialize_value(v) for k, v in value.items()}
+        elif isinstance(value, (list, tuple)):
+            return [self._serialize_value(item) for item in value]
+        else:
+            try:
+                json.dumps(value)
+                return value
+            except (TypeError, ValueError):
+                return str(value)
+
+    async def process_message_stream(self, message: str, thread_id: str, model_name: str, user_id: str = "default_user"):
+        """流式处理消息，返回异步生成器"""
+        graph = self.graphs.get(model_name)
+        if not graph:
+            raise ValueError(f"模型 '{model_name}' 未找到或未初始化")
+
+        config = {
+            "configurable": {"thread_id": thread_id},
+            "metadata": {"user_id": user_id}
+        }
+        input_state = {"messages": [{"role": "user", "content": message}]}
+        context = GraphContext(user_id=user_id)
+
+        async for chunk in graph.astream(
+            input_state,
+            config=config,
+            context=context,
+            stream_mode=["messages", "updates", "custom"],
+            version="v2",
+            subgraphs=True
+        ):
+            chunk_type = chunk["type"]
+            processed_event = {}
+
+            if chunk_type == "messages":
+                message_chunk, metadata = chunk["data"]
+                node_name = metadata.get("langgraph_node", "unknown")
+                token_content = getattr(message_chunk, 'content', str(message_chunk))
+                reasoning_token = ""
+                if hasattr(message_chunk, 'additional_kwargs'):
+                    reasoning_token = message_chunk.additional_kwargs.get("reasoning_content", "")
+
+                processed_event = {
+                    "type": "llm_token",
+                    "node": node_name,
+                    "token": token_content,
+                    "reasoning_token": reasoning_token,
+                    "metadata": metadata
+                }
+            elif chunk_type == "updates":
+                updates_data = chunk["data"]
+                serialized_data = self._serialize_value(updates_data)
+                processed_event = {
+                    "type": "state_update",
+                    "data": serialized_data
+                }
+                if "messages" in serialized_data:
+                    processed_event["messages"] = serialized_data["messages"]
+            elif chunk_type == "custom":
+                serialized_data = self._serialize_value(chunk["data"])
+                processed_event = {
+                    "type": "custom",
+                    "data": serialized_data
+                }
+            else:
+                continue
+
+            if processed_event:
+                yield processed_event
\ No newline at end of file
diff --git a/backend/app/backend.py b/backend/app/backend.py
new file mode 100644
index 0000000..e7544b7
--- /dev/null
+++ b/backend/app/backend.py
@@ -0,0 +1,212 @@
+"""
+FastAPI 后端 - 支持动态模型切换，使用 PostgreSQL 持久化记忆
+采用依赖注入模式，优雅管理资源生命周期
+"""
+
+import os
+from .config import DB_URI, BACKEND_PORT
+import uuid
+import json
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI, HTTPException, WebSocket, WebSocketDisconnect, Depends, Request, Query
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
+from .agent.service import AIAgentService
+from .agent.history import ThreadHistoryService
+from .logger import info, error
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """应用生命周期管理：创建并注入全局服务"""
+    # 1. 创建数据库连接池并初始化表（仅 checkpointer）
+    async with AsyncPostgresSaver.from_conn_string(DB_URI) as checkpointer:
+        await checkpointer.setup()
+        
+        # 2. 构建 AI Agent 服务
+        agent_service = AIAgentService(checkpointer)
+        await agent_service.initialize()
+        
+        # 3. 创建历史查询服务
+        history_service = ThreadHistoryService(checkpointer)
+        
+        # 4. 将服务实例存入 app.state
+        app.state.agent_service = agent_service
+        app.state.history_service = history_service
+        
+        # 应用运行中...
+        yield
+        
+        # 5. 关闭时自动清理数据库连接（async with 负责）
+        info("🛑 应用关闭，数据库连接池已释放")
+
+app = FastAPI(lifespan=lifespan)
+
+# CORS 中间件（允许前端跨域）
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# ========== 健康检查端点 ==========
+@app.get("/health")
+async def health_check():
+    """健康检查端点，用于 Docker 和 CI/CD 监控"""
+    return {"status": "ok", "service": "ai-agent-backend"}
+
+# ========== Pydantic 模型 ==========
+class ChatRequest(BaseModel):
+    message: str
+    thread_id: str | None = None
+    model: str = "zhipu"
+    user_id: str = "default_user"
+
+class ChatResponse(BaseModel):
+    reply: str
+    thread_id: str
+    model_used: str
+    input_tokens: int = 0
+    output_tokens: int = 0
+    total_tokens: int = 0
+    elapsed_time: float = 0.0
+
+# ========== 依赖注入函数 ==========
+def get_agent_service(request: Request) -> AIAgentService:
+    """从 app.state 中获取全局 AIAgentService 实例"""
+    return request.app.state.agent_service
+
+def get_history_service(request: Request) -> ThreadHistoryService:
+    """从 app.state 中获取全局 ThreadHistoryService 实例"""
+    return request.app.state.history_service
+
+# ========== HTTP 端点 ==========
+@app.post("/chat", response_model=ChatResponse)
+async def chat_endpoint(
+    request: ChatRequest,
+    agent_service: AIAgentService = Depends(get_agent_service)
+):
+    """同步对话接口，支持模型选择"""
+    if not request.message:
+        raise HTTPException(status_code=400, detail="message required")
+
+    thread_id = request.thread_id or str(uuid.uuid4())
+    result = await agent_service.process_message(
+        request.message, thread_id, request.model, request.user_id
+    )
+    
+    # 提取 token 统计信息
+    token_usage = result.get("token_usage", {})
+    input_tokens = token_usage.get('prompt_tokens', token_usage.get('input_tokens', 0))
+    output_tokens = token_usage.get('completion_tokens', token_usage.get('output_tokens', 0))
+    elapsed_time = result.get("elapsed_time", 0.0)
+    
+    actual_model = request.model if request.model in agent_service.graphs else next(iter(agent_service.graphs.keys()))
+    
+    return ChatResponse(
+        reply=result["reply"],
+        thread_id=thread_id,
+        model_used=actual_model,
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        total_tokens=input_tokens + output_tokens,
+        elapsed_time=elapsed_time
+    )
+
+# ========== 历史查询接口 ==========
+@app.get("/threads")
+async def list_threads(
+    user_id: str = Query("default_user", description="用户 ID"),
+    limit: int = Query(50, ge=1, le=200, description="返回数量限制"),
+    history_service: ThreadHistoryService = Depends(get_history_service)
+):
+    """获取当前用户的对话历史列表"""
+    threads = await history_service.get_user_threads(user_id, limit)
+    return {"threads": threads}
+
+@app.get("/thread/{thread_id}/messages")
+async def get_thread_messages(
+    thread_id: str,
+    user_id: str = Query("default_user", description="用户 ID"),
+    history_service: ThreadHistoryService = Depends(get_history_service)
+):
+    """获取指定线程的完整消息历史"""
+    messages = await history_service.get_thread_messages(thread_id)
+    return {"messages": messages}
+
+@app.get("/thread/{thread_id}/summary")
+async def get_thread_summary(
+    thread_id: str,
+    user_id: str = Query("default_user", description="用户 ID"),
+    history_service: ThreadHistoryService = Depends(get_history_service)
+):
+    """获取指定线程的摘要信息"""
+    summary = await history_service.get_thread_summary(thread_id)
+    return summary
+
+# ========== 流式对话接口 ==========
+@app.post("/chat/stream")
+async def chat_stream_endpoint(
+    request: ChatRequest,
+    agent_service: AIAgentService = Depends(get_agent_service)
+):
+    """流式对话接口（SSE）"""
+    if not request.message:
+        raise HTTPException(status_code=400, detail="message required")
+
+    thread_id = request.thread_id or str(uuid.uuid4())
+    
+    async def event_generator():
+        try:
+            async for chunk in agent_service.process_message_stream(
+                request.message, thread_id, request.model, request.user_id
+            ):
+                yield f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n"
+            yield "data: [DONE]\n\n"
+        except Exception as e:
+            error(f"流式响应异常: {e}")
+            yield f"data: {json.dumps({'type': 'error', 'message': str(e)}, ensure_ascii=False)}\n\n"
+            yield "data: [DONE]\n\n"
+    
+    return StreamingResponse(
+        event_generator(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",  # 禁用 Nginx 缓冲
+        }
+    )
+
+# ========== WebSocket 端点（可选） ==========
+@app.websocket("/ws")
+async def websocket_endpoint(
+    websocket: WebSocket,
+    agent_service: AIAgentService = Depends(get_agent_service)
+):
+    await websocket.accept()
+    try:
+        while True:
+            data = await websocket.receive_json()
+            message = data.get("message")
+            thread_id = data.get("thread_id", str(uuid.uuid4()))
+            model = data.get("model", "zhipu")
+            user_id = data.get("user_id", "default_user")
+            if not message:
+                await websocket.send_json({"error": "missing message"})
+                continue
+            reply = await agent_service.process_message(message, thread_id, model, user_id)
+            actual_model = model if model in agent_service.graphs else next(iter(agent_service.graphs.keys()))
+            await websocket.send_json({"reply": reply, "thread_id": thread_id, "model_used": actual_model})
+    except WebSocketDisconnect:
+        pass
+
+if __name__ == "__main__":
+    import uvicorn
+    # 使用环境变量或默认端口 8079（避免与 llama.cpp 的 8081 端口冲突）
+    port = int(BACKEND_PORT)
+    uvicorn.run(app, host="0.0.0.0", port=port)
diff --git a/backend/app/config.py b/backend/app/config.py
new file mode 100644
index 0000000..5b3ea11
--- /dev/null
+++ b/backend/app/config.py
@@ -0,0 +1,50 @@
+"""
+环境变量集中管理模块
+所有配置项统一定义，避免散落在各个文件中
+"""
+
+import os
+
+
+# ========== Graph 执行追踪配置 ==========
+# 是否启用 Graph 流转追踪（通过环境变量控制）
+ENABLE_GRAPH_TRACE = os.getenv("ENABLE_GRAPH_TRACE", "true").lower() == "true"
+
+# ========== 记忆提取配置 ==========
+# 记忆提取间隔：每 N 轮对话生成一次摘要
+MEMORY_SUMMARIZE_INTERVAL = int(os.getenv("MEMORY_SUMMARIZE_INTERVAL", "10"))
+
+# ========== Mem0 记忆层配置 ==========
+# Qdrant 向量数据库地址
+QDRANT_URL = os.getenv("QDRANT_URL", "http://127.0.0.1:6333")
+QDRANT_COLLECTION_NAME = os.getenv("QDRANT_COLLECTION_NAME", "mem0_user_memories")
+QDRANT_API_KEY = os.getenv("QDRANT_API_KEY", "your-qdrant-api-key")
+
+# ==========    llm 配置 ==========
+# LLM 模型配置
+VLLM_BASE_URL = os.getenv("VLLM_BASE_URL", "http://127.0.0.1:8081/v1")
+LLM_API_KEY = os.getenv("LLM_API_KEY", "your-ai-api-key")
+
+# llama.cpp Embedding 服务地址 (用于 Mem0 的向量化)
+LLAMACPP_EMBEDDING_URL = os.getenv("LLAMACPP_EMBEDDING_URL", "http://127.0.0.1:8082/v1")
+LLAMACPP_API_KEY = os.getenv("LLAMACPP_API_KEY", "your-llamacpp-api-key")
+
+# ========== 后端服务配置 ==========
+# 数据库连接字符串
+DB_URI = os.getenv(
+    "DB_URI",
+    "postgresql://postgres:huang1998@ai-postgres:5432/langgraph_db?sslmode=disable"
+)
+# 后端服务端口
+BACKEND_PORT = int(os.getenv("BACKEND_PORT", "8079"))
+
+# ========== 日志配置 ==========
+LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO").upper()
+DEBUG = os.getenv("DEBUG", "false").lower() == "true"
+
+# ========== Reranker 服务配置 ==========
+LLAMACPP_RERANKER_URL = os.getenv("LLAMACPP_RERANKER_URL", "http://127.0.0.1:8083")
+
+# ========== 第三方 API 密钥 ==========
+ZHIPUAI_API_KEY = os.getenv("ZHIPUAI_API_KEY", "")
+DEEPSEEK_API_KEY = os.getenv("DEEPSEEK_API_KEY", "")
diff --git a/backend/app/graph/__init__.py b/backend/app/graph/__init__.py
new file mode 100644
index 0000000..a4c7afc
--- /dev/null
+++ b/backend/app/graph/__init__.py
@@ -0,0 +1,8 @@
+"""
+Graph 子模块
+"""
+
+from .graph_builder import GraphBuilder
+from .state import MessagesState, GraphContext
+
+__all__ = ["GraphBuilder", "MessagesState", "GraphContext"]
diff --git a/backend/app/graph/graph_builder.py b/backend/app/graph/graph_builder.py
new file mode 100644
index 0000000..21abe76
--- /dev/null
+++ b/backend/app/graph/graph_builder.py
@@ -0,0 +1,83 @@
+"""
+LangGraph 状态图构建模块 - 精简版，仅负责组装图
+所有节点逻辑已拆分到独立模块
+"""
+
+from langchain_core.language_models import BaseLLM
+from langgraph.graph import StateGraph, START, END
+from .state import MessagesState, GraphContext
+from ..nodes import (
+    should_continue,
+    create_llm_call_node,
+    create_tool_call_node,
+    create_retrieve_memory_node,
+    create_summarize_node,
+    finalize_node,
+)
+from ..nodes.memory_trigger import memory_trigger_node, set_mem0_client
+from ..memory import Mem0Client
+
+
+class GraphBuilder:
+    """LangGraph 状态图构建器 - 仅负责组装图"""
+
+    def __init__(self, llm: BaseLLM, tools: list, tools_by_name: dict):
+        """
+        初始化构建器
+        
+        Args:
+            llm: 大语言模型实例
+            tools: 工具列表
+            tools_by_name: 名称到工具函数的映射
+        """
+        self.llm = llm
+        self.tools = tools
+        self.tools_by_name = tools_by_name
+        
+        # ⭐ 创建 Mem0 客户端（懒加载，首次使用时初始化）
+        self.mem0_client = Mem0Client(llm)
+
+    def build(self) -> StateGraph:
+        """
+        构建未编译的状态图
+        
+        Returns:
+            StateGraph 实例
+        """
+        # 注入全局客户端
+        set_mem0_client(self.mem0_client)
+        
+        builder = StateGraph(MessagesState, context_schema=GraphContext)
+        
+        # ⭐ 通过工厂函数创建节点（依赖注入）
+        retrieve_memory_node = create_retrieve_memory_node(self.mem0_client)
+        llm_call_node = create_llm_call_node(self.llm, self.tools)
+        tool_call_node = create_tool_call_node(self.tools_by_name)
+        summarize_node = create_summarize_node(self.mem0_client)
+        
+        # 添加节点
+        builder.add_node("retrieve_memory", retrieve_memory_node)
+        builder.add_node("memory_trigger", memory_trigger_node)
+        builder.add_node("llm_call", llm_call_node)
+        builder.add_node("tool_node", tool_call_node)
+        builder.add_node("summarize", summarize_node)
+        builder.add_node("finalize", finalize_node)
+
+        # 添加边
+        builder.add_edge(START, "retrieve_memory")
+        builder.add_edge("retrieve_memory", "memory_trigger")
+        builder.add_edge("memory_trigger", "llm_call")
+        builder.add_conditional_edges(
+            "llm_call",
+            should_continue,
+            {
+                "tool_node": "tool_node",
+                "summarize": "summarize",
+                "finalize": "finalize"
+            }
+        )
+        builder.add_edge("tool_node", "llm_call")
+        builder.add_edge("summarize", "finalize")
+        builder.add_edge("finalize", END)
+        
+        return builder
\ No newline at end of file
diff --git a/backend/app/graph/graph_tools.py b/backend/app/graph/graph_tools.py
new file mode 100644
index 0000000..1cc1e17
--- /dev/null
+++ b/backend/app/graph/graph_tools.py
@@ -0,0 +1,95 @@
+"""
+工具定义模块 - 纯函数工具，无依赖 AIAgent 类
+"""
+
+# 标准库
+from pathlib import Path
+
+# 第三方库
+import pandas as pd
+import pypdf
+import requests
+from bs4 import BeautifulSoup
+from langchain_core.tools import tool
+
+def _file_allow_check(filename: str) -> Path:
+    """检查用户文件名是否位于允许目录 './user_docs' 下，防止路径遍历攻击。"""
+    allowed_dir = Path("./user_docs").resolve()
+    allowed_dir.mkdir(exist_ok=True)
+
+    file_path = (allowed_dir / filename).resolve()
+    if not str(file_path).startswith(str(allowed_dir)):
+        raise ValueError("错误：非法文件路径。")
+
+    if not file_path.exists():
+        raise FileNotFoundError(f"错误：文件 '{filename}' 不存在。")
+
+    return file_path
+
+@tool
+def get_current_temperature(location: str) -> str:
+    """获取指定地点的当前温度。"""
+    return f'当前{location}的温度为25℃'
+
+@tool
+def read_local_file(filename: str) -> str:
+    """读取用户指定名称的本地文本文件内容并返回摘要。"""
+    try:
+        file_path = _file_allow_check(filename)
+        with open(file_path, 'r', encoding='utf-8') as f:
+            content = f.read()
+        return f"文件 '{filename}' 的内容开头：\n{content[:1000]}..."
+    except Exception as e:
+        return f"读取文件时出错：{str(e)}"
+
+@tool
+def read_pdf_summary(filename: str) -> str:
+    """读取PDF文件并返回内容文本摘要。"""
+    try:
+        file_path = _file_allow_check(filename)
+        text = ""
+        with open(file_path, 'rb') as f:
+            reader = pypdf.PdfReader(f)
+            for page in reader.pages[:3]:
+                text += page.extract_text()
+        return f"PDF文件 '{filename}' 的前几页内容：\n{text[:2000]}..."
+    except Exception as e:
+        return f"读取PDF出错：{e}"
+
+@tool
+def read_excel_as_markdown(filename: str) -> str:
+    """读取Excel文件，并将其主要数据转换为Markdown表格格式。"""
+    try:
+        file_path = _file_allow_check(filename)
+        df = pd.read_excel(file_path)
+        markdown_table = df.head(10).to_markdown(index=False)
+        return f"Excel文件 '{filename}' 的数据预览（前10行）：\n{markdown_table}"
+    except Exception as e:
+        return f"读取Excel出错：{e}"
+
+@tool
+def fetch_webpage_content(url: str) -> str:
+    """抓取给定URL的网页正文内容，并返回清晰的纯文本。"""
+    try:
+        response = requests.get(url, timeout=10)
+        response.raise_for_status()
+        soup = BeautifulSoup(response.text, 'html.parser')
+        for script in soup(["script", "style"]):
+            script.decompose()
+        text = soup.get_text()
+        lines = (line.strip() for line in text.splitlines())
+        chunks = (phrase.strip() for line in lines for phrase in line.split("  "))
+        text = '\n'.join(chunk for chunk in chunks if chunk)
+        return f"成功抓取网页 {url}，正文内容开头：\n{text[:1500]}..."
+    except Exception as e:
+        return f"抓取网页时出错：{str(e)}"
+
+# 工具列表和映射（全局常量）
+AVAILABLE_TOOLS = [
+    get_current_temperature,
+    read_local_file,
+    fetch_webpage_content,
+    read_pdf_summary,
+    read_excel_as_markdown
+]
+TOOLS_BY_NAME = {tool.name: tool for tool in AVAILABLE_TOOLS}
diff --git a/backend/app/graph/retrieve_memory.py b/backend/app/graph/retrieve_memory.py
new file mode 100644
index 0000000..36453b0
--- /dev/null
+++ b/backend/app/graph/retrieve_memory.py
@@ -0,0 +1,76 @@
+"""
+记忆检索节点模块
+负责从 Mem0 检索相关长期记忆
+"""
+
+from typing import Any, Dict
+
+# 本地模块
+from .state import MessagesState
+from ..memory.mem0_client import Mem0Client
+from ..utils.logging import log_state_change
+from ..logger import debug
+
+def create_retrieve_memory_node(mem0_client: Mem0Client):
+    """
+    工厂函数：创建记忆检索节点
+    
+    Args:
+        mem0_client: Mem0 客户端实例
+        
+    Returns:
+        异步节点函数
+    """
+    
+    from langchain_core.runnables.config import RunnableConfig
+    
+    async def retrieve_memory(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+        """
+        记忆检索节点 - 使用 Mem0
+        
+        Args:
+            state: 当前对话状态
+            config: 运行时配置
+            
+        Returns:
+            包含 memory_context 的状态更新
+        """
+        log_state_change("retrieve_memory", state, "进入")
+
+        # 从 metadata 中获取 user_id
+        user_id = config.get("metadata", {}).get("user_id", "default_user")
+        
+        # 兼容 dict 和对象两种消息格式
+        last_msg = state["messages"][-1]
+        if isinstance(last_msg, dict):
+            query = str(last_msg.get("content", ""))
+        else:
+            query = str(last_msg.content)
+        memory_text_parts = []
+
+        # 确保 Mem0 已初始化（懒加载）
+        if not mem0_client._initialized:
+            await mem0_client.initialize()
+
+        if mem0_client.mem0:
+            try:
+                # 异步调用 Mem0 语义检索
+                facts = await mem0_client.search_memories(query, user_id=user_id, limit=5)
+                
+                if facts:
+                    memory_text_parts.append(f"【相关长期记忆】\n" + "\n".join(f"- {f}" for f in facts))
+                else:
+                    debug("🔍 [记忆检索] 未找到相关记忆")
+            except Exception as e:
+                from app.logger import warning
+                warning(f"⚠️ Mem0 检索失败: {e}")
+        else:
+            from app.logger import warning
+            warning("⚠️ Mem0 未初始化，跳过记忆检索")
+
+        memory_context = "\n\n".join(memory_text_parts) if memory_text_parts else "暂无用户信息"
+        result = {"memory_context": memory_context}
+        log_state_change("retrieve_memory", {**state, **result}, "离开")
+        return result
+    
+    return retrieve_memory
diff --git a/backend/app/graph/state.py b/backend/app/graph/state.py
new file mode 100644
index 0000000..2fd214e
--- /dev/null
+++ b/backend/app/graph/state.py
@@ -0,0 +1,25 @@
+"""
+LangGraph 状态定义模块
+包含 MessagesState 和 GraphContext
+"""
+
+import operator
+from typing import Annotated
+from typing_extensions import TypedDict
+from dataclasses import dataclass
+from langchain_core.messages import AnyMessage
+
+class MessagesState(TypedDict):
+    """对话状态类型定义"""
+    messages: Annotated[list[AnyMessage], operator.add]
+    llm_calls: int
+    memory_context: str
+    last_token_usage: dict  # 本次调用的 token 使用详情
+    last_elapsed_time: float  # 本次调用耗时（秒）
+    turns_since_last_summary: int  # 距离上次生成摘要的轮数
+
+@dataclass
+class GraphContext:
+    """图执行上下文"""
+    user_id: str
+    # 可扩展更多上下文信息
diff --git a/backend/app/logger.py b/backend/app/logger.py
new file mode 100644
index 0000000..777aeec
--- /dev/null
+++ b/backend/app/logger.py
@@ -0,0 +1,56 @@
+"""
+统一的日志模块 - 基于环境变量控制日志级别
+类似 C# 的条件编译效果，开发时打印详细调试信息，生产环境只输出关键信息
+"""
+
+import os
+from .config import LOG_LEVEL, DEBUG
+import logging
+from typing import Any
+from dotenv import load_dotenv
+
+# 先加载环境变量
+load_dotenv()
+
+# 从环境变量读取日志级别，默认 INFO
+
+
+# 根据环境变量控制是否显示详细调试信息
+DEBUG_MODE = DEBUG
+
+# 创建统一的日志器
+logger = logging.getLogger("ai_agent")
+logger.setLevel(getattr(logging, LOG_LEVEL, logging.INFO))
+
+# 避免重复添加 handler
+if not logger.handlers:
+    handler = logging.StreamHandler()
+    # 重要：handler 也需要设置级别，否则可能继承根 logger 的级别
+    handler.setLevel(getattr(logging, LOG_LEVEL, logging.INFO))
+    formatter = logging.Formatter(
+        fmt="%(asctime)s | %(levelname)-8s | %(name)s | %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S"
+    )
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+
+
+def debug(msg: Any, *args, **kwargs):
+    """调试日志，仅在 DEBUG 环境变量为 true 时打印"""
+    if DEBUG_MODE:
+        logger.debug(msg, *args, **kwargs)
+
+
+def info(msg: Any, *args, **kwargs):
+    """信息日志"""
+    logger.info(msg, *args, **kwargs)
+
+
+def warning(msg: Any, *args, **kwargs):
+    """警告日志"""
+    logger.warning(msg, *args, **kwargs)
+
+
+def error(msg: Any, *args, **kwargs):
+    """错误日志"""
+    logger.error(msg, *args, **kwargs)
diff --git a/backend/app/memory/__init__.py b/backend/app/memory/__init__.py
new file mode 100644
index 0000000..ba1f389
--- /dev/null
+++ b/backend/app/memory/__init__.py
@@ -0,0 +1,7 @@
+"""
+Mem0 记忆层模块
+"""
+
+from .mem0_client import Mem0Client
+
+__all__ = ["Mem0Client"]
diff --git a/backend/app/memory/mem0_client.py b/backend/app/memory/mem0_client.py
new file mode 100644
index 0000000..4a54ec2
--- /dev/null
+++ b/backend/app/memory/mem0_client.py
@@ -0,0 +1,146 @@
+from ..config import LLM_API_KEY
+from ..config import VLLM_BASE_URL
+import time
+"""
+Mem0 记忆层客户端封装模块
+负责 Mem0 的初始化、检索和存储
+"""
+
+import asyncio
+from typing import Optional, List, Dict
+from mem0 import AsyncMemory
+
+from ..config import (
+    QDRANT_URL,QDRANT_COLLECTION_NAME,QDRANT_API_KEY,
+    VLLM_BASE_URL, LLM_API_KEY,
+    LLAMACPP_EMBEDDING_URL, LLAMACPP_API_KEY
+)
+from ..logger import info, warning, error
+
+class Mem0Client:
+    """Mem0 异步客户端封装类"""
+
+    def __init__(self, llm_instance):
+        """
+        初始化 Mem0 客户端
+
+        Args:
+            llm_instance: LangChain LLM 实例（用于事实提取）
+        """
+        self.llm = llm_instance
+        self.mem0: Optional[AsyncMemory] = None
+        self._initialized = False
+
+    async def initialize(self):
+        """异步初始化 Mem0 客户端，并进行实际连接测试"""
+        if self._initialized:
+            return
+
+        try:
+            # Mem0 配置
+            config = {
+                "vector_store": {
+                    "provider": "qdrant",
+                    "config": {
+                        "url": QDRANT_URL,                     # 直接使用完整 URL
+                        "api_key": QDRANT_API_KEY,
+                        "collection_name": QDRANT_COLLECTION_NAME,
+                        "embedding_model_dims": 1024,
+                    }
+                },
+                "llm": {
+                    "provider": "openai",
+                    "config": {
+                        "model": "LLM_MODEL",
+                        "api_key": LLM_API_KEY,
+                        "openai_base_url": VLLM_BASE_URL,  
+                        "temperature": 0.1,
+                        "max_tokens": 2000,
+                    }
+                },
+                "embedder": {
+                    "provider": "openai",
+                    "config": {
+                        "model": "Qwen3-Embedding-0.6B-Q8_0",
+                        "api_key": LLAMACPP_API_KEY,
+                        "openai_base_url": LLAMACPP_EMBEDDING_URL,
+                    },
+                },
+                "version": "v1.1"
+            }
+
+            self.mem0 = AsyncMemory.from_config(config)
+            info("✅ Mem0 配置加载成功，开始连接测试...")
+
+            # 实际连接测试：调用一次 search 确保 Qdrant 和 Embedding 都可达
+            await asyncio.wait_for(
+                self.mem0.search("ping", user_id="test", limit=1),
+                timeout=60.0
+            )
+            info("✅ Mem0 实际连接测试成功，初始化完成")
+            self._initialized = True
+
+        except asyncio.TimeoutError:
+            error("❌ Mem0 连接测试超时 (10s)，请检查 Qdrant 或 Embedding 服务响应")
+            self.mem0 = None
+            self._initialized = False
+        except Exception as e:
+            error(f"❌ Mem0 初始化或连接测试失败: {e}")
+            import traceback
+            error(f"详细错误信息:\n{traceback.format_exc()}")
+            self.mem0 = None
+            self._initialized = False
+
+    async def search_memories(self, query: str, user_id: str, limit: int = 5) -> List[str]:
+        """
+        检索相关记忆
+
+        Args:
+            query: 查询文本
+            user_id: 用户 ID
+            limit: 返回结果数量限制
+
+        Returns:
+            List[str]: 记忆事实列表
+        """
+        if not self.mem0:
+            warning("⚠️ Mem0 未初始化，跳过记忆检索")
+            return []
+
+        try:
+            memories = await asyncio.wait_for(
+                self.mem0.search(query, user_id=user_id, limit=limit),
+                timeout=30.0
+            )
+
+            if memories and "results" in memories:
+                facts = [m["memory"] for m in memories["results"] if m.get("memory")]
+                if facts:
+                    info(f"🔍 [记忆检索] Mem0 返回 {len(facts)} 条记忆")
+                    return facts
+
+            info("🔍 [记忆检索] 未找到相关记忆")
+            return []
+
+        except asyncio.TimeoutError:
+            warning("⚠️ Mem0 检索超时 (30s)，跳过本次记忆检索")
+            return []
+        except Exception as e:
+            warning(f"⚠️ Mem0 检索失败: {e}")
+            return []
+
+    async def add_memories(self, messages, user_id):
+      if not self.mem0:
+        return False
+      try:
+        start = time.time()
+        info(f"📝 开始 Mem0 add，消息数: {len(messages)}")
+        await asyncio.wait_for(
+            self.mem0.add(messages, user_id=user_id, metadata={"type": "conversation"}),
+            timeout=60.0
+        )
+        info(f"✅ Mem0 add 完成，耗时: {time.time() - start:.2f}s")
+        return True
+      except asyncio.TimeoutError:
+        error(f"❌ Mem0 记忆添加超时 (60s)，已等待 {time.time() - start:.2f}s")
+        return False
\ No newline at end of file
diff --git a/backend/app/nodes/__init__.py b/backend/app/nodes/__init__.py
new file mode 100644
index 0000000..345883a
--- /dev/null
+++ b/backend/app/nodes/__init__.py
@@ -0,0 +1,19 @@
+"""
+节点模块 - 导出所有 LangGraph 节点函数
+"""
+
+from .router import should_continue
+from .llm_call import create_llm_call_node
+from .tool_call import create_tool_call_node
+from ..graph.retrieve_memory import create_retrieve_memory_node
+from .summarize import create_summarize_node
+from .finalize import finalize_node
+
+__all__ = [
+    "should_continue",
+    "create_llm_call_node",
+    "create_tool_call_node",
+    "create_retrieve_memory_node",
+    "create_summarize_node",
+    "finalize_node",
+]
diff --git a/backend/app/nodes/finalize.py b/backend/app/nodes/finalize.py
new file mode 100644
index 0000000..2203a0b
--- /dev/null
+++ b/backend/app/nodes/finalize.py
@@ -0,0 +1,45 @@
+"""
+完成事件节点模块
+负责发送完成事件，包含token使用情况和耗时信息
+"""
+
+from typing import Any, Dict
+from langgraph.config import get_stream_writer
+
+# 本地模块
+from ..graph.state import MessagesState
+from ..utils.logging import log_state_change
+from ..logger import info, error
+
+from langchain_core.runnables.config import RunnableConfig
+
+async def finalize_node(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+    """
+    完成事件节点 - 发送完成事件，包含token使用情况和耗时信息
+    
+    Args:
+        state: 当前对话状态
+        config: 运行时配置
+        
+    Returns:
+        空字典（完成节点，无状态更新）
+    """
+    log_state_change("finalize", state, "进入")
+
+    try:
+        # 获取流式写入器并发送完成事件
+        writer = get_stream_writer()
+        writer({
+            "type": "custom", 
+            "data": {
+                "type": "done",
+                "token_usage": state.get("last_token_usage", {}),
+                "elapsed_time": state.get("last_elapsed_time", 0.0)
+            }
+        })
+        info("🏁 [完成事件] 已发送完成事件，包含token使用情况和耗时信息")
+    except Exception as e:
+        error(f"❌ [完成事件] 发送完成事件时发生异常: {e}")
+    
+    log_state_change("finalize", state, "离开")
+    return {}
\ No newline at end of file
diff --git a/backend/app/nodes/llm_call.py b/backend/app/nodes/llm_call.py
new file mode 100644
index 0000000..f413d85
--- /dev/null
+++ b/backend/app/nodes/llm_call.py
@@ -0,0 +1,150 @@
+"""
+LLM 调用节点模块
+负责调用大语言模型并处理响应
+"""
+
+import time
+from typing import Any, Dict
+from langchain_core.language_models import BaseLLM
+from langchain_core.messages import AIMessage
+
+# 本地模块
+from ..graph.state import MessagesState
+from ..agent.prompts import create_system_prompt
+from ..utils.logging import log_state_change
+from ..logger import debug, info, error
+
+def create_llm_call_node(llm: BaseLLM, tools: list):
+    """
+    工厂函数：创建 LLM 调用节点
+    
+    Args:
+        llm: LangChain LLM 实例
+        tools: 工具列表
+        
+    Returns:
+        异步节点函数
+    """
+    # 构建调用链
+    prompt = create_system_prompt(tools)
+    llm_with_tools = llm.bind_tools(tools)
+    
+    # 恢复带 RunnableLambda 的链，并在下方使用 astream 遍历
+    chain = prompt  | llm_with_tools
+    
+    from langchain_core.runnables.config import RunnableConfig
+    
+    async def call_llm(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+        """
+        LLM 调用节点（异步方法）
+        
+        Args:
+            state: 当前对话状态
+            config: LangChain/LangGraph 自动注入的配置，包含 callbacks 等信息
+            
+        Returns:
+            更新后的状态字典
+        """
+        log_state_change("llm_call", state, "进入")
+        
+        memory_context = state.get("memory_context", "暂无用户信息")
+        start_time = time.time()
+        
+        try:
+            # 恢复为：手动进行 astream，并将所有的 chunk 拼接成最终的 response 返回。
+            # LangGraph 会自动监听这期间产生的所有 token。
+            chunks = []
+            async for chunk in chain.astream(
+                {
+                    "messages": state["messages"],
+                    "memory_context": memory_context
+                },
+                config=config
+            ):
+                chunks.append(chunk)
+
+            # 将所有 chunk 合并成最终的 AIMessage
+            if chunks:
+                response = chunks[0]
+                for chunk in chunks[1:]:
+                    response = response + chunk
+            else:
+                response = AIMessage(content="")
+            
+            elapsed_time = time.time() - start_time
+            
+            # 提取 token 用量（兼容不同 LLM 提供商的元数据格式）
+            token_usage = {}
+            input_tokens = 0
+            output_tokens = 0
+            
+            # 尝试从 response_metadata 中提取
+            if hasattr(response, 'response_metadata') and response.response_metadata:
+                meta = response.response_metadata
+                if 'token_usage' in meta:
+                    token_usage = meta['token_usage']
+                elif 'usage' in meta:
+                    token_usage = meta['usage']
+            
+            # 尝试从 additional_kwargs 中提取
+            if not token_usage and hasattr(response, 'additional_kwargs'):
+                add_kwargs = response.additional_kwargs
+                if 'llm_output' in add_kwargs and 'token_usage' in add_kwargs['llm_output']:
+                    token_usage = add_kwargs['llm_output']['token_usage']
+            
+            # 提取具体的 token 数值
+            if token_usage:
+                input_tokens = token_usage.get('prompt_tokens', token_usage.get('input_tokens', 0))
+                output_tokens = token_usage.get('completion_tokens', token_usage.get('output_tokens', 0))
+           
+            # 打印 LLM 的完整输出
+            debug("\n" + "="*80)
+            debug("📥 [LLM输出] 大模型返回的完整响应:")
+            debug(f"   消息类型: {response.type.upper()}")
+            debug(f"   内容长度: {len(str(response.content))} 字符")
+            debug("-"*80)
+            debug(f"{response.content}")
+
+            # 打印响应统计信息
+            info(f"⏱️  [LLM统计] 调用耗时: {elapsed_time:.2f}秒")
+            info(f"📊 [LLM统计] Token用量: 输入={input_tokens}, 输出={output_tokens}, 总计={input_tokens + output_tokens}")
+            if token_usage:
+                debug(f"📋 [LLM统计] 详细用量: {token_usage}")
+            debug("="*80 + "\n")
+            
+            result = {
+                "messages": [response],
+                "llm_calls": state.get('llm_calls', 0) + 1,
+                "last_token_usage": token_usage,
+                "last_elapsed_time": elapsed_time,
+                "turns_since_last_summary": state.get('turns_since_last_summary', 0) + 1  # 递增计数器
+            }
+            
+            log_state_change("llm_call", {**state, **result}, "离开")
+            return result
+            
+        except Exception as e:
+            elapsed_time = time.time() - start_time
+            error(f"\n❌ [LLM错误] 调用失败 (耗时: {elapsed_time:.2f}秒)")
+            error(f"   错误类型: {type(e).__name__}")
+            error(f"   错误信息: {str(e)}")
+            import traceback
+            traceback.print_exc()
+            debug("="*80 + "\n")
+            
+            # 返回一个友好的错误消息
+            error_response = AIMessage(
+                content="抱歉，模型暂时无法响应，可能是网络超时或服务繁忙，请稍后再试。"
+            )
+            error_result = {
+                "messages": [error_response],
+                "llm_calls": state.get('llm_calls', 0),
+                "last_token_usage": {},
+                "last_elapsed_time": elapsed_time,
+                "turns_since_last_summary": state.get('turns_since_last_summary', 0) + 1  # 即使出错也递增计数器
+            }
+            
+            log_state_change("llm_call", state, "离开（异常）")
+            return error_result
+    
+    return call_llm
diff --git a/backend/app/nodes/memory_trigger.py b/backend/app/nodes/memory_trigger.py
new file mode 100644
index 0000000..be5a65c
--- /dev/null
+++ b/backend/app/nodes/memory_trigger.py
@@ -0,0 +1,38 @@
+from typing import Any, Dict
+from langchain_core.runnables.config import RunnableConfig
+from ..graph.state import MessagesState
+from ..memory.mem0_client import Mem0Client
+from ..logger import info
+
+# 全局变量，在 GraphBuilder 中注入
+_mem0_client: Mem0Client = None
+
+def set_mem0_client(client: Mem0Client):
+    global _mem0_client
+    _mem0_client = client
+
+async def memory_trigger_node(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+    """检测用户消息中的记忆指令，若命中则主动调用 Mem0 存储"""
+    if _mem0_client is None:
+        return {}
+
+    messages = state.get("messages", [])
+    if not messages:
+        return {}
+
+    last_msg = messages[-1]
+    content = last_msg.content if hasattr(last_msg, 'content') else str(last_msg)
+
+    # 触发词（可自行扩展）
+    trigger_words = ["记住", "记下", "保存", "备忘", "记录下", "别忘了"]
+    if any(word in content for word in trigger_words):
+        user_id = config.get("metadata", {}).get("user_id", "default_user")
+        # 确保 Mem0 已初始化
+        if not _mem0_client._initialized:
+            await _mem0_client.initialize()
+        # 将用户消息作为事实来源提交给 Mem0
+        info(f"📌 检测到记忆指令，已主动触发 Mem0 存储")
+        mem0_messages = [{"role": "user", "content": content}]
+        await _mem0_client.add_memories(mem0_messages, user_id=user_id)
+
+    return {}  # 不修改状态
\ No newline at end of file
diff --git a/backend/app/nodes/router.py b/backend/app/nodes/router.py
new file mode 100644
index 0000000..6fab51a
--- /dev/null
+++ b/backend/app/nodes/router.py
@@ -0,0 +1,48 @@
+"""
+路由决策节点
+根据当前状态决定下一步走向
+"""
+
+from typing import Literal
+from langchain_core.messages import AIMessage
+
+# 本地模块
+from ..config import ENABLE_GRAPH_TRACE, MEMORY_SUMMARIZE_INTERVAL
+from ..graph.state import MessagesState
+from ..logger import info
+
+
+def should_continue(state: MessagesState) -> Literal['tool_node', 'summarize', 'finalize']:
+    """
+    决定下一步：工具调用、生成摘要还是结束
+    
+    Args:
+        state: 当前对话状态
+        
+    Returns:
+        下一个节点名称
+    """
+    last_message = state["messages"][-1]
+
+    # 1. 如果需要调用工具，优先进入工具节点
+    if isinstance(last_message, AIMessage) and last_message.tool_calls:
+        if ENABLE_GRAPH_TRACE:
+            info(f"🔀 [路由决策] 检测到 {len(last_message.tool_calls)} 个工具调用 → 转向 'tool_node'")
+        return 'tool_node'
+
+    # 2. 如果是 AI 的最终回复，判断是否达到摘要生成阈值
+    if isinstance(last_message, AIMessage):
+        turns = state.get("turns_since_last_summary", 0)
+        if turns >= MEMORY_SUMMARIZE_INTERVAL:
+            if ENABLE_GRAPH_TRACE:
+                info(f"🔀 [路由决策] 收到 AI 最终回复，已达摘要阈值({turns}/{MEMORY_SUMMARIZE_INTERVAL}) → 转向 'summarize'")
+            return 'summarize'
+        else:
+            if ENABLE_GRAPH_TRACE:
+                info(f"🔀 [路由决策] 收到 AI 最终回复，未达摘要阈值({turns}/{MEMORY_SUMMARIZE_INTERVAL}) → 结束流程")
+            return 'finalize'
+
+    # 3. 其他情况（如只有用户消息）直接结束
+    if ENABLE_GRAPH_TRACE:
+        info(f"🔀 [路由决策] 非 AI 消息（如纯用户消息） → 结束流程")
+    return 'finalize'
diff --git a/backend/app/nodes/summarize.py b/backend/app/nodes/summarize.py
new file mode 100644
index 0000000..2c9856b
--- /dev/null
+++ b/backend/app/nodes/summarize.py
@@ -0,0 +1,87 @@
+"""
+记忆存储节点模块
+负责将对话历史提交给 Mem0 进行事实提取和存储
+"""
+
+from typing import Any, Dict
+
+# 本地模块
+from ..graph.state import MessagesState
+from ..memory.mem0_client import Mem0Client
+from ..utils.logging import log_state_change
+from ..logger import debug, info, error, warning
+
+def create_summarize_node(mem0_client: Mem0Client):
+    """
+    工厂函数：创建记忆存储节点
+    
+    Args:
+        mem0_client: Mem0 客户端实例
+        
+    Returns:
+        异步节点函数
+    """
+    
+    from langchain_core.runnables.config import RunnableConfig
+    
+    async def summarize_conversation(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+        """
+        记忆存储节点 - 使用 Mem0
+        
+        Args:
+            state: 当前对话状态
+            config: 运行时配置
+            
+        Returns:
+            重置计数器的状态更新
+        """
+        log_state_change("summarize", state, "进入")
+
+        messages = state["messages"]
+        if len(messages) < 4:
+            debug("📝 [记忆添加] 对话过短，跳过")
+            return {"turns_since_last_summary": 0}
+
+        # 从 metadata 中获取 user_id
+        user_id = config.get("metadata", {}).get("user_id", "default_user")
+
+        # 确保 Mem0 已初始化（懒加载）
+        if not mem0_client._initialized:
+            await mem0_client.initialize()
+
+        # 将整个对话历史转换为 Mem0 需要的消息格式
+        mem0_messages = []
+        for msg in messages:
+            # 兼容 dict 和对象两种格式
+            if isinstance(msg, dict):
+                msg_type = msg.get("type", "")
+                msg_content = msg.get("content", "")
+            else:
+                msg_type = getattr(msg, 'type', '')
+                msg_content = getattr(msg, 'content', '')
+            
+            if msg_type == "human":
+                mem0_messages.append({"role": "user", "content": msg_content})
+            elif msg_type == "ai":
+                mem0_messages.append({"role": "assistant", "content": msg_content})
+            elif msg_type == "tool":
+                mem0_messages.append({"role": "system", "content": f"[工具返回] {msg_content}"})
+
+        if mem0_client.mem0:
+            try:
+                # 异步调用 Mem0 自动提取并存储事实
+                success = await mem0_client.add_memories(
+                    mem0_messages,
+                    user_id=user_id
+                )
+                if success:
+                    info(f"📝 [记忆添加] 已提交给 Mem0 进行事实提取")
+            except Exception as e:
+                error(f"❌ Mem0 记忆添加失败: {e}")
+        else:
+            warning("⚠️ Mem0 未初始化，跳过记忆添加")
+
+        log_state_change("summarize", state, "离开")
+        return {"turns_since_last_summary": 0}
+    
+    return summarize_conversation
\ No newline at end of file
diff --git a/backend/app/nodes/tool_call.py b/backend/app/nodes/tool_call.py
new file mode 100644
index 0000000..69ad2e6
--- /dev/null
+++ b/backend/app/nodes/tool_call.py
@@ -0,0 +1,101 @@
+"""
+工具执行节点模块
+负责执行 AI 调用的工具函数
+"""
+
+import asyncio
+from typing import Any, Dict
+from langchain_core.messages import AIMessage, ToolMessage
+from langgraph.config import get_stream_writer
+
+# 本地模块
+from ..graph.state import MessagesState
+from ..utils.logging import log_state_change
+from ..logger import debug, info
+
+def create_tool_call_node(tools_by_name: Dict[str, Any]):
+    """
+    工厂函数：创建工具执行节点
+    
+    Args:
+        tools_by_name: 名称到工具函数的映射字典
+        
+    Returns:
+        异步节点函数
+    """
+    
+    from langchain_core.runnables.config import RunnableConfig
+    
+    async def call_tools(state: MessagesState, config: RunnableConfig) -> Dict[str, Any]:
+        """
+        工具执行节点（异步方法）
+        
+        Args:
+            state: 当前对话状态
+            config: 运行时配置
+            
+        Returns:
+            包含 ToolMessage 的状态更新
+        """
+        log_state_change("tool_node", state, "进入")
+        
+        last_message = state['messages'][-1]
+        if not isinstance(last_message, AIMessage) or not last_message.tool_calls:
+            log_state_change("tool_node", state, "离开（无工具调用）")
+            return {"messages": []}
+
+        results = []
+        loop = asyncio.get_event_loop()
+        
+        info(f"🛠️  [工具调用] 准备执行 {len(last_message.tool_calls)} 个工具")
+        
+        for tool_call in last_message.tool_calls:
+            tool_name = tool_call["name"]
+            tool_args = tool_call["args"]
+            tool_id = tool_call["id"]
+            tool_func = tools_by_name.get(tool_name)
+
+            debug(f"   ├─ 调用工具: {tool_name} 参数: {tool_args}")
+
+            if tool_func is None:
+                err_msg = f"Tool {tool_name} not found"
+                debug(f"   └─ ❌ {err_msg}")
+                results.append(ToolMessage(content=err_msg, tool_call_id=tool_id))
+                continue
+
+            # 获取流式写入器并发送工具调用开始事件
+            writer = get_stream_writer()
+            writer({"type": "custom", "data": {"type": "tool_start", "tool": tool_name}})
+            
+            try:
+                # 修复闭包问题：将变量作为默认参数传入 lambda
+                # 如果工具支持异步 (ainvoke)，优先使用异步调用
+                if hasattr(tool_func, 'ainvoke'):
+                    observation = await tool_func.ainvoke(tool_args)
+                else:
+                    observation = await loop.run_in_executor(
+                        None,
+                        lambda args=tool_args: tool_func.invoke(args)  # 默认参数捕获当前值
+                    )
+                
+                # 字符打印
+                result_preview = str(observation).replace("\n", " ")
+                debug(f"   └─ ✅ 结果: {result_preview}")
+                results.append(ToolMessage(content=str(observation), tool_call_id=tool_id))
+                
+                # 发送工具调用完成事件
+                writer({"type": "custom", "data": {"type": "tool_end", "tool": tool_name, "success": True}})
+            except Exception as e:
+                debug(f"   └─ ❌ 异常: {e}")
+                results.append(ToolMessage(content=f"Error: {e}", tool_call_id=tool_id))
+                
+                # 发送工具调用失败事件
+                writer({"type": "custom", "data": {"type": "tool_end", "tool": tool_name, "success": False, "error": str(e)}})
+
+        info(f"🛠️  [工具调用] 执行完成，返回 {len(results)} 条 ToolMessage")
+        
+        result = {"messages": results}
+        log_state_change("tool_node", {**state, **result}, "离开")
+        return result
+    
+    return call_tools
\ No newline at end of file
diff --git a/backend/app/rag/README.md b/backend/app/rag/README.md
new file mode 100644
index 0000000..a91d8f6
--- /dev/null
+++ b/backend/app/rag/README.md
@@ -0,0 +1,391 @@
+# 在线 RAG 检索与生成系统 (Online RAG Retriever)
+
+该模块负责 RAG 系统的阶段二：**在线检索与生成**。它接收用户提问，从知识库中检索出上下文，利用各种高级策略去噪、融合，并作为增强上下文输入给大语言模型 (LLM)。
+
+## 🎯 核心架构
+
+### 技术栈
+
+| 组件 | 技术选型 | 版本 | 说明 |
+|:-----|:---------|:-----|:-----|
+| **基础检索** | `Qdrant` | 1.17+ | HNSW 稠密向量检索 |
+| **混合检索** | `Qdrant` + `BM25` | 内置 | 稠密 + 稀疏向量融合 |
+| **查询改写** | `LangChain` | 内置 | `MultiQueryGenerator` 多路改写 |
+| **RRF 融合** | 自实现 | - | `reciprocal_rank_fusion` 倒数排名融合 |
+| **重排序** | `llama.cpp` | 本地服务 | OpenAI 兼容 Rerank API |
+| **编排框架** | `asyncio` | Python 3.10+ | 异步并行检索 |
+
+### 检索流水线
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                      用户提问                                │
+└──────────────────────┬──────────────────────────────────────┘
+                       │
+                       ▼
+┌─────────────────────────────────────────────────────────────┐
+│              MultiQueryGenerator                             │
+│           多路查询改写 (num_queries=3)                        │
+│   "如何申请项目资金？" → ["项目资金申请流程", "经费申请步骤"]  │
+└──────────────────────┬──────────────────────────────────────┘
+                       │
+                       ▼
+┌─────────────────────────────────────────────────────────────┐
+│              并行检索 (asyncio.gather)                        │
+│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐       │
+│  │  查询1 检索   │  │  查询2 检索   │  │  查询3 检索   │       │
+│  │  (k=20)      │  │  (k=20)      │  │  (k=20)      │       │
+│  └──────────────┘  └──────────────┘  └──────────────┘       │
+└──────────────────────┬──────────────────────────────────────┘
+                       │
+                       ▼
+┌─────────────────────────────────────────────────────────────┐
+│           reciprocal_rank_fusion (RRF)                       │
+│   RRF_score(d) = Σ 1/(k + rank_q(d))  (k=60)                │
+│   融合多路检索结果，去重排序                                  │
+└──────────────────────┬──────────────────────────────────────┘
+                       │
+                       ▼
+┌─────────────────────────────────────────────────────────────┐
+│              LLaMaCPPReranker                                │
+│   远程重排序 (bge-reranker-v2-m3)                             │
+│   返回 Top-N (top_n=5) 最相关文档                             │
+└──────────────────────┬──────────────────────────────────────┘
+                       │
+                       ▼
+┌─────────────────────────────────────────────────────────────┐
+│              返回增强上下文                                  │
+│   format_context() → 格式化输出                               │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### 技术特性
+
+- ✅ **多路查询改写**：通过 LLM 将单一问题改写为多个不同角度的查询
+- ✅ **RRF 融合算法**：Reciprocal Rank Fusion，无需评分归一化的融合算法
+- ✅ **远程重排序**：使用 llama.cpp 服务的 OpenAI 兼容 Rerank API
+- ✅ **混合检索支持**：稠密向量 + BM25 稀疏向量混合检索
+- ✅ **异步并行检索**：多路查询并行执行，提升检索速度
+- ✅ **优雅降级**：重排序器不可用时自动降级到基础融合结果
+
+## 📂 架构与文件结构
+
+```
+app/rag/
+├── __init__.py
+├── retriever.py          # Qdrant 基础检索与混合检索
+├── reranker.py           # llama.cpp 远程重排序器
+├── query_transform.py    # 多路查询改写生成器
+├── fusion.py             # RRF 倒数排名融合算法
+├── pipeline.py           # RAG 流水线编排
+└── tools.py              # LangChain Tool 封装
+```
+
+## 🎯 演进路线与算法详解 (Roadmap)
+
+### Level 1: 基础向量搜索 (Basic Similarity Search)
+
+- **核心算法**: 近似最近邻搜索 (ANN, 常用 HNSW 算法)。将用户问题转化为向量后，计算它与库中向量的余弦相似度 (Cosine Similarity)，取距离最近的 K 个块。
+- **优缺点**: 速度极快。但只能捕捉"语义相似"，如果用户搜索特定专有名词、编号、订单号，纯向量检索往往会失效（产生"幻觉"匹配）。
+- **实现指南**:
+  - 使用 `rag_indexer.embedders.LlamaCppEmbedder` 作为嵌入模型
+  - 使用 `app/rag/retriever.py` 中的 `create_base_retriever` 创建基础检索器
+  - 配置 `search_kwargs={"k": 20}` 进行初步召回
+
+```python
+from app.rag.retriever import create_base_retriever
+
+retriever = create_base_retriever(
+    collection_name="rag_documents",
+    embeddings=embeddings,
+    search_kwargs={"k": 20}
+)
+docs = retriever.invoke("什么是 RAG？")
+```
+
+### Level 2: 混合检索与重排序 (Hybrid Search + Reranker)
+
+混合检索旨在结合向量的"语义泛化"与关键词的"精准匹配"，随后利用重排序模型过滤噪声。
+
+**1. 基础召回 (混合检索)**
+
+- **核心原理**: 结合基于 HNSW 的 Dense Vector 相似度搜索与基于 TF-IDF 的 BM25 稀疏检索 (Sparse Vector)。
+- **实现指南**: 使用 `app/rag/retriever.py` 中的 `create_hybrid_retriever` 函数，配置 `dense_k=10` 和 `sparse_k=10`，总召回 20 条结果。
+
+```python
+from app.rag.retriever import create_hybrid_retriever
+
+retriever = create_hybrid_retriever(
+    collection_name="rag_documents",
+    embeddings=embeddings,
+    dense_k=10,
+    sparse_k=10,
+    score_threshold=0.3
+)
+```
+
+**2. 二次精排 (Cross-Encoder)**
+
+- **核心原理**: 不同于双塔模型（分别算向量再求距离），交叉编码器将"用户问题 + 检索到的单例文档"拼接后整体输入 Transformer 模型，由模型直接输出 0~1 的相关性得分，精度极高。
+- **实现指南**:
+  - 使用 `app/rag/reranker.py` 中的 `LLaMaCPPReranker` 类，加载 `bge-reranker-v2-m3` 模型
+  - 设置 `top_n=5` 保留最相关的 5 条结果
+
+```python
+from app.rag.reranker import LLaMaCPPReranker
+
+reranker = LLaMaCPPReranker(
+    base_url="http://127.0.0.1:8083",
+    api_key="your-api-key",
+    top_n=5
+)
+sorted_docs = reranker.compress_documents(documents, query)
+```
+
+### Level 3: RAG-Fusion (多路改写与倒数排名融合)
+
+RAG-Fusion 通过大模型发散思维，将单一问题改写为多个相似问题，扩大搜索面，再利用数学统计算法合并结果。
+
+**1. 多路查询改写**
+
+- **核心原理**: 克服用户初始提问词不达意或视角受限的问题。
+- **实现指南**: 使用 `app/rag/query_transform.py` 中的 `MultiQueryGenerator` 类，配置 `num_queries=3` 生成 3 个不同角度的查询。
+
+```python
+from app.rag.query_transform import MultiQueryGenerator
+
+generator = MultiQueryGenerator(llm=llm, num_queries=3)
+queries = await generator.agenerate("如何申请项目资金？")
+# 返回：["如何申请项目资金？", "项目资金申请流程是什么？", "申请项目经费需要哪些步骤？"]
+```
+
+**2. 倒数排名融合 (RRF)**
+
+- **核心原理**: RRF (Reciprocal Rank Fusion) 是一种无需评分归一化的融合算法。公式为 `RRF_score(d) = Σ 1/(k + rank_q(d))`，有效避免某一极端检索结果主导全局。
+- **实现指南**: 使用 `app/rag/fusion.py` 中的 `reciprocal_rank_fusion` 函数，配置 `k=60` 实现倒数排名融合。
+
+```python
+from app.rag.fusion import reciprocal_rank_fusion
+
+# 多个查询的检索结果
+doc_lists = [result1, result2, result3]
+fused_docs = reciprocal_rank_fusion(doc_lists, k=60)
+```
+
+### Level 4: Agentic RAG / Self-RAG (智能体与自我反思)
+
+- **核心原理**: 基于 LangGraph 的 ReAct (Reasoning and Acting) 状态机路由。大模型并非每次都去死板地执行检索，而是先判断问题："这是闲聊？还是需要查知识库？"。如果是后者，模型输出一个 `ToolCall` 指令，触发检索。
+- **实现指南**: 使用 `app/rag/tools.py` 中的 `search_knowledge_base` 工具，将其绑定到 LangGraph 状态机中。
+
+- **示意图**:
+
+```
+┌──────────┐     ┌──────────────┐     ┌──────────┐     ┌────────
+│   User   │────>│ LangGraph    │────>│ RAG_Tool │────>│ Qdrant │
+│          │     │ Agent        │     │          │     │        │
+│ "公司报  │     │ 思考: 这是   │     │ ToolCall │     │ RAG-   │
+│ 销流程？"│     │ 内部规章问题 │     │ search_  │     │ Fusion │
+│          │     │ 需要查资料   │     │ knowledge│     │ & 混合 │
+│          │<────│ 资料充分，   │<────│ 返回最相 │<────│ 检索   │
+│ "根据知  │     │ 开始撰写回答 │     │ 关5条规定 │     │ Cross- │
+│ 识库规定 │     │              │     │          │     │ Encoder│
+│ ..."     │     │              │     │          │     │ 重排   │
+└──────────     └──────────────     └──────────┘     └────────┘
+```
+
+### Level 5: GraphRAG 集成 (基于图和关系的 RAG)
+
+- **核心原理**: 结合知识图谱的结构化关系和向量检索的语义相似度，解决跨文档复杂关系推理问题。
+- **实现指南**:
+  - 使用 `langchain_community.graphs` 模块构建知识图谱
+  - 配置本地大模型（如 `Gemma-4-E4B`）用于实体关系抽取
+  - 实现混合检索逻辑，结合向量相似度和图路径分析
+
+```python
+from langchain_community.graphs import Neo4jGraph
+from langchain_experimental.graph_transformers import LLMGraphTransformer
+
+# 实体关系抽取
+transformer = LLMGraphTransformer(llm=local_llm)
+graph_documents = transformer.convert_to_graph_documents(documents)
+
+# 存储到图数据库
+graph = Neo4jGraph(url="bolt://localhost:7687")
+graph.add_graph_documents(graph_documents)
+```
+
+## 🔧 核心组件详解
+
+### 1. 检索器 (retriever.py)
+
+提供基于 Qdrant 的向量检索能力。
+
+**基础检索器**：
+```python
+from app.rag.retriever import create_base_retriever
+
+retriever = create_base_retriever(
+    collection_name="rag_documents",
+    embeddings=embeddings,
+    search_kwargs={"k": 20}
+)
+```
+
+**混合检索器**：
+```python
+from app.rag.retriever import create_hybrid_retriever
+
+retriever = create_hybrid_retriever(
+    collection_name="rag_documents",
+    embeddings=embeddings,
+    dense_k=10,
+    sparse_k=10,
+    score_threshold=0.3
+)
+```
+
+### 2. 多路查询改写 (query_transform.py)
+
+通过 LLM 将用户问题改写为多个不同版本，扩大搜索面。
+
+```python
+from app.rag.query_transform import MultiQueryGenerator
+
+generator = MultiQueryGenerator(llm=llm, num_queries=3)
+queries = await generator.agenerate("如何申请项目资金？")
+```
+
+### 3. RRF 融合算法 (fusion.py)
+
+Reciprocal Rank Fusion 算法，公式：`RRF_score(d) = Σ 1/(k + rank_q(d))`
+
+```python
+from app.rag.fusion import reciprocal_rank_fusion
+
+# 多个查询的检索结果
+doc_lists = [result1, result2, result3]
+fused_docs = reciprocal_rank_fusion(doc_lists, k=60)
+```
+
+### 4. 重排序器 (reranker.py)
+
+使用 llama.cpp 服务的 OpenAI 兼容 Rerank API 对检索结果重排序。
+
+```python
+from app.rag.reranker import LLaMaCPPReranker
+
+reranker = LLaMaCPPReranker(
+    base_url="http://127.0.0.1:8083",
+    api_key="your-api-key",
+    top_n=5
+)
+sorted_docs = reranker.compress_documents(documents, query)
+```
+
+### 5. RAG 流水线 (pipeline.py)
+
+组合上述组件的完整检索流水线。
+
+```python
+from app.rag.pipeline import RAGPipeline
+
+pipeline = RAGPipeline(
+    retriever=retriever,
+    llm=llm,
+    num_queries=3,
+    rerank_top_n=5,
+)
+
+# 异步检索
+docs = await pipeline.aretrieve("如何申请项目资金？")
+
+# 格式化上下文
+context = pipeline.format_context(docs)
+```
+
+## 🔄 与 Agent 系统集成
+
+### 封装为 LangChain Tool
+
+```python
+from langchain_core.tools import tool
+from app.rag.pipeline import RAGPipeline
+
+@tool
+def search_knowledge_base(query: str) -> str:
+    """搜索知识库获取相关信息"""
+    docs = pipeline.retrieve(query)
+    return pipeline.format_context(docs)
+```
+
+### 绑定到 LangGraph
+
+```python
+from app.graph.graph_builder import GraphBuilder
+
+# 将 RAG 工具添加到工具列表
+tools = AVAILABLE_TOOLS + [search_knowledge_base]
+
+# 构建图
+builder = GraphBuilder(llm, tools, tools_by_name)
+graph = builder.build().compile(checkpointer=checkpointer)
+```
+
+## ⚙️ 环境配置
+
+| 变量名 | 说明 | 默认值 |
+|:-------|:-----|:-------|
+| `QDRANT_URL` | Qdrant 向量数据库地址 | `http://127.0.0.1:6333` |
+| `QDRANT_API_KEY` | Qdrant API 密钥 | - |
+| `LLAMACPP_RERANKER_URL` | llama.cpp 重排序服务地址 | `http://127.0.0.1:8083` |
+| `LLAMACPP_API_KEY` | llama.cpp API 密钥 | - |
+
+## 🚀 快速开始
+
+```python
+# 1. 初始化嵌入模型
+from rag_core.embedders import LlamaCppEmbedder
+embedder = LlamaCppEmbedder()
+embeddings = embedder.as_langchain_embeddings()
+
+# 2. 创建检索器
+from app.rag.retriever import create_base_retriever
+retriever = create_base_retriever(
+    collection_name="rag_documents",
+    embeddings=embeddings,
+    search_kwargs={"k": 20}
+)
+
+# 3. 创建 RAG 流水线
+from app.rag.pipeline import RAGPipeline
+pipeline = RAGPipeline(
+    retriever=retriever,
+    llm=llm,
+    num_queries=3,
+    rerank_top_n=5,
+)
+
+# 4. 执行检索
+docs = pipeline.retrieve("如何申请项目资金？")
+
+# 5. 格式化上下文
+context = pipeline.format_context(docs)
+print(context)
+```
+
+## 📊 检索策略对比
+
+| 策略 | 优点 | 缺点 | 适用场景 |
+|:-----|:-----|:-----|:---------|
+| **基础向量检索** | 速度快，语义理解好 | 专有名词匹配差 | 通用问答 |
+| **混合检索** | 语义 + 关键词匹配 | 需要配置稀疏向量 | 专业术语查询 |
+| **多路改写 + RRF** | 搜索面广，结果稳定 | 延迟略高 | 复杂问题 |
+| **重排序** | 精度高 | 依赖额外模型 | 最终精排 |
+
+## 🤝 与 rag_indexer 集成
+
+- **向量存储**：共享 Qdrant 集合，确保嵌入模型一致
+- **文档存储**：使用 PostgreSQL 存储父块，通过 UUID 映射
+- **集合名称**：默认使用 `rag_documents` 集合
+
+详见 [rag_indexer/README.md](../../rag_indexer/README.md)
diff --git a/backend/app/rag/__init__.py b/backend/app/rag/__init__.py
new file mode 100644
index 0000000..4f86e0b
--- /dev/null
+++ b/backend/app/rag/__init__.py
@@ -0,0 +1,69 @@
+"""
+RAG 检索与生成模块
+
+提供在线检索与生成功能，包括：
+- 基础向量检索（稠密向量 / 混合检索）
+- 重排序（Cross-Encoder）
+- 多路查询改写（Multi-Query）
+- RRF 融合（Reciprocal Rank Fusion）
+- 完整的 RAG 流水线
+- Agent 工具封装
+
+固定流水线：
+    用户查询 → 多路改写 → 并行检索 → RRF 融合 → 重排序 → 返回父文档
+
+示例用法：
+    >>> from app.rag.rag import RAGPipeline, create_rag_tool
+    >>> from rag_indexer.builder import IndexBuilder, IndexBuilderConfig
+    >>> from langchain_openai import ChatOpenAI
+    >>>
+    >>> # 获取基础检索器（如父子块检索器）
+    >>> config = IndexBuilderConfig(collection_name="my_docs")
+    >>> builder = IndexBuilder(config)
+    >>> retriever = builder.retriever
+    >>>
+    >>> # 创建 LLM 和流水线
+    >>> llm = ChatOpenAI(model="gpt-3.5-turbo")
+    >>> pipeline = RAGPipeline(retriever=retriever, llm=llm)
+    >>>
+    >>> # 检索
+    >>> docs = await pipeline.aretrieve("什么是 RAG？")
+    >>> context = pipeline.format_context(docs)
+    >>>
+    >>> # 创建 Agent 工具
+    >>> rag_tool = create_rag_tool(retriever=retriever, llm=llm)
+"""
+
+from .retriever import (
+    create_base_retriever,
+    create_hybrid_retriever,
+    create_qdrant_client,
+)
+from .reranker import LLaMaCPPReranker
+from .query_transform import MultiQueryGenerator
+from .fusion import reciprocal_rank_fusion
+from .pipeline import RAGPipeline
+from .tools import create_rag_tool_sync
+
+
+__all__ = [
+    # 检索器工厂函数
+    "create_base_retriever",
+    "create_hybrid_retriever",
+    "create_qdrant_client",
+    
+    # 重排序器
+    "LLaMaCPPReranker",
+    
+    # 查询改写生成器
+    "MultiQueryGenerator",
+    
+    # 融合算法
+    "reciprocal_rank_fusion",
+    
+    # 主流水线
+    "RAGPipeline",
+    
+    # 工具创建（供 Agent 使用）
+    "create_rag_tool_sync",
+]
\ No newline at end of file
diff --git a/backend/app/rag/fusion.py b/backend/app/rag/fusion.py
new file mode 100644
index 0000000..ddf8f42
--- /dev/null
+++ b/backend/app/rag/fusion.py
@@ -0,0 +1,36 @@
+# rag/fusion.py
+
+from typing import List, Dict
+from langchain_core.documents import Document
+
+def reciprocal_rank_fusion(
+    doc_lists: List[List[Document]],
+    k: int = 60
+) -> List[Document]:
+    """
+    对多个检索结果列表进行 RRF 融合。
+    
+    Args:
+        doc_lists: 多个检索结果列表，每个列表来自一个查询
+        k: RRF 常数，通常设为 60
+    
+    Returns:
+        融合后按 RRF 得分降序排列的文档列表
+    """
+    # 使用文档内容作为唯一标识（如果内容相同但 metadata 不同，视为同一文档）
+    # 更好的做法是用 docstore 的 ID，这里简化处理：用内容 hash
+    doc_to_score: Dict[str, float] = {}
+    doc_map: Dict[str, Document] = {}
+    
+    for docs in doc_lists:
+        for rank, doc in enumerate(docs, start=1):
+            # 生成唯一标识符（内容+来源组合，避免不同文件相同内容混淆）
+            doc_id = f"{doc.page_content[:200]}_{doc.metadata.get('source', '')}"
+            if doc_id not in doc_map:
+                doc_map[doc_id] = doc
+            score = doc_to_score.get(doc_id, 0.0) + 1.0 / (k + rank)
+            doc_to_score[doc_id] = score
+    
+    # 按得分排序
+    sorted_ids = sorted(doc_to_score.keys(), key=lambda x: doc_to_score[x], reverse=True)
+    return [doc_map[doc_id] for doc_id in sorted_ids]
\ No newline at end of file
diff --git a/backend/app/rag/pipeline.py b/backend/app/rag/pipeline.py
new file mode 100644
index 0000000..41f4186
--- /dev/null
+++ b/backend/app/rag/pipeline.py
@@ -0,0 +1,91 @@
+# rag/pipeline.py
+
+import asyncio
+import os
+from ..config import LLAMACPP_RERANKER_URL, LLAMACPP_API_KEY
+from typing import List
+from langchain_core.documents import Document
+from langchain_core.language_models import BaseLanguageModel
+
+from .reranker import LLaMaCPPReranker
+from .query_transform import MultiQueryGenerator
+from .fusion import reciprocal_rank_fusion
+
+class RAGPipeline:
+    """
+    固定流程的 RAG 检索流水线：
+    多路改写 → 并行检索 → RRF融合 → 重排序 → 返回父文档
+    """
+    
+    def __init__(
+        self,
+        retriever,  # 基础检索器（应返回父文档，例如 ParentDocumentRetriever 实例）
+        llm: BaseLanguageModel,
+        num_queries: int = 3,
+        rerank_top_n: int = 5,
+    ):
+        """
+        Args:
+            retriever: 基础检索器对象，需实现 ainvoke(query) 异步方法
+            llm: 用于生成多路查询的语言模型
+            num_queries: 生成的查询变体数量
+            rerank_top_n: 最终返回的文档数量
+            rerank_model: 重排序模型名称
+        """
+        self.retriever = retriever
+        self.llm = llm
+        self.num_queries = num_queries
+        self.rerank_top_n = rerank_top_n
+        
+        # 初始化组件
+        self.query_generator = MultiQueryGenerator(llm=llm, num_queries=num_queries)
+        self.reranker = LLaMaCPPReranker(
+                base_url=LLAMACPP_RERANKER_URL,
+                api_key=LLAMACPP_API_KEY,
+                top_n=rerank_top_n,
+            )
+    
+    async def aretrieve(self, query: str) -> List[Document]:
+        """
+        异步执行完整检索流程
+        """
+        # Step 1: 生成多路查询
+        queries = await self.query_generator.agenerate(query)
+        # 包含原始查询，确保至少有一条
+        if query not in queries:
+            queries.insert(0, query)
+        else:
+            # 如果原始查询已在列表中，将其移至首位
+            queries.remove(query)
+            queries.insert(0, query)
+        
+        # Step 2: 并行检索（每个查询获取文档列表）
+        tasks = [self.retriever.ainvoke(q) for q in queries]
+        doc_lists = await asyncio.gather(*tasks)
+        
+        # Step 3: RRF 融合
+        fused_docs = reciprocal_rank_fusion(doc_lists)
+        
+        # Step 4: 重排序
+        try:
+            final_docs = self.reranker.compress_documents(fused_docs, query)
+        except Exception:
+            # 若重排序器不可用，直接返回融合后的前 N 条
+            final_docs = fused_docs[:self.rerank_top_n]
+        
+        return final_docs
+    
+    def retrieve(self, query: str) -> List[Document]:
+        """同步检索入口（内部调用异步方法）"""
+        return asyncio.run(self.aretrieve(query))
+    
+    def format_context(self, documents: List[Document]) -> str:
+        """将文档列表格式化为上下文字符串"""
+        if not documents:
+            return ""
+        
+        parts = []
+        for i, doc in enumerate(documents, 1):
+            source = doc.metadata.get("source", "未知来源")
+            parts.append(f"【资料 {i}】来源：{source}\n{doc.page_content}\n---\n")
+        return "\n".join(parts)
\ No newline at end of file
diff --git a/backend/app/rag/query_transform.py b/backend/app/rag/query_transform.py
new file mode 100644
index 0000000..38f9fd1
--- /dev/null
+++ b/backend/app/rag/query_transform.py
@@ -0,0 +1,43 @@
+# rag/query_transform.py
+
+from typing import List
+from langchain_core.language_models import BaseLanguageModel
+from langchain_core.prompts import PromptTemplate
+
+MULTI_QUERY_PROMPT = PromptTemplate.from_template(
+    """你是一个专业的查询改写助手。你的任务是将用户的问题改写成 {num_queries} 个不同的版本。
+这些版本应该从不同的角度、使用不同的关键词来表达相同或相关的意图。
+
+原始问题: {question}
+
+请生成 {num_queries} 个不同版本的查询，每个版本一行。
+确保每个版本都是独立、完整的查询语句。
+
+生成 {num_queries} 个查询:"""
+)
+
+class MultiQueryGenerator:
+    """多路查询生成器（不依赖 LangChain 的 MultiQueryRetriever）"""
+    
+    def __init__(self, llm: BaseLanguageModel, num_queries: int = 3):
+        self.llm = llm
+        self.num_queries = num_queries
+        self.prompt = MULTI_QUERY_PROMPT
+    
+    def generate(self, query: str) -> List[str]:
+        """同步生成多个查询变体"""
+        prompt_str = self.prompt.format(num_queries=self.num_queries, question=query)
+        response = self.llm.invoke(prompt_str)
+        # 处理响应内容，按行分割并去除空行和首尾空白
+        lines = response.content.strip().split('\n')
+        queries = [line.strip() for line in lines if line.strip()]
+        # 确保至少返回原始查询
+        return queries[:self.num_queries] if queries else [query]
+    
+    async def agenerate(self, query: str) -> List[str]:
+        """异步生成多个查询变体"""
+        prompt_str = self.prompt.format(num_queries=self.num_queries, question=query)
+        response = await self.llm.ainvoke(prompt_str)
+        lines = response.content.strip().split('\n')
+        queries = [line.strip() for line in lines if line.strip()]
+        return queries[:self.num_queries] if queries else [query]
\ No newline at end of file
diff --git a/backend/app/rag/reranker.py b/backend/app/rag/reranker.py
new file mode 100644
index 0000000..925e283
--- /dev/null
+++ b/backend/app/rag/reranker.py
@@ -0,0 +1,75 @@
+"""
+重排序器模块 (适配版)
+使用远程 llama.cpp 服务 (兼容 OpenAI Rerank API) 替代本地 Cross-Encoder
+"""
+import requests
+from typing import List
+from langchain_core.documents import Document
+
+class LLaMaCPPReranker:
+    """使用远程 llama.cpp 服务对检索结果重排序。"""
+
+    def __init__(self, 
+                 base_url: str, 
+                 api_key: str,
+                 top_n: int = 5,
+                 timeout: int = 60):
+        """
+        初始化远程重排序器
+        
+        Args:
+            base_url: llama.cpp 服务的地址和端口，默认为环境变量 LLAMACPP_RERANKER_URL 或 "http://127.0.0.1:8083"。
+            top_n: 返回前 N 个结果。
+            api_key: API 密钥，默认为环境变量 LLAMACPP_API_KEY 或 "huang1998"。
+            timeout: 请求超时时间（秒）。
+        """
+        self.base_url = base_url 
+        self.api_key = api_key 
+        self.top_n = top_n
+        self.timeout = timeout
+        self.endpoint = f"{self.base_url}/rerank"
+        
+    def compress_documents(
+        self, documents: List[Document], query: str
+    ) -> List[Document]:
+        """
+        对文档进行重排序
+        
+        Args:
+            documents: 待排序的文档列表
+            query: 查询字符串
+        
+        Returns:
+            排序后的文档列表
+        """
+        if not documents:
+            return []
+
+        # 准备请求体
+        # 根据 llama.cpp 的 OpenAI 兼容性，文档是一个字符串列表
+        payload = {
+            "model": "bge-reranker-v2-m3",
+            "query": query,
+            "documents": [doc.page_content for doc in documents],
+            "top_n": self.top_n
+        }
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {self.api_key}"
+        }
+
+        try:
+            response = requests.post(self.endpoint, json=payload, headers=headers, timeout=self.timeout)
+            response.raise_for_status()  # 检查请求是否成功
+            results = response.json()
+            
+            # 解析返回结果
+            # 返回格式: {"results": [{"index": 0, "document": "...", "relevance_score": 0.8}, ...]}
+            # 按相关性得分降序排列
+            sorted_indices = [item["index"] for item in results["results"]]
+            sorted_docs = [documents[idx] for idx in sorted_indices]
+            return sorted_docs
+            
+        except Exception as e:
+            print(f"警告: 远程重排序过程出错，将使用原始排序。错误: {e}")
+            return documents[:self.top_n]
\ No newline at end of file
diff --git a/backend/app/rag/retriever.py b/backend/app/rag/retriever.py
new file mode 100644
index 0000000..483c8b9
--- /dev/null
+++ b/backend/app/rag/retriever.py
@@ -0,0 +1,199 @@
+"""
+Qdrant 向量检索器模块
+
+提供基于 Qdrant 的基础向量检索和混合检索（Dense + Sparse）功能。
+
+核心原理：
+- 基础检索：将查询文本转换为向量，在 Qdrant 中进行近似最近邻（ANN）搜索，
+  使用余弦相似度返回最相似的 k 个文档。
+- 混合检索：结合稠密向量检索（语义相似）和 BM25 稀疏向量检索（关键词匹配），
+  通过加权或分数融合提高召回精度。
+
+使用示例：
+    >>> from rag_core import LlamaCppEmbedder
+    >>> embedder = LlamaCppEmbedder()
+    >>> embeddings = embedder.as_langchain_embeddings()
+    >>>
+    >>> # 创建基础检索器
+    >>> retriever = create_base_retriever(
+    ...     collection_name="my_docs",
+    ...     embeddings=embeddings,
+    ...     search_kwargs={"k": 10}
+    ... )
+    >>>
+    >>> # 执行检索
+    >>> docs = retriever.invoke("什么是 RAG？")
+"""
+
+from typing import Optional, Dict, Any
+from qdrant_client import QdrantClient
+from qdrant_client.http.exceptions import UnexpectedResponse
+from langchain_qdrant import QdrantVectorStore
+from langchain_core.embeddings import Embeddings
+from langchain_core.retrievers import BaseRetriever
+
+from rag_core import QDRANT_URL, QDRANT_API_KEY
+
+# 模块级常量
+DEFAULT_SEARCH_K = 20
+DEFAULT_SCORE_THRESHOLD = 0.3
+
+
+def create_qdrant_client(
+    url: Optional[str] = None,
+    api_key: Optional[str] = None,
+    timeout: int = 30,
+) -> QdrantClient:
+    """
+    创建并返回一个配置好的 Qdrant 客户端。
+
+    优先使用传入参数，若未提供则回退到环境变量 QDRANT_URL 和 QDRANT_API_KEY。
+
+    Args:
+        url: Qdrant 服务地址，例如 "http://localhost:6333"。
+             默认从环境变量 QDRANT_URL 读取。
+        api_key: API 密钥（若 Qdrant 启用了认证）。
+                 默认从环境变量 QDRANT_API_KEY 读取。
+        timeout: 请求超时时间（秒），默认 30 秒。
+
+    Returns:
+        配置好的 QdrantClient 实例。
+
+    Raises:
+        ValueError: 如果 url 为空且环境变量也未设置。
+    """
+    effective_url = url or QDRANT_URL
+    if not effective_url:
+        raise ValueError(
+            "Qdrant URL 未提供，请设置参数 url 或环境变量 QDRANT_URL"
+        )
+
+    effective_api_key = api_key or QDRANT_API_KEY
+
+    client_kwargs = {
+        "url": effective_url,
+        "timeout": timeout,
+    }
+    if effective_api_key:
+        client_kwargs["api_key"] = effective_api_key
+
+    return QdrantClient(**client_kwargs)
+
+
+def create_base_retriever(
+    collection_name: str,
+    embeddings: Embeddings,
+    search_kwargs: Optional[Dict[str, Any]] = None,
+    client: Optional[QdrantClient] = None,
+) -> BaseRetriever:
+    """
+    创建基础向量检索器（仅稠密向量检索）。
+
+    该检索器使用嵌入模型将查询转为向量，在 Qdrant 集合中执行 ANN 搜索，
+    返回语义上最相似的文档块。
+
+    Args:
+        collection_name: Qdrant 集合名称（需预先创建并索引）。
+        embeddings: LangChain 兼容的嵌入模型实例。
+        search_kwargs: 搜索参数，可包含：
+            - k (int): 返回的文档数量，默认 20。
+            - score_threshold (float): 相似度阈值，仅返回高于此分数的文档。
+            - filter (dict): Qdrant 过滤条件。
+            若为 None，则使用默认值 {"k": 20}。
+        client: 可选的 Qdrant 客户端实例。若未提供，将自动创建。
+
+    Returns:
+        BaseRetriever 实例，可直接调用 .invoke(query) 或 .ainvoke(query) 检索。
+
+    Raises:
+        ValueError: 如果集合不存在或嵌入模型无效。
+    """
+    # 合并默认搜索参数
+    merged_search_kwargs = {"k": DEFAULT_SEARCH_K}
+    if search_kwargs:
+        merged_search_kwargs.update(search_kwargs)
+
+    # 创建或复用 Qdrant 客户端
+    if client is None:
+        client = create_qdrant_client()
+
+    # 验证集合是否存在（可选，便于提前发现问题）
+    try:
+        client.get_collection(collection_name)
+    except UnexpectedResponse as e:
+        if e.status_code == 404:
+            raise ValueError(
+                f"Qdrant 集合 '{collection_name}' 不存在，请先创建并索引文档。"
+            )
+        raise
+
+    # 构建向量存储
+    vector_store = QdrantVectorStore(
+        client=client,
+        collection_name=collection_name,
+        embedding=embeddings,
+    )
+
+    # 返回检索器
+    return vector_store.as_retriever(search_kwargs=merged_search_kwargs)
+
+
+def create_hybrid_retriever(
+    collection_name: str,
+    embeddings: Embeddings,
+    dense_k: int = 10,
+    sparse_k: int = 10,
+    score_threshold: Optional[float] = DEFAULT_SCORE_THRESHOLD,
+    client: Optional[QdrantClient] = None,
+) -> BaseRetriever:
+    """
+    创建混合检索器（稠密向量 + BM25 稀疏向量）。
+
+    混合检索结合了语义相似度（Dense）和关键词匹配（Sparse），
+    能够更好地处理专有名词、精确匹配等场景。
+
+    注意：此功能要求 Qdrant 集合已配置稀疏向量字段并生成了 BM25 索引。
+    若集合未配置稀疏向量，将回退到纯稠密检索（不会报错，但检索效果降级）。
+
+    Args:
+        collection_name: Qdrant 集合名称。
+        embeddings: 嵌入模型（用于稠密向量）。
+        dense_k: 稠密向量检索返回数量，默认 10。
+        sparse_k: 稀疏向量检索返回数量，默认 10。
+        score_threshold: 相似度阈值，默认 0.3。
+        client: 可选的 Qdrant 客户端实例。
+
+    Returns:
+        BaseRetriever 实例，配置了混合搜索参数。
+    """
+    total_k = dense_k + sparse_k
+
+    search_kwargs = {
+        "k": total_k,
+    }
+    if score_threshold is not None:
+        search_kwargs["score_threshold"] = score_threshold
+
+    # 复用基础检索器创建逻辑，只需调整搜索参数
+    return create_base_retriever(
+        collection_name=collection_name,
+        embeddings=embeddings,
+        search_kwargs=search_kwargs,
+        client=client,
+    )
+
+
+# 可选：提供异步友好的辅助函数
+async def acreate_base_retriever(
+    collection_name: str,
+    embeddings: Embeddings,
+    search_kwargs: Optional[Dict[str, Any]] = None,
+    client: Optional[QdrantClient] = None,
+) -> BaseRetriever:
+    """
+    异步创建基础向量检索器（与同步版本功能相同）。
+
+    适用于需要异步初始化的场景（例如在 FastAPI 启动事件中）。
+    """
+    # 由于 QdrantVectorStore 初始化本身是同步的，这里直接调用同步版本即可
+    return create_base_retriever(collection_name, embeddings, search_kwargs, client)
\ No newline at end of file
diff --git a/backend/app/rag/test.py b/backend/app/rag/test.py
new file mode 100644
index 0000000..5e5d19e
--- /dev/null
+++ b/backend/app/rag/test.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+"""
+RAG 系统使用示例（重构版）
+
+演示：
+1. 使用 IndexBuilder 获取父子块检索器
+2. 创建固定流程的 RAGPipeline（多路改写 → RRF融合 → 重排序 → 返回父文档）
+3. 将流水线封装为 LangChain 工具，供 Agent 调用
+"""
+
+import asyncio
+import sys
+import os
+
+from dotenv import load_dotenv
+
+# 加载环境变量（Qdrant URL、PostgreSQL 连接等）
+load_dotenv()
+
+# 添加项目根目录到路径
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../.."))
+
+from rag_indexer.index_builder import IndexBuilderConfig
+from rag_indexer.splitters import SplitterType
+from .pipeline import RAGPipeline
+from .tools import create_rag_tool_sync
+from pydantic import SecretStr
+# 使用本地 LLM（通过 OpenAI 兼容接口）
+from langchain_openai import ChatOpenAI
+from rag_core.retriever_factory import create_parent_retriever
+
+load_dotenv()
+
+def create_llm():
+        """创建本地 vLLM 服务 LLM"""
+        vllm_base_url = os.getenv(
+            "VLLM_BASE_URL",
+            "http://127.0.0.1:8081/v1"
+        )
+        
+        return ChatOpenAI(
+            base_url=vllm_base_url,
+            api_key=SecretStr(os.getenv("LLAMACPP_API_KEY", "token-abc123")),
+            model="gemma-4-E2B-it",
+            timeout=60.0,          # 请求超时时间（秒）
+            max_retries=2,          # 失败后自动重试次数
+            streaming=True,         # 确保开启流式输出
+        )
+
+async def demonstrate_full_pipeline():
+    """
+    完整流水线演示：
+    - 从 IndexBuilder 获取 ParentDocumentRetriever
+    - 创建 RAGPipeline
+    - 执行检索并打印结果
+    """
+    print("=" * 60)
+    print("演示：固定流程 RAG 检索（多路改写 + RRF + 重排序 + 父文档）")
+    print("=" * 60)
+
+    retriever = create_parent_retriever(collection_name="rag_documents", search_k=5)
+
+    if retriever is None:
+        print("错误：检索器未初始化，请确保索引已构建。")
+        return
+
+    # 3. 创建 LLM 用于查询改写
+    llm = create_llm()
+
+    # 4. 创建 RAGPipeline（固定流程）
+    pipeline = RAGPipeline(
+        retriever=retriever,
+        llm=llm,
+        num_queries=3,          # 生成 3 个查询变体
+        rerank_top_n=5,         # 最终返回 5 个父文档
+    )
+
+    # 5. 执行检索
+    query = "打虎英雄是谁？"
+    print(f"\n查询: {query}")
+    print("-" * 40)
+
+    try:
+        documents = await pipeline.aretrieve(query)
+        print(f"返回 {len(documents)} 个父文档\n")
+
+        # 打印结果预览
+        for i, doc in enumerate(documents, 1):
+            content_preview = doc.page_content.replace("\n", " ")[:150]
+            source = doc.metadata.get("source", "未知来源")
+            print(f"{i}. 【来源：{source}】")
+            print(f"   {content_preview}...\n")
+
+        # 可选：格式化完整上下文
+        # context = pipeline.format_context(documents)
+        # print(context)
+
+    except Exception as e:
+        print(f"检索失败: {e}")
+        import traceback
+        traceback.print_exc()
+
+async def demonstrate_tool_creation():
+    """
+    演示创建 RAG 工具（供 Agent 使用）
+    """
+    print("\n" + "=" * 60)
+    print("演示：创建 RAG 工具（供 LangGraph Agent 调用）")
+    print("=" * 60)
+
+    # 1. 获取检索器（同上）
+    config = IndexBuilderConfig(
+        collection_name="rag_documents",
+        splitter_type=SplitterType.PARENT_CHILD,
+    )
+    retriever = retriever = create_parent_retriever(collection_name="rag_documents", search_k=5)
+
+    # 2. 创建 LLM
+    llm = create_llm()
+
+    # 3. 创建工具
+    rag_tool = create_rag_tool_sync(
+        retriever=retriever,
+        llm=llm,
+        num_queries=3,
+        rerank_top_n=5,
+        collection_name="rag_documents",
+    )
+
+    print(f"工具名称: {rag_tool.name}")
+    print(f"工具描述: {rag_tool.description[:100]}...")
+
+    # 4. 模拟 Agent 调用工具
+    query = "请告诉我 打虎英雄是谁？"
+    print(f"\n模拟调用: {query}")
+    print("-" * 40)
+
+    result = await rag_tool.ainvoke({"query": query})
+    print(result[:800] + "..." if len(result) > 800 else result)
+
+async def main():
+    await demonstrate_full_pipeline()
+    await demonstrate_tool_creation()
+
+if __name__ == "__main__":
+    asyncio.run(main())
\ No newline at end of file
diff --git a/backend/app/rag/tools.py b/backend/app/rag/tools.py
new file mode 100644
index 0000000..33f79c2
--- /dev/null
+++ b/backend/app/rag/tools.py
@@ -0,0 +1,54 @@
+"""
+RAG 工具模块
+
+将检索功能封装为 LangChain Tool，供 Agent 调用。
+采用固定流水线：多路改写 → 并行检索 → RRF 融合 → 重排序 → 返回父文档。
+"""
+from typing import Callable
+from langchain_core.tools import tool
+from langchain_core.language_models import BaseLanguageModel
+from langchain_core.retrievers import BaseRetriever
+from .pipeline import RAGPipeline
+
+def create_rag_tool_sync(
+    retriever: BaseRetriever,
+    llm: BaseLanguageModel,
+    num_queries: int = 3,
+    rerank_top_n: int = 5,
+    collection_name: str = "rag_documents",
+) -> Callable:
+    """
+    创建一个配置好的 RAG 检索工具（同步版本，用于不支持异步的旧版 Agent）。
+
+    参数同 create_rag_tool。
+    """
+    pipeline = RAGPipeline(
+        retriever=retriever,
+        llm=llm,
+        num_queries=num_queries,
+        rerank_top_n=rerank_top_n,
+    )
+
+    @tool
+    def search_knowledge_base_sync(query: str) -> str:
+        """在知识库中搜索与查询相关的文档片段（同步版本）。
+
+        功能与异步版本相同：多路改写 → RRF融合 → 重排序 → 返回父文档。
+
+        Args:
+            query: 用户提出的问题或查询字符串
+
+        Returns:
+            格式化后的相关文档内容。
+        """
+        try:
+            documents = pipeline.retrieve(query)  # 内部调用异步方法并等待
+            if not documents:
+                return f"在知识库 '{collection_name}' 中未找到与 '{query}' 相关的信息。"
+
+            context = pipeline.format_context(documents)
+            return context
+        except Exception as e:
+            return f"检索过程中发生错误: {str(e)}"
+
+    return search_knowledge_base_sync
\ No newline at end of file
diff --git a/backend/app/test_backend.py b/backend/app/test_backend.py
new file mode 100644
index 0000000..6af60d2
--- /dev/null
+++ b/backend/app/test_backend.py
@@ -0,0 +1,304 @@
+#!/usr/bin/env python3
+"""
+完整后端测试 - 验证 Agent 所有功能
+包括：短期记忆、长期记忆、工具调用、流式对话、历史查询
+"""
+
+import asyncio
+import os
+from .config import DB_URI
+import sys
+import uuid
+from dotenv import load_dotenv
+
+# 添加项目根目录到 Python 路径 (现在文件在 backend/app/ 下，backend 就是根)
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+load_dotenv()
+
+from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
+from ..agent import AIAgentService
+from ..agent.history import ThreadHistoryService
+from ..logger import info, warning, error
+
+# PostgreSQL 连接字符串
+
+async def print_section(title):
+    """打印测试区块标题"""
+    print("\n" + "=" * 70)
+    print(f"  {title}")
+    print("=" * 70)
+
+async def test_short_term_memory(agent_service):
+    """测试短期记忆（同一 thread_id 继续对话）"""
+    await print_section("测试 1: 短期记忆（Short-term Memory）")
+    
+    thread_id = str(uuid.uuid4())
+    user_id = "test_user_memory"
+    
+    print(f"\n使用 thread_id: {thread_id[:8]}...")
+    print(f"使用 user_id: {user_id}")
+    
+    # 第一轮对话
+    print("\n[第一轮] 发送消息: '我叫张三，今年28岁'")
+    result1 = await agent_service.process_message(
+        "我叫张三，今年28岁", thread_id, "local", user_id
+    )
+    print(f"回复: {result1['reply'][:100]}...")
+    
+    # 第二轮对话 - 测试记忆
+    print("\n[第二轮] 发送消息: '我叫什么名字？今年多大？'")
+    result2 = await agent_service.process_message(
+        "我叫什么名字？今年多大？", thread_id, "local", user_id
+    )
+    print(f"回复: {result2['reply']}")
+    
+    # 验证记忆是否存在
+    if "张三" in result2['reply'] or "28" in result2['reply']:
+        print("\n✅ 短期记忆测试通过！")
+        return True
+    else:
+        print("\n❌ 短期记忆测试失败！")
+        return False
+
+async def test_tool_calling(agent_service):
+    """测试工具调用（RAG 搜索）"""
+    await print_section("测试 2: 工具调用（Tool Calling）")
+    
+    thread_id = str(uuid.uuid4())
+    user_id = "test_user_tools"
+    
+    print(f"\n使用 thread_id: {thread_id[:8]}...")
+    print(f"使用 user_id: {user_id}")
+    
+    # 发送需要 RAG 搜索的问题
+    print("\n发送消息: '请告诉我，打虎英雄是谁？'")
+    result = await agent_service.process_message(
+        "请告诉我，打虎英雄是谁？", thread_id, "local", user_id
+    )
+    print(f"回复: {result['reply'][:200]}...")
+    
+    # 检查是否调用了 RAG 工具（回复中会有水浒传相关内容）
+    if "武松" in result['reply'] or "李忠" in result['reply'] or "水浒传" in result['reply']:
+        print("\n✅ 工具调用测试通过！")
+        return True
+    else:
+        print("\n⚠️  工具调用测试结果不确定，需要手动验证")
+        return None
+
+async def test_streaming(agent_service):
+    """测试流式对话"""
+    await print_section("测试 3: 流式对话（Streaming）")
+    
+    thread_id = str(uuid.uuid4())
+    user_id = "test_user_stream"
+    
+    print(f"\n使用 thread_id: {thread_id[:8]}...")
+    print(f"使用 user_id: {user_id}")
+    
+    print("\n发送消息: '用100字介绍一下AI人工智能' (流式)...")
+    print("流式输出: ", end="", flush=True)
+    
+    full_reply = ""
+    chunk_count = 0
+    
+    try:
+        async for chunk in agent_service.process_message_stream(
+            "用100字介绍一下AI人工智能", thread_id, "local", user_id
+        ):
+            chunk_count += 1
+            if chunk.get("type") == "llm_token":
+                token = chunk.get("token", "")
+                print(token, end="", flush=True)
+                full_reply += token
+            elif chunk.get("type") == "state_update":
+                pass  # 状态更新不显示
+        
+        print(f"\n\n共收到 {chunk_count} 个 chunk")
+        print(f"完整回复长度: {len(full_reply)} 字")
+        
+        if chunk_count > 0 and len(full_reply) > 10:
+            print("\n✅ 流式对话测试通过！")
+            return True
+        else:
+            print("\n❌ 流式对话测试失败！")
+            return False
+            
+    except Exception as e:
+        print(f"\n❌ 流式对话异常: {e}")
+        return False
+
+async def test_history_service(agent_service, history_service):
+    """测试历史查询服务"""
+    await print_section("测试 4: 历史查询服务（History Service）")
+    
+    user_id = "test_user_history"
+    
+    # 先创建几个对话
+    print(f"\n为 user_id={user_id} 创建测试对话...")
+    
+    thread_ids = []
+    for i in range(3):
+        thread_id = str(uuid.uuid4())
+        thread_ids.append(thread_id)
+        
+        await agent_service.process_message(
+            f"这是第 {i+1} 个测试对话", thread_id, "local", user_id
+        )
+        print(f"  创建线程 {i+1}: {thread_id[:8]}...")
+    
+    # 1. 测试获取用户线程列表
+    print("\n[4.1] 测试获取用户线程列表...")
+    threads = await history_service.get_user_threads(user_id, limit=10)
+    print(f"  找到 {len(threads)} 个线程")
+    
+    if len(threads) >= 3:
+        print("  ✅ 线程列表查询通过")
+    else:
+        print("  ⚠️  线程数量少于预期")
+    
+    # 2. 测试获取单个线程的消息历史
+    if thread_ids:
+        test_thread_id = thread_ids[0]
+        print(f"\n[4.2] 测试获取线程消息历史 (thread_id={test_thread_id[:8]}...)")
+        messages = await history_service.get_thread_messages(test_thread_id)
+        print(f"  找到 {len(messages)} 条消息")
+        
+        if len(messages) >= 2:  # 至少有一问一答
+            print("  ✅ 消息历史查询通过")
+        else:
+            print("  ⚠️  消息数量少于预期")
+        
+        # 3. 测试获取线程摘要
+        print(f"\n[4.3] 测试获取线程摘要...")
+        summary = await history_service.get_thread_summary(test_thread_id)
+        print(f"  摘要: {summary.get('summary', '')[:50]}...")
+        print(f"  消息数: {summary.get('message_count', 0)}")
+        
+        if summary.get('message_count', 0) > 0:
+            print("  ✅ 线程摘要查询通过")
+        else:
+            print("  ⚠️  摘要查询结果不确定")
+    
+    return len(threads) >= 3
+
+async def test_long_term_memory(agent_service):
+    """测试长期记忆（mem0）"""
+    await print_section("测试 5: 长期记忆（Long-term Memory - mem0）")
+    
+    thread_id1 = str(uuid.uuid4())
+    thread_id2 = str(uuid.uuid4())  # 不同的线程
+    user_id = "test_user_longterm"
+    
+    print(f"\n使用 user_id: {user_id}")
+    print(f"线程 1: {thread_id1[:8]}...")
+    print(f"线程 2: {thread_id2[:8]}...")
+    
+    # 在第一个线程中保存信息
+    print("\n[线程 1] 发送消息: '记住，我的宠物名字叫小白，是一只猫'")
+    result1 = await agent_service.process_message(
+        "记住，我的宠物名字叫小白，是一只猫", thread_id1, "local", user_id
+    )
+    print(f"回复: {result1['reply'][:100]}...")
+    
+    # 等待一下，让 mem0 保存
+    await asyncio.sleep(1)
+    
+    # 在第二个线程中询问（不同的 thread_id）
+    print("\n[线程 2] 发送消息: '我的宠物叫什么名字？是什么动物？'")
+    result2 = await agent_service.process_message(
+        "我的宠物叫什么名字？是什么动物？", thread_id2, "local", user_id
+    )
+    print(f"回复: {result2['reply']}")
+    
+    # 验证长期记忆
+    if "小白" in result2['reply'] or "猫" in result2['reply']:
+        print("\n✅ 长期记忆测试通过！")
+        return True
+    else:
+        print("\n⚠️  长期记忆可能未启用，或需要手动验证")
+        return None
+
+async def main():
+    """主测试函数"""
+    print("\n" + "=" * 70)
+    print("  后端完整功能测试")
+    print("=" * 70)
+    
+    results = {}
+    
+    try:
+        # 创建数据库连接和服务
+        print("\n正在初始化数据库连接...")
+        async with AsyncPostgresSaver.from_conn_string(DB_URI) as checkpointer:
+            await checkpointer.setup()
+            print("✅ 数据库连接成功")
+            
+            # 创建服务实例
+            print("\n正在初始化 Agent 服务...")
+            agent_service = AIAgentService(checkpointer)
+            await agent_service.initialize()
+            print("✅ Agent 服务初始化成功")
+            
+            history_service = ThreadHistoryService(checkpointer)
+            print("✅ 历史服务初始化成功")
+            
+            print(f"\n可用模型: {list(agent_service.graphs.keys())}")
+            
+            # 运行测试
+            results["短期记忆"] = await test_short_term_memory(agent_service)
+            await asyncio.sleep(1)
+            
+            results["工具调用"] = await test_tool_calling(agent_service)
+            await asyncio.sleep(1)
+            
+            results["流式对话"] = await test_streaming(agent_service)
+            await asyncio.sleep(1)
+            
+            results["历史查询"] = await test_history_service(agent_service, history_service)
+            await asyncio.sleep(1)
+            
+            results["长期记忆"] = await test_long_term_memory(agent_service)
+            await asyncio.sleep(1)
+            
+            # 打印总结
+            await print_section("测试总结")
+            print("\n测试结果:")
+            print("-" * 40)
+            
+            pass_count = 0
+            fail_count = 0
+            skip_count = 0
+            
+            for test_name, result in results.items():
+                if result is True:
+                    status = "✅ 通过"
+                    pass_count += 1
+                elif result is False:
+                    status = "❌ 失败"
+                    fail_count += 1
+                else:
+                    status = "⚠️  待验证"
+                    skip_count += 1
+                print(f"  {test_name:12s}: {status}")
+            
+            print("-" * 40)
+            print(f"总计: {len(results)} 个测试")
+            print(f"通过: {pass_count}, 失败: {fail_count}, 待验证: {skip_count}")
+            
+            if fail_count == 0:
+                print("\n🎉 所有核心测试通过！")
+            else:
+                print(f"\n⚠️  有 {fail_count} 个测试失败")
+            
+    except Exception as e:
+        error(f"\n❌ 测试运行异常: {e}")
+        import traceback
+        traceback.print_exc()
+        return 1
+    
+    return 0 if fail_count == 0 else 1
+
+if __name__ == "__main__":
+    exit_code = asyncio.run(main())
+    sys.exit(exit_code)
diff --git a/backend/app/utils/__init__.py b/backend/app/utils/__init__.py
new file mode 100644
index 0000000..d8bd0f1
--- /dev/null
+++ b/backend/app/utils/__init__.py
@@ -0,0 +1,7 @@
+"""
+工具模块
+"""
+
+from .logging import log_state_change, print_llm_input
+
+__all__ = ["log_state_change", "print_llm_input"]
diff --git a/backend/app/utils/logging.py b/backend/app/utils/logging.py
new file mode 100644
index 0000000..ae513fd
--- /dev/null
+++ b/backend/app/utils/logging.py
@@ -0,0 +1,61 @@
+"""
+LangGraph 节点日志工具模块
+提供状态流转追踪和 LLM 输入输出打印功能
+"""
+
+from ..config import ENABLE_GRAPH_TRACE
+from ..logger import debug, info
+
+
+def log_state_change(node_name: str, state: dict, prefix: str = "进入"):
+    """
+    记录状态变化日志
+    
+    Args:
+        node_name: 节点名称
+        state: 当前状态
+        prefix: 日志前缀（"进入" 或 "离开"）
+    """
+    from app.logger import info
+    
+    messages = state.get("messages", [])
+    msg_count = len(messages)
+    last_msg = messages[-1] if messages else None
+    last_info = ""
+    if last_msg:
+        # 兼容 dict 和对象两种格式
+        if isinstance(last_msg, dict):
+            content_preview = str(last_msg.get("content", ""))[:10].replace("\n", " ")
+            msg_type = last_msg.get("type", "unknown")
+        else:
+            content_preview = str(last_msg.content)[:10].replace("\n", " ")
+            msg_type = getattr(last_msg, 'type', 'unknown')
+        last_info = f"{msg_type.upper()}: {content_preview}"
+    info(f"🔄 [{node_name}] {prefix} | 消息数:{msg_count} | 最后一条:{last_info}")
+
+
+def print_llm_input(prompt_value):
+    """
+    RunnableLambda 回调函数：打印格式化后发送给 LLM 的完整消息
+    
+    Args:
+        prompt_value: ChatPromptValue 对象，包含格式化后的消息列表
+        
+    Returns:
+        原样返回 prompt_value，不影响链式调用
+    """
+    if not ENABLE_GRAPH_TRACE:
+        return prompt_value
+
+    messages = prompt_value.messages  # ChatPromptValue 提供 .messages 属性
+
+    debug("\n" + "=" * 80)
+    debug("📤 [LLM输入] 格式化后发送给大模型的完整消息:")
+    debug(f"   总消息数: {len(messages)}")
+    debug("-" * 80)
+    for i, msg in enumerate(messages):
+        content_preview = str(msg.content)  # 完整输出
+        debug(f"   [{i}] {msg.type.upper():10s}: {content_preview}")
+    debug("\n" + "=" * 80 + "\n")
+
+    return prompt_value
diff --git a/backend/rag_core/__init__.py b/backend/rag_core/__init__.py
new file mode 100644
index 0000000..318a066
--- /dev/null
+++ b/backend/rag_core/__init__.py
@@ -0,0 +1,21 @@
+"""
+RAG Core - 公共 RAG 组件包
+
+提供嵌入模型、向量存储和文档存储的公共功能，被 rag_indexer 和 app/rag 共用。
+"""
+
+from .embedders import LlamaCppEmbedder
+from .vector_store import QdrantVectorStore, QDRANT_URL, QDRANT_API_KEY
+from .store import PostgresDocStore, create_docstore
+from .retriever_factory import create_parent_retriever
+
+
+__all__ = [
+    "LlamaCppEmbedder",
+    "QdrantVectorStore",
+    "QDRANT_URL",
+    "QDRANT_API_KEY",
+    "PostgresDocStore",
+    "create_docstore",
+    "create_parent_retriever",
+]
diff --git a/backend/rag_core/client.py b/backend/rag_core/client.py
new file mode 100644
index 0000000..d689a1f
--- /dev/null
+++ b/backend/rag_core/client.py
@@ -0,0 +1,27 @@
+# rag_core/client.py
+import os
+from .config import QDRANT_URL, QDRANT_API_KEY
+from typing import Optional
+from qdrant_client import QdrantClient
+
+
+
+def create_qdrant_client(
+    url: Optional[str] = None,
+    api_key: Optional[str] = None,
+    timeout: int = 300,  # 索引构建需要较长超时
+) -> QdrantClient:
+    effective_url = url or QDRANT_URL
+    effective_api_key = api_key or QDRANT_API_KEY
+
+    if not effective_url:
+        raise ValueError("Qdrant URL 未配置")
+
+    client_kwargs = {
+        "url": effective_url,
+        "timeout": timeout,
+    }
+    if effective_api_key:
+        client_kwargs["api_key"] = effective_api_key
+
+    return QdrantClient(**client_kwargs)
\ No newline at end of file
diff --git a/backend/rag_core/config.py b/backend/rag_core/config.py
new file mode 100644
index 0000000..0d06575
--- /dev/null
+++ b/backend/rag_core/config.py
@@ -0,0 +1,24 @@
+"""
+RAG Core 配置管理模块
+集中管理所有环境变量配置项，避免散落在各个文件中
+"""
+
+import os
+
+# ========== 向量数据库配置 ==========
+QDRANT_URL = os.getenv("QDRANT_URL", "http://127.0.0.1:6333")
+QDRANT_API_KEY = os.getenv("QDRANT_API_KEY", "")
+
+# ========== 嵌入服务配置 ==========
+LLAMACPP_EMBEDDING_URL = os.getenv("LLAMACPP_EMBEDDING_URL", "http://127.0.0.1:8082")
+LLAMACPP_API_KEY = os.getenv("LLAMACPP_API_KEY", "")
+
+# ========== 文档存储配置 ==========
+DB_URI = os.getenv(
+    "DB_URI",
+    "postgresql://postgres:***@ai-postgres:5432/langgraph_db?sslmode=disable"
+)
+DOCSTORE_URI = os.getenv("DOCSTORE_URI", DB_URI)
+
+# ========== 其他配置 ==========
+# 可以在此添加其他 RAG Core 专用的配置项
\ No newline at end of file
diff --git a/backend/rag_core/embedders.py b/backend/rag_core/embedders.py
new file mode 100644
index 0000000..eff11a0
--- /dev/null
+++ b/backend/rag_core/embedders.py
@@ -0,0 +1,83 @@
+"""
+嵌入模型包装器，用于 llama.cpp 服务。
+"""
+
+import os
+from .config import LLAMACPP_EMBEDDING_URL, LLAMACPP_API_KEY
+import httpx
+from typing import List, Optional
+
+from langchain_core.embeddings import Embeddings
+
+class LlamaCppEmbedder:
+    """通过 OpenAI 兼容 API 封装 llama.cpp 嵌入服务。"""
+
+    def __init__(
+        self,
+        base_url: Optional[str] = None,
+        api_key: Optional[str] = None,
+        model: str = "Qwen3-Embedding-0.6B-Q8_0",
+    ):
+        self.base_url = base_url or LLAMACPP_EMBEDDING_URL
+        self.api_key = api_key or LLAMACPP_API_KEY
+        self.model = model
+
+    def as_langchain_embeddings(self) -> Embeddings:
+        """创建 LangChain 兼容的嵌入实例。"""
+        return _LlamaCppLangchainAdapter(self)
+
+    def embed_documents(self, texts: List[str]) -> List[List[float]]:
+        """嵌入一批文档。"""
+        return self._call_embedding_api(texts)
+
+    def embed_query(self, text: str) -> List[float]:
+        """嵌入单个查询。"""
+        return self._call_embedding_api([text])[0]
+
+    def get_embedding_dimension(self) -> int:
+        """通过嵌入测试字符串获取嵌入维度。"""
+        test_embedding = self.embed_query("test")
+        return len(test_embedding)
+
+    def _call_embedding_api(self, texts: List[str]) -> List[List[float]]:
+        """直接调用 llama.cpp 嵌入 API。"""
+        base = self.base_url.rstrip("/")
+        if not base.endswith("/v1"):
+            base = base + "/v1"
+
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        payload = {
+            "input": texts,
+            "model": self.model,
+        }
+
+        with httpx.Client(timeout=120) as client:
+            response = client.post(
+                f"{base}/embeddings",
+                headers=headers,
+                json=payload,
+            )
+            response.raise_for_status()
+            data = response.json()
+
+            if isinstance(data, list):
+                return [item["embedding"] for item in data]
+            elif isinstance(data, dict) and "data" in data:
+                return [item["embedding"] for item in sorted(data["data"], key=lambda x: x["index"])]
+            else:
+                raise ValueError(f"未知的嵌入 API 响应格式: {data}")
+
+class _LlamaCppLangchainAdapter(Embeddings):
+    """将 LlamaCppEmbedder 适配为 LangChain Embeddings 接口。"""
+
+    def __init__(self, embedder: LlamaCppEmbedder):
+        self._embedder = embedder
+
+    def embed_documents(self, texts: List[str]) -> List[List[float]]:
+        return self._embedder.embed_documents(texts)
+
+    def embed_query(self, text: str) -> List[float]:
+        return self._embedder.embed_query(text)
\ No newline at end of file
diff --git a/backend/rag_core/retriever_factory.py b/backend/rag_core/retriever_factory.py
new file mode 100644
index 0000000..25dab1c
--- /dev/null
+++ b/backend/rag_core/retriever_factory.py
@@ -0,0 +1,59 @@
+# rag_core/retriever_factory.py 
+from langchain_core.embeddings import Embeddings
+from langchain_classic.retrievers import ParentDocumentRetriever
+from langchain_text_splitters import RecursiveCharacterTextSplitter
+from typing import Optional
+from langchain_core.embeddings import Embeddings
+from langchain_core.stores import BaseStore
+from langchain_text_splitters import RecursiveCharacterTextSplitter, TextSplitter
+from langchain_classic.retrievers import ParentDocumentRetriever
+
+from rag_core import LlamaCppEmbedder, QdrantVectorStore, create_docstore
+
+def create_parent_retriever(
+    collection_name: str = "rag_documents",
+    embeddings: Optional[Embeddings] = None,
+    parent_splitter: Optional[TextSplitter] = None,
+    child_splitter: Optional[TextSplitter] = None,
+    docstore: Optional[BaseStore] = None,
+    search_k: int = 5,
+    # 若未传入切分器，则用以下参数创建默认切分器
+    parent_chunk_size: int = 1000,
+    parent_chunk_overlap: int = 100,
+    child_chunk_size: int = 200,
+    child_chunk_overlap: int = 20,
+) -> ParentDocumentRetriever:
+    # 嵌入模型
+    if embeddings is None:
+        embedder = LlamaCppEmbedder()
+        embeddings = embedder.as_langchain_embeddings()
+
+    # 向量存储（只读）
+    vector_store = QdrantVectorStore(
+        collection_name=collection_name,
+        embeddings=embeddings,
+    )
+
+    # 切分器（若未提供则创建默认）
+    if parent_splitter is None:
+        parent_splitter = RecursiveCharacterTextSplitter(
+            chunk_size=parent_chunk_size,
+            chunk_overlap=parent_chunk_overlap,
+        )
+    if child_splitter is None:
+        child_splitter = RecursiveCharacterTextSplitter(
+            chunk_size=child_chunk_size,
+            chunk_overlap=child_chunk_overlap,
+        )
+
+    # 文档存储
+    if docstore is None:
+        docstore, _ = create_docstore()  # 从环境变量读取连接
+
+    return ParentDocumentRetriever(
+        vectorstore=vector_store.get_langchain_vectorstore(),
+        docstore=docstore,
+        child_splitter=child_splitter,
+        parent_splitter=parent_splitter,
+        search_kwargs={"k": search_k},
+    )
\ No newline at end of file
diff --git a/backend/rag_core/store/__init__.py b/backend/rag_core/store/__init__.py
new file mode 100644
index 0000000..359db76
--- /dev/null
+++ b/backend/rag_core/store/__init__.py
@@ -0,0 +1,31 @@
+"""
+文档存储模块 - 用于 ParentDocumentRetriever 的父文档存储。
+
+提供 PostgreSQL 存储后端：
+- PostgresDocStore: PostgreSQL 数据库存储（生产环境）
+
+示例用法：
+    >>> from rag_core.store import create_docstore
+    
+    >>> # 创建 PostgreSQL 存储
+    >>> store, conn = create_docstore(
+    ...     connection_string="postgresql://user:pass@host:5432/db",
+    ...     table_name="parent_docs"
+    ... )
+"""
+
+
+from .postgres import PostgresDocStore
+from .factory import create_docstore, get_docstore_uri, DEFAULT_DB_URI
+
+__version__ = "2.0.0"
+
+__all__ = [
+    # 具体实现
+    "PostgresDocStore",
+    
+    # 工厂函数
+    "create_docstore",
+    "get_docstore_uri",
+    "DEFAULT_DB_URI",
+]
diff --git a/backend/rag_core/store/factory.py b/backend/rag_core/store/factory.py
new file mode 100644
index 0000000..6c87ac9
--- /dev/null
+++ b/backend/rag_core/store/factory.py
@@ -0,0 +1,71 @@
+"""
+文档存储工厂 - 创建不同类型的存储实例。
+
+提供统一的接口来创建本地文件存储或 PostgreSQL 存储。
+"""
+
+import os
+from ..config import DB_URI, DOCSTORE_URI
+import logging
+from typing import Optional, Tuple
+
+from langchain_core.stores import BaseStore
+from .postgres import PostgresDocStore
+
+logger = logging.getLogger(__name__)        
+
+# 默认连接字符串（从环境变量读取）
+DEFAULT_DB_URI = DB_URI
+
+
+def get_docstore_uri() -> str:
+    """获取 docstore 专用的数据库连接字符串（可与主库相同）"""
+    return DOCSTORE_URI
+
+
+def create_docstore(
+    store_type: str = "postgres",
+    connection_string: Optional[str] = None,
+    table_name: str = "parent_documents",
+    pool_config: Optional[dict] = None,
+    max_concurrency: Optional[int] = None
+) -> Tuple[BaseStore, Optional[str]]:
+    """
+    工厂函数，创建 PostgreSQL 文档存储。
+    
+    Args:
+        store_type: 存储类型，目前仅支持 "postgres"（默认）
+        connection_string: PostgreSQL 连接字符串
+        table_name: PostgreSQL 表名（默认：parent_documents）
+        pool_config: 连接池配置
+        max_concurrency: 最大并发操作数，如果为 None 则不限制
+    
+    Returns:
+        元组 (存储实例, 连接字符串)
+    
+    Raises:
+        ValueError: 不支持的存储类型
+        ImportError: 缺少必要的依赖
+    
+    Example:
+        >>> # 创建 PostgreSQL 存储
+        >>> store, conn = create_docstore(
+        ...     connection_string="postgresql://user:pass@host:5432/db",
+        ...     table_name="parent_docs",
+        ...     max_concurrency=10
+        ... )
+    """
+    store_type = store_type.lower()
+    
+    if store_type == "postgres":
+        conn_str = connection_string or get_docstore_uri()
+        store = PostgresDocStore(
+            connection_string=conn_str,
+            table_name=table_name,
+            pool_config=pool_config,
+            max_concurrency=max_concurrency
+        )
+        return store, conn_str
+    
+    else:
+        raise ValueError(f"不支持的存储类型: {store_type}。目前仅支持: postgres")
diff --git a/backend/rag_core/store/postgres.py b/backend/rag_core/store/postgres.py
new file mode 100644
index 0000000..23b7153
--- /dev/null
+++ b/backend/rag_core/store/postgres.py
@@ -0,0 +1,246 @@
+"""
+异步 PostgreSQL 存储实现 - 用于生产环境。
+
+使用 asyncpg 实现真正的异步 PostgreSQL 文档存储，支持高并发访问。
+"""
+
+import asyncio
+import json
+import logging
+from typing import List, Dict, Any, Optional, Iterator, Tuple, Sequence
+
+from langchain_core.documents import Document
+from langchain_core.stores import BaseStore
+
+import asyncpg
+
+logger = logging.getLogger(__name__)
+
+class PostgresDocStore(BaseStore[str, Any]):
+    """
+    异步 PostgreSQL 文档存储实现。
+    
+    使用 asyncpg 作为异步 PostgreSQL 客户端，支持：
+    - 真正的异步操作
+    - 连接池管理
+    - 自动表创建
+    - 批量操作（amget/amset/amdelete）
+    - JSONB 数据存储
+    - 并发控制
+    
+    适用于生产环境，提供高性能的异步数据持久化。
+    
+    Attributes:
+        dsn: PostgreSQL 连接字符串
+        table_name: 存储表名，默认为 "parent_documents"
+        _pool: asyncpg 连接池实例
+        _semaphore: 控制并发数的信号量（可选）
+    """
+    
+    def __init__(
+        self,
+        connection_string: str,
+        table_name: str = "parent_documents",
+        pool_config: Optional[Dict[str, Any]] = None,
+        max_concurrency: Optional[int] = None
+    ):
+        """
+        初始化异步 PostgreSQL 文档存储。
+        
+        Args:
+            connection_string: PostgreSQL 连接 URL，格式：
+                "postgresql://user:password@host:port/database?sslmode=disable"
+            table_name: 存储表名，默认为 "parent_documents"
+            pool_config: 连接池配置字典，包含：
+                - min_size: 最小连接数（默认 2）
+                - max_size: 最大连接数（默认 10）
+            max_concurrency: 最大并发操作数，如果为 None 则不限制
+        
+        Raises:
+            ImportError: 未安装 asyncpg 时抛出
+        
+        Example:
+            >>> store = PostgresDocStore(
+            ...     "postgresql://user:pass@localhost:5432/mydb",
+            ...     table_name="parent_docs",
+            ...     pool_config={"min_size": 5, "max_size": 20},
+            ...     max_concurrency=10
+            ... )
+        """
+
+        
+        self.dsn = connection_string
+        self.table_name = table_name
+        self._pool: Optional["asyncpg.Pool"] = None
+        self._pool_config = pool_config or {}
+        
+        # 并发控制信号量
+        self._semaphore = None
+        if max_concurrency is not None and max_concurrency > 0:
+            self._semaphore = asyncio.Semaphore(max_concurrency)
+        
+        # 注意：连接池的异步初始化延迟到第一次使用时
+        # 表结构创建也延迟到第一次操作时
+    
+    async def _get_pool(self):
+        """获取或创建 asyncpg 连接池。"""
+        if self._pool is None:
+            import asyncpg
+            min_size = self._pool_config.get("min_size", 2)
+            max_size = self._pool_config.get("max_size", 10)
+            
+            try:
+                self._pool = await asyncpg.create_pool(
+                    dsn=self.dsn,
+                    min_size=min_size,
+                    max_size=max_size
+                )
+                logger.info(f"PostgreSQL 异步连接池已创建: {self.table_name}")
+                
+                # 初始化表结构
+                await self._create_table()
+            except Exception as e:
+                raise RuntimeError(f"PostgreSQL 异步连接池创建失败: {e}")
+        
+        return self._pool
+    
+    async def _create_table(self):
+        """创建存储表（如果不存在）。"""
+        pool = await self._get_pool()
+        async with pool.acquire() as conn:
+            async with conn.transaction():
+                await conn.execute(f"""
+                    CREATE TABLE IF NOT EXISTS {self.table_name} (
+                        key TEXT PRIMARY KEY,
+                        value JSONB NOT NULL,
+                        created_at TIMESTAMPTZ DEFAULT NOW()
+                    )
+                """)
+                logger.info(f"表 {self.table_name} 已就绪")
+    
+    async def _with_concurrency_control(self, coro):
+        """使用信号量控制并发执行。"""
+        if self._semaphore is None:
+            return await coro
+        async with self._semaphore:
+            return await coro
+    
+    # --- 同步方法（保持兼容性，但功能有限）---
+    
+    def mget(self, keys: Sequence[str]) -> List[Optional[Any]]:
+        """不支持同步操作，请使用异步 amget 方法。"""
+        raise NotImplementedError("不支持同步操作，请使用异步 amget 方法。")
+    
+    def mset(self, key_value_pairs: Sequence[Tuple[str, Any]]) -> None:
+        """不支持同步操作，请使用异步 amset 方法。"""
+        raise NotImplementedError("不支持同步操作，请使用异步 amset 方法。")
+    
+    def mdelete(self, keys: Sequence[str]) -> None:
+        """不支持同步操作，请使用异步 amdelete 方法。"""
+        raise NotImplementedError("不支持同步操作，请使用异步 amdelete 方法。")
+    
+    def yield_keys(self, *, prefix: str | None = None) -> Iterator[str]:
+        """不支持同步操作，请使用异步 ayield_keys 方法。"""
+        raise NotImplementedError("不支持同步操作，请使用异步 ayield_keys 方法。")
+    
+    # --- 异步方法（真正的实现）---
+    
+    async def amget(self, keys: Sequence[str]) -> List[Optional[Any]]:
+        """异步批量获取文档。"""
+        if not keys:
+            return []
+        
+        async def _amget():
+            pool = await self._get_pool()
+            async with pool.acquire() as conn:
+                rows = await conn.fetch(
+                    f"SELECT key, value FROM {self.table_name} WHERE key = ANY($1)",
+                    keys
+                )
+                result_map = {}
+                for row in rows:
+                    val = row['value']
+                    if isinstance(val, str):
+                        val = json.loads(val)
+                    if isinstance(val, dict) and 'page_content' in val:
+                        result_map[row['key']] = Document(**val)
+                    else:
+                        result_map[row['key']] = val
+                return [result_map.get(key) for key in keys]
+        
+        return await self._with_concurrency_control(_amget())
+    
+    async def amset(self, key_value_pairs: Sequence[Tuple[str, Any]]) -> None:
+        """异步批量设置文档。"""
+        if not key_value_pairs:
+            return
+        
+        async def _amset():
+            pool = await self._get_pool()
+            async with pool.acquire() as conn:
+                async with conn.transaction():
+                    await conn.executemany(
+                        f"""
+                        INSERT INTO {self.table_name} (key, value)
+                        VALUES ($1, $2)
+                        ON CONFLICT (key) DO UPDATE SET value = EXCLUDED.value
+                        """,
+                        [
+                            (k, json.dumps(v.dict() if isinstance(v, Document) else v, ensure_ascii=False))
+                            for k, v in key_value_pairs
+                        ]
+                    )
+                    logger.debug(f"已异步批量设置 {len(key_value_pairs)} 个文档")
+        
+        await self._with_concurrency_control(_amset())
+    
+    async def amdelete(self, keys: Sequence[str]) -> None:
+        """异步批量删除文档。"""
+        if not keys:
+            return
+        
+        async def _amdelete():
+            pool = await self._get_pool()
+            async with pool.acquire() as conn:
+                async with conn.transaction():
+                    await conn.execute(
+                        f"DELETE FROM {self.table_name} WHERE key = ANY($1)",
+                        keys
+                    )
+                    logger.debug(f"已异步批量删除 {len(keys)} 个文档")
+        
+        await self._with_concurrency_control(_amdelete())
+    
+    async def ayield_keys(self, *, prefix: str | None = None) -> Iterator[str]:
+        """异步迭代所有键。
+        
+        注意：这是一个异步生成器，需要使用 async for 迭代。
+        """
+        pool = await self._get_pool()
+        async with pool.acquire() as conn:
+            if prefix:
+                rows = await conn.fetch(
+                    f"SELECT key FROM {self.table_name} WHERE key LIKE $1 ORDER BY key",
+                    f"{prefix}%"
+                )
+            else:
+                rows = await conn.fetch(
+                    f"SELECT key FROM {self.table_name} ORDER BY key"
+                )
+            
+            for row in rows:
+                yield row['key']
+    
+    async def aclose(self) -> None:
+        """异步关闭连接池，释放资源。"""
+        if self._pool:
+            await self._pool.close()
+            self._pool = None
+            logger.info("PostgreSQL 异步连接池已关闭")
+    
+    def close(self) -> None:
+        """同步关闭连接池（功能有限）。
+        
+        注意：在异步环境中，请使用 aclose 方法。
+        """
+        pass
diff --git a/backend/rag_core/vector_store.py b/backend/rag_core/vector_store.py
new file mode 100644
index 0000000..b2ecd20
--- /dev/null
+++ b/backend/rag_core/vector_store.py
@@ -0,0 +1,180 @@
+"""
+Qdrant 向量数据库包装器。
+"""
+
+import logging
+import os
+from .config import QDRANT_URL, QDRANT_API_KEY
+import time
+from typing import List, Optional, Dict, Any
+
+from langchain_core.documents import Document
+from langchain_qdrant import QdrantVectorStore as LangchainQdrantVS
+from qdrant_client import QdrantClient
+from qdrant_client.http.models import Distance, VectorParams
+from httpx import RemoteProtocolError
+from qdrant_client.http.exceptions import ResponseHandlingException
+from .client import create_qdrant_client
+
+logger = logging.getLogger(__name__)
+
+
+
+class QdrantVectorStore:
+    """Qdrant 向量数据库操作包装器。"""
+
+    def __init__(
+        self,
+        collection_name: str,
+        embeddings: Optional[Any] = None,
+    ):
+        self.collection_name = collection_name
+        self._client: Optional[QdrantClient] = None
+        self._connection_attempts = 0
+        self._last_connection_time: Optional[float] = None
+
+        if embeddings is None:
+            from rag_core.embedders import LlamaCppEmbedder
+            embedder = LlamaCppEmbedder()
+            self.embeddings = embedder.as_langchain_embeddings()
+        else:
+            self.embeddings = embeddings
+
+        self.create_collection()
+
+        self.vector_store = LangchainQdrantVS(
+            client=self.get_client(),
+            collection_name=self.collection_name,
+            embedding=self.embeddings,
+        )
+
+    def get_client(self) -> QdrantClient:
+        if self._client is None:
+            self._client = create_qdrant_client(timeout=300)
+            self._connection_attempts += 1
+            self._last_connection_time = time.time()
+            logger.debug("Qdrant 客户端已创建 (第 %d 次连接)", self._connection_attempts)
+        return self._client
+
+    def refresh_client(self):
+        """关闭旧连接，创建新连接。"""
+        if self._client is not None:
+            try:
+                self._client.close()
+                logger.debug("Qdrant 旧连接已关闭")
+            except Exception as e:
+                logger.warning("关闭 Qdrant 连接时出现异常: %s", e)
+            finally:
+                self._client = None
+                self._last_connection_time = None
+
+    def check_connection_health(self) -> bool:
+        """检查连接健康状态，如果连接已失效则自动重建。"""
+        if self._client is None:
+            logger.info("Qdrant 客户端未初始化，将创建新连接")
+            return False
+        
+        try:
+            client = self.get_client()
+            client.get_collections()
+            logger.debug("Qdrant 连接健康检查通过")
+            return True
+        except (RemoteProtocolError, ConnectionError, OSError, ResponseHandlingException) as e:
+            logger.warning("Qdrant 连接健康检查失败: %s", e)
+            self.refresh_client()
+            return False
+
+    def get_connection_stats(self) -> Dict[str, Any]:
+        """获取连接统计信息。"""
+        return {
+            "connection_attempts": self._connection_attempts,
+            "last_connection_time": self._last_connection_time,
+            "client_initialized": self._client is not None,
+        }
+
+    def create_collection(self, vector_size: Optional[int] = None, force_recreate: bool = False):
+        """创建集合，设置合适的向量维度。"""
+        if vector_size is None:
+            from rag_core.embedders import LlamaCppEmbedder
+            embedder = LlamaCppEmbedder()
+            vector_size = embedder.get_embedding_dimension()
+
+        max_retries = 3
+        base_delay = 2
+        for attempt in range(max_retries):
+            try:
+                client = self.get_client()
+                collections = client.get_collections().collections
+                exists = any(c.name == self.collection_name for c in collections)
+
+                if exists and force_recreate:
+                    client.delete_collection(self.collection_name)
+                    exists = False
+
+                if not exists:
+                    client.create_collection(
+                        collection_name=self.collection_name,
+                        vectors_config=VectorParams(size=vector_size, distance=Distance.COSINE),
+                    )
+                    logger.info("集合 '%s' 已创建（维度=%d）", self.collection_name, vector_size)
+                else:
+                    logger.info("集合 '%s' 已存在", self.collection_name)
+                return
+            except (RemoteProtocolError, ConnectionError, OSError, ResponseHandlingException) as e:
+                if attempt == max_retries - 1:
+                    logger.error("创建集合 '%s' 重试 %d 次后仍然失败: %s", self.collection_name, max_retries, e)
+                    raise
+                wait_time = base_delay * (2 ** attempt)
+                error_type = type(e).__name__
+                logger.warning(
+                    "创建集合 '%s' 遇到网络异常 [%s]，%d秒后重试 (%d/%d): %s",
+                    self.collection_name, error_type, wait_time, attempt + 1, max_retries, e
+                )
+                self.refresh_client()
+                logger.debug("已刷新 Qdrant 客户端连接")
+                time.sleep(wait_time)
+
+    def add_documents(self, documents: List[Document], batch_size: int = 100):
+        """将文档添加到向量数据库。"""
+        if not documents:
+            return []
+        self.create_collection()
+        ids = self.vector_store.add_documents(documents, batch_size=batch_size)
+        logger.info("已向 '%s' 添加 %d 个文档", self.collection_name, len(ids))
+        return ids
+
+    def similarity_search(self, query: str, k: int = 5) -> List[Document]:
+        return self.vector_store.similarity_search(query, k=k)
+
+    def similarity_search_with_score(self, query: str, k: int = 5) -> List[tuple[Document, float]]:
+        return self.vector_store.similarity_search_with_score(query, k=k)
+
+    def delete_collection(self):
+        self.get_client().delete_collection(self.collection_name)
+        logger.info("集合 '%s' 已删除", self.collection_name)
+
+    def get_collection_info(self) -> Dict[str, Any]:
+        info = self.get_client().get_collection(self.collection_name)
+        vectors_config = info.config.params.vectors
+        if isinstance(vectors_config, dict):
+            first_config = next(iter(vectors_config.values()), None)
+            vector_size = first_config.size if first_config else 0
+        else:
+            vector_size = vectors_config.size if vectors_config else 0
+        return {
+            "name": self.collection_name,
+            "vectors_count": info.points_count or 0,
+            "status": info.status,
+            "vector_size": vector_size,
+        }
+
+    def as_langchain_vectorstore(self):
+        return self.vector_store
+
+    def get_langchain_vectorstore(self):
+        """返回 LangChain Qdrant 向量存储对象（别名）"""
+        return self.vector_store
+
+    def get_qdrant_client(self):
+        """返回原生 Qdrant 客户端（如需手动管理 collection）"""
+        return self.get_client()
\ No newline at end of file
diff --git a/backend/requirements.txt b/backend/requirements.txt
new file mode 100644
index 0000000..e648caa
--- /dev/null
+++ b/backend/requirements.txt
@@ -0,0 +1,46 @@
+# Core
+pydantic==2.12.5
+python-dotenv==1.2.2
+typing-extensions==4.15.0
+
+# LangChain
+langchain==1.2.15
+langchain-community==0.4.1
+langchain-core==1.2.28
+langchain-openai==1.1.12
+langchain-qdrant==1.1.0
+langgraph==1.1.6
+langgraph-checkpoint-postgres==3.0.5
+tiktoken>=0.12.0
+
+# Vector DB
+qdrant-client==1.17.1
+
+# Memory
+mem0ai==1.0.11
+
+# Backend
+fastapi==0.135.3
+uvicorn[standard]==0.44.0
+
+# Database
+asyncpg==0.31.0
+psycopg[binary]==3.3.3
+
+# HTTP
+httpx==0.28.1
+aiohttp==3.13.5
+
+# Utilities
+tenacity==9.1.4
+rich==15.0.0
+PyYAML==6.0.3
+numpy>=1.26.2
+
+# Document Processing
+unstructured==0.22.21
+pypdf==6.10.0
+beautifulsoup4==4.14.3
+lxml==6.1.0
+pandas==3.0.2        # 若需Excel保留，否则移除
+spacy==3.8.14         # unstructured 可能依赖
diff --git a/frontend/src/config.py b/frontend/src/config.py
index da4c995..7b34700 100644
--- a/frontend/src/config.py
+++ b/frontend/src/config.py
@@ -57,6 +57,13 @@ class FrontendConfig:
         api_url = os.getenv("API_URL", "http://127.0.0.1:8079")
         self.api_base = api_url.replace("/chat", "").rstrip("/")
 
+        # 日志配置
+        self.log_level = os.getenv("LOG_LEVEL", "INFO").upper()
+        self.debug = os.getenv("DEBUG", "false").lower() == "true"
 
+
+        # 日志配置
+        self.log_level = os.getenv("LOG_LEVEL", "INFO").upper()
+        self.debug = os.getenv("DEBUG", "false").lower() == "true"
 # 全局配置实例（单例模式）
 config = FrontendConfig()
\ No newline at end of file
diff --git a/frontend/src/logger.py b/frontend/src/logger.py
index 1f3aefc..f5cacec 100644
--- a/frontend/src/logger.py
+++ b/frontend/src/logger.py
@@ -4,6 +4,7 @@
 """
 
 import os
+from .config import config
 import logging
 from typing import Any
 from dotenv import load_dotenv
@@ -14,10 +15,10 @@ load_dotenv()
 # ==================== 日志配置 ====================
 
 # 从环境变量读取日志级别，默认 INFO
-LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO").upper()
+LOG_LEVEL = config.log_level
 
 # 根据环境变量控制是否显示详细调试信息
-DEBUG_MODE = os.getenv("DEBUG", "false").lower() == "true"
+DEBUG_MODE = config.debug
 
 # 创建统一的日志器
 logger = logging.getLogger("ai_agent_frontend")
diff --git a/rag_indexer/config.py b/rag_indexer/config.py
new file mode 100644
index 0000000..218a431
--- /dev/null
+++ b/rag_indexer/config.py
@@ -0,0 +1,32 @@
+"""
+RAG Indexer 配置管理模块
+集中管理所有环境变量配置项，避免散落在各个文件中
+"""
+
+import os
+
+# 尝试从 rag_core 导入配置（如果可用）
+try:
+    from rag_core.config import (
+        QDRANT_URL,
+        QDRANT_API_KEY,
+        LLAMACPP_EMBEDDING_URL,
+        LLAMACPP_API_KEY,
+        DB_URI,
+        DOCSTORE_URI,
+    )
+except ImportError:
+    # 如果 rag_core 不可用，则直接读取环境变量
+    QDRANT_URL = os.getenv("QDRANT_URL", "http://127.0.0.1:6333")
+    QDRANT_API_KEY = os.getenv("QDRANT_API_KEY", "")
+    LLAMACPP_EMBEDDING_URL = os.getenv("LLAMACPP_EMBEDDING_URL", "http://127.0.0.1:8082")
+    LLAMACPP_API_KEY = os.getenv("LLAMACPP_API_KEY", "")
+    DB_URI = os.getenv(
+        "DB_URI",
+        "postgresql://postgres:huang1998@ai-postgres:5432/langgraph_db?sslmode=disable"
+    )
+    DOCSTORE_URI = os.getenv("DOCSTORE_URI", DB_URI)
+
+# ========== 索引器专用配置 ==========
+# 默认索引存储路径
+INDEX_STORAGE_PATH = os.getenv("INDEX_STORAGE_PATH", "./index_storage")
\ No newline at end of file