knowledge-agent/agent.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
使用 FastAPI + LangGraph 重构的 Agent 服务
提供强大的工作流管理和状态控制
"""

import json
import sys
import os
import time
from typing import Any, Dict, List, Optional, TypedDict, Annotated
from contextlib import asynccontextmanager

# 保证可以导入本项目模块
sys.path.append(os.path.dirname(os.path.abspath(__file__)))

# 禁用 LangSmith 追踪，避免网络连接错误
os.environ["LANGCHAIN_TRACING_V2"] = "false"
os.environ["LANGCHAIN_ENDPOINT"] = ""
os.environ["LANGCHAIN_API_KEY"] = ""

from fastapi import FastAPI, HTTPException, BackgroundTasks
from fastapi.responses import JSONResponse
from pydantic import BaseModel, Field
import uvicorn

# LangGraph 相关导入
try:
    from langgraph.graph import StateGraph, END
    HAS_LANGGRAPH = True
except ImportError:
    HAS_LANGGRAPH = False
    print("警告: LangGraph 未安装，将使用传统模式")

from utils.logging_config import get_logger
from agent_tools import QueryDataTool, IdentifyTool, StructureTool

# 创建 logger
logger = get_logger('Agent')

# 状态定义
class AgentState(TypedDict):
    request_id: str
    items: List[Dict[str, Any]]
    details: List[Dict[str, Any]]
    processed: int
    success: int
    current_index: int
    current_item: Optional[Dict[str, Any]]
    identify_result: Optional[Dict[str, Any]]
    error: Optional[str]
    status: str

# 请求模型
class TriggerRequest(BaseModel):
    requestId: str = Field(..., description="请求ID")

# 响应模型
class TriggerResponse(BaseModel):
    requestId: str
    processed: int
    success: int
    details: List[Dict[str, Any]]

# 全局变量
identify_tool = None

@asynccontextmanager
async def lifespan(app: FastAPI):
    """应用生命周期管理"""
    # 启动时初始化
    global identify_tool
    identify_tool = IdentifyTool()
    logger.info("Agent 服务启动完成")
    
    yield
    
    # 关闭时清理
    logger.info("Agent 服务正在关闭")

# 创建 FastAPI 应用
app = FastAPI(
    title="Knowledge Agent API",
    description="基于 LangGraph 的智能内容识别和结构化处理服务",
    version="2.0.0",
    lifespan=lifespan
)

# =========================
# LangGraph 工作流定义
# =========================

def create_langgraph_workflow():
    """创建 LangGraph 工作流"""
    if not HAS_LANGGRAPH:
        return None
    
    # 工作流节点定义
    
    def fetch_data(state: AgentState) -> AgentState:
        """获取待处理数据"""
        try:
            request_id = state["request_id"]
            logger.info(f"开始获取数据: requestId={request_id}")
            
            items = QueryDataTool.fetch_crawl_data_list(request_id)
            state["items"] = items
            state["processed"] = len(items)
            state["status"] = "data_fetched"
            
            logger.info(f"数据获取完成: requestId={request_id}, 数量={len(items)}")
            return state
            
        except Exception as e:
            logger.error(f"获取数据失败: {e}")
            state["error"] = str(e)
            state["status"] = "error"
            return state
    
    def process_item(state: AgentState) -> AgentState:
        """处理单个数据项"""
        try:
            items = state["items"]
            current_index = state.get("current_index", 0)
            
            if current_index >= len(items):
                state["status"] = "completed"
                return state
            
            item = items[current_index]
            state["current_item"] = item
            state["content_id"] = item.get('content_id') or ''
            state["task_id"] = item.get('task_id') or ''
            state["current_index"] = current_index + 1
            
            # 处理当前项
            crawl_data = item.get('crawl_data') or {}
            
            # Step 1: 识别
            identify_result = identify_tool.run(
                crawl_data if isinstance(crawl_data, dict) else {}
            )
            state["identify_result"] = identify_result
            
            # Step 2: 结构化并入库
            affected = StructureTool.store_parsing_result(
                state["request_id"], 
                {
                    "content_id": state["content_id"],
                    "task_id": state["task_id"]
                }, 
                identify_result
            )
            
            ok = affected is not None and affected > 0
            if ok:
                state["success"] += 1
            
            # 记录处理详情
            detail = {
                "index": current_index + 1,
                "dbInserted": ok,
                "identifyError": identify_result.get('error'),
                "status": "success" if ok else "failed"
            }
            state["details"].append(detail)
            
            state["status"] = "item_processed"
            logger.info(f"处理进度: {current_index + 1}/{len(items)} - {'成功' if ok else '失败'}")
            
            return state
            
        except Exception as e:
            logger.error(f"处理第 {current_index + 1} 项时出错: {e}")
            detail = {
                "index": current_index + 1,
                "dbInserted": False,
                "identifyError": str(e),
                "status": "error"
            }
            state["details"].append(detail)
            state["status"] = "item_error"
            return state
    
    def should_continue(state: AgentState) -> str:
        """判断是否继续处理"""
        if state.get("error"):
            return "end"
        
        current_index = state.get("current_index", 0)
        items = state.get("items", [])
        
        if current_index >= len(items):
            return "end"
        
        return "continue"
    
    # 构建工作流图
    workflow = StateGraph(AgentState)
    
    # 添加节点
    workflow.add_node("fetch_data", fetch_data)
    workflow.add_node("process_item", process_item)
    
    # 设置入口点
    workflow.set_entry_point("fetch_data")
    
    # 添加边
    workflow.add_edge("fetch_data", "process_item")
    workflow.add_conditional_edges(
        "process_item",
        should_continue,
        {
            "continue": "process_item",
            "end": END
        }
    )
    
    # 编译工作流，禁用 LangSmith 追踪
    return workflow.compile()

# 全局工作流实例
WORKFLOW = create_langgraph_workflow() if HAS_LANGGRAPH else None

# =========================
# FastAPI 接口定义
# =========================

@app.get("/")
async def root():
    """根路径，返回服务信息"""
    return {
        "service": "Knowledge Agent API",
        "version": "2.0.0",
        "status": "running",
        "langgraph_enabled": HAS_LANGGRAPH,
        "endpoints": {
            "parse": "/parse",
            "parse/async": "/parse/async",
            "health": "/health",
            "docs": "/docs"
        }
    }

@app.get("/health")
async def health_check():
    """健康检查接口"""
    return {
        "status": "healthy", 
        "timestamp": time.time(),
        "langgraph_enabled": HAS_LANGGRAPH
    }

@app.post("/parse", response_model=TriggerResponse)
async def parse_processing(request: TriggerRequest, background_tasks: BackgroundTasks):
    """
    解析内容处理
    
    - **requestId**: 请求ID，用于标识处理任务
    """
    try:
        logger.info(f"收到解析请求: requestId={request.requestId}")
        
        if WORKFLOW and HAS_LANGGRAPH:
            # 使用 LangGraph 工作流
            logger.info("使用 LangGraph 工作流处理")
            
            # 初始化状态
            initial_state = AgentState(
                request_id=request.requestId,
                items=[],
                details=[],
                processed=0,
                success=0,
                current_index=0,
                current_item=None,
                identify_result=None,
                error=None,
                status="started"
            )
            
            # 执行工作流
            final_state = WORKFLOW.invoke(
                initial_state,
                config={"configurable": {"thread_id": f"thread_{request.requestId}"}}
            )
            
            # 构建响应
            result = TriggerResponse(
                requestId=request.requestId,
                processed=final_state.get("processed", 0),
                success=final_state.get("success", 0),
                details=final_state.get("details", [])
            )
            
        else:
            # 回退到传统模式
            logger.info("使用传统模式处理")
            
            # 获取待处理数据
            items = QueryDataTool.fetch_crawl_data_list(request.requestId)
            if not items:
                return TriggerResponse(
                    requestId=request.requestId,
                    processed=0,
                    success=0,
                    details=[]
                )

            # 处理数据
            success_count = 0
            details: List[Dict[str, Any]] = []
            
            for idx, item in enumerate(items, start=1):
                try:
                    crawl_data = item.get('crawl_data') or {}
                    
                    # Step 1: 识别
                    identify_result = identify_tool.run(
                        crawl_data if isinstance(crawl_data, dict) else {}
                    )
                    
                    # Step 2: 结构化并入库
                    affected = StructureTool.store_parsing_result(
                        request.requestId, 
                        {
                            content_id: item.get('content_id') or '',
                            task_id: item.get('task_id') or ''
                        }, 
                        identify_result
                    )
                    
                    ok = affected is not None and affected > 0
                    if ok:
                        success_count += 1
                    
                    details.append({
                        "index": idx,
                        "dbInserted": ok,
                        "identifyError": identify_result.get('error'),
                        "status": "success" if ok else "failed"
                    })
                    
                except Exception as e:
                    logger.error(f"处理第 {idx} 项时出错: {e}")
                    details.append({
                        "index": idx,
                        "dbInserted": False,
                        "identifyError": str(e),
                        "status": "error"
                    })

            result = TriggerResponse(
                requestId=request.requestId,
                processed=len(items),
                success=success_count,
                details=details
            )
        
        logger.info(f"处理完成: requestId={request.requestId}, processed={result.processed}, success={result.success}")
        return result
        
    except Exception as e:
        logger.error(f"处理请求失败: {e}")
        raise HTTPException(status_code=500, detail=f"处理失败: {str(e)}")

@app.post("/parse/async")
async def parse_processing_async(request: TriggerRequest, background_tasks: BackgroundTasks):
    """
    异步解析内容处理（后台任务）
    
    - **requestId**: 请求ID，用于标识处理任务
    """
    try:
        logger.info(f"收到异步解析请求: requestId={request.requestId}")
        
        # 添加后台任务
        background_tasks.add_task(process_request_background, request.requestId)
        
        return {
            "requestId": request.requestId,
            "status": "processing",
            "message": "任务已提交到后台处理",
            "langgraph_enabled": HAS_LANGGRAPH
        }
        
    except Exception as e:
        logger.error(f"提交异步任务失败: {e}")
        raise HTTPException(status_code=500, detail=f"提交任务失败: {str(e)}")

async def process_request_background(request_id: str):
    """后台处理请求"""
    try:
        logger.info(f"开始后台处理: requestId={request_id}")
        
        if WORKFLOW and HAS_LANGGRAPH:
            # 使用 LangGraph 工作流
            initial_state = AgentState(
                request_id=request_id,
                items=[],
                details=[],
                processed=0,
                success=0,
                current_index=0,
                current_item=None,
                identify_result=None,
                error=None,
                status="started"
            )
            
            final_state = WORKFLOW.invoke(
                initial_state,
                config={"configurable": {"thread_id": f"thread_{request_id}"}}
            )
            logger.info(f"LangGraph 后台处理完成: requestId={request_id}, processed={final_state.get('processed', 0)}, success={final_state.get('success', 0)}")
            
        else:
            # 传统模式
            items = QueryDataTool.fetch_crawl_data_list(request_id)
            if not items:
                logger.info(f"后台处理完成: requestId={request_id}, 无数据需要处理")
                return

            success_count = 0
            for idx, item in enumerate(items, start=1):
                try:
                    crawl_data = item.get('crawl_data') or {}
                    content_id = item.get('content_id') or ''
                    
                    identify_result = identify_tool.run(
                        crawl_data if isinstance(crawl_data, dict) else {}
                    )
                    
                    affected = StructureTool.store_parsing_result(
                        request_id, 
                        {
                            content_id: item.get('content_id') or '',
                            task_id: item.get('task_id') or ''
                        },
                        identify_result
                    )
                    
                    if affected is not None and affected > 0:
                        success_count += 1
                    
                    logger.info(f"后台处理进度: {idx}/{len(items)} - {'成功' if affected else '失败'}")
                    
                except Exception as e:
                    logger.error(f"后台处理第 {idx} 项时出错: {e}")

            logger.info(f"传统模式后台处理完成: requestId={request_id}, processed={len(items)}, success={success_count}")
        
    except Exception as e:
        logger.error(f"后台处理失败: requestId={request_id}, error={e}")

if __name__ == "__main__":
    # 启动服务
    uvicorn.run(
        "agent:app",
        host="0.0.0.0",
        port=8080,
        reload=True,  # 开发模式，自动重载
        log_level="info"
    )