langgraph调用mcp，langsmith调试

简介

一份示例代码，使用langgraph构建一个简单的graph，模型加载mcp工具，使用langsmith查看效果

1, 代码示例

文件名：main_demo_mcp.py

python 复制代码

import main_utils

llm = main_utils.llm_ali  # 加载一个llm

from langgraph.graph import StateGraph, START, END
from typing import TypedDict, Literal
import asyncio
from langchain.messages import HumanMessage
from langchain_mcp_adapters.client import MultiServerMCPClient

# modelscope 的 MCP广场
mcp_config_gaode = {"transport": "streamable_http",
                    "url": "https://mcp.api-inference.modelscope.net/******/mcp"}

mcp_config_12306 = {"transport": "streamable_http",
                    "url": "https://mcp.api-inference.modelscope.net/*****/mcp"
                    }
# mcp client
mcp_client_ = MultiServerMCPClient({"mcp_config_gaode": mcp_config_gaode,
                                    "mcp_config_12306": mcp_config_12306})


class State(TypedDict):
    query: str
    answer: str | None
    intent: Literal["chat", "qa", "code"] | None


def detect_intent(state: State) -> State:
    query = state["query"]
    prompt = f"""根据用户的问题， 判断用户的意图是聊天、问答还是代码
    用户问题：{query}， 回答"chat", "qa", "code"中的一个"""
    response = llm.invoke(prompt).content

    response = response.strip()
    if response not in ["chat", "qa", "code"]:
        response = "qa"
    return {"intent": response.strip()}


def route_func(state: State) -> str | None:
    intent = state["intent"]
    if intent == "chat":
        return "chat_"
    elif intent == "qa":
        return "qa_"
    elif intent == "code":
        return "code_"
    else:
        return None


def chat_(state: State) -> State:
    query = state["query"]
    response = llm.invoke(query).content
    return {"answer": response.strip()}


async def qa_(state: State) -> State:
    query = state["query"]

    from langchain.agents import create_agent
    mcp_tools_list = await mcp_client_.get_tools()  #获取mcp工具列表

    # 打印 观看tool说明
    for itemid, item in enumerate(mcp_tools_list):
        print(itemid, item)

    # 定义带mcp工具的智能体
    agent_ = create_agent(model=llm, tools=mcp_tools_list)
    response = await agent_.ainvoke({"messages": HumanMessage(content=query)})
    response_1 = response["messages"]
    response_2 = response_1[-1].content #最终的结果
    return {"answer": response_2}


def code_(state: State) -> State:
    query = state["query"]
    prompt = f"""根据用户的问题， 生成对应的代码
    用户问题：{query}， 只输出代码，不输出解释"""
    response = llm.invoke(prompt).content
    return {"answer": response.strip()}


builder = StateGraph(State)
builder.add_node("detect_intent", detect_intent)
builder.add_node("chat_", chat_)
builder.add_node("code_", code_)
builder.add_node("qa_", qa_)

builder.add_edge(START, "detect_intent")
builder.add_edge("chat_", END)
builder.add_edge("qa_", END)
builder.add_edge("code_", END)

builder.add_conditional_edges("detect_intent",
                              route_func,
                              {
                                  "chat_": "chat_",
                                  "qa_": "qa_",
                                  "code_": "code_"}
                              )

graph = builder.compile()

# 画结构图
pic = graph.get_graph().draw_mermaid_png()
with open("graph.png", "wb") as f:
    f.write(pic)

# 如果只用langsmith的话，下面的代码都可以注释掉
async def main():
    query_ = "明天济南到郑州的火车"
    query_ = "郑州市的天气"
    #加载mcp为异步方式，graph也只能使用这种
    r = await graph.ainvoke({"query": query_}) 方式了
    for key in r:
        print("-----------   output key&value: ", key, r[key])


if __name__ == "__main__":
    asyncio.run(main())

graph结构如下

2, 配置langgraph.json

文件名：langgraph.json

上面的main_demo_mcp.py 文件，代码中的图名称为graph，需对应好。dependencies随便填写个值即可

python 复制代码

{
  "dependencies": [
    "langgraph"
  ],
  "graphs": {
    "eval_graph": "main_demo_mcp:graph"
  }
}

3, langsmith调试

终端下执行 langgraph dev 命令，即可，自动打开页面，登录即可。

如下图所示，eval_graph为langgraph.json中定义的字段，如果有多个图，可以在这里选择。

左侧方框为图的结构。

右上为第一轮对话，看到可以得到天气的信息，如果没有加载的高德的mcp工具，无法得到天气信息。

右下为第二轮对话，看到可以得到火车信息，如果没有加载12306的mcp工具的话，无法得到火车信息。