langchain_agent.py

import os
from langchain_openai import ChatOpenAI
from langchain.agents import create_agent
from langchain.tools import tool
from langgraph.checkpoint.memory import MemorySaver  # ← 加记忆

# ===== 第一步：定义工具 =====
@tool
def get_weather(city: str) -> str:
    """查询指定城市的天气情况"""
    data = {"北京": "晴, 25°C", "上海": "小雨, 22°C", "广州": "多云, 30°C"}
    return data.get(city, f"{city}天气未收录")

@tool
def get_order(order_id: str) -> str:
    """查询订单的状态信息"""
    orders = {"A001": "iPhone 16, 已发货", "B002": "机械键盘, 打包中"}
    return orders.get(order_id, f"订单{order_id}未找到")

tools = [get_weather, get_order]

# ===== 第二步：初始化模型 =====
llm = ChatOpenAI(
    model="deepseek-chat",
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
    temperature=0
)

# ===== 第三步：拼装 Agent + 记忆 =====
memory = MemorySaver()  # ← 保存对话状态的组件

agent_executor = create_agent(
    model=llm,
    tools=tools,
    system_prompt="你是一个全能助手，可以查天气、查订单。",
    checkpointer=memory  # ← 传入记忆，否则每次 invoke 都是新对话
)

# ===== 运行 =====
config = {"configurable": {"thread_id": "user-001"}}  # ← 线程 ID，同一 ID 共享记忆

print("=== LangChain Agent 有记忆+思维链版（输入 quit 退出）===")
while True:
    user_input = input("\n你: ")
    if user_input.lower() == "quit":
        break
    try:
        # stream 替代 invoke：每步实时返回，而不是等全部跑完
        for event in agent_executor.stream(
            {"messages": [("user", user_input)]},
            config=config,
            stream_mode="values"  # 返回每一步的完整状态
        ):
            msg = event["messages"][-1]

            # 判断这一步在做什么
            msg_type = getattr(msg, "type", "")

            if msg_type == "ai":
                # 模型的消息：可能是 tool_calls（决定调工具），也可能是 final answer
                tool_calls = getattr(msg, "tool_calls", []) or []
                if tool_calls:
                    for tc in tool_calls:
                        print(f"  ⚡ Action: {tc['name']}({tc['args']})")
                elif msg.content:
                    print(f"  💬 Answer: {msg.content}")

            elif msg_type == "tool":
                print(f"  👁️ Observation: {msg.content}")

    except Exception as e:
        print(f"执行出错: {e}")