136 lines
5.0 KiB
Python
136 lines
5.0 KiB
Python
from typing import Dict,Any,Type
|
|
from langchain_core.tools import StructuredTool
|
|
from m11_mcp_advanced.mcp_client import MCPClient
|
|
from pydantic import Field,create_model
|
|
from contextlib import AsyncExitStack
|
|
|
|
|
|
class LangChainMCPAdapter:
|
|
"""
|
|
MCP适配器：将MCP客户端无缝转换为LangChain可用的工具集。
|
|
实现了上下文管理器协议，
|
|
"""
|
|
def __init__(self,mcp_client:MCPClient):
|
|
self.client = mcp_client
|
|
|
|
async def __aenter__(self):
|
|
"""进入上下文，自动建立连接"""
|
|
await self.client.connect()
|
|
return self
|
|
|
|
async def __aexit__(self,exc_type,exc_value,exc_tb):
|
|
"""退出上下文，自动清理资源"""
|
|
await self.client.cleanup()
|
|
|
|
@staticmethod
|
|
def _schema_to_pydantic(name:str,schema:Dict[str,Any]):
|
|
"""
|
|
将MCP的JSON Schema动态转换为Pydantic模型
|
|
这是让LLM理解参数要求的关键
|
|
"""
|
|
# print(f"🔧 调试: 工具 '{name}' 的 inputSchema = {schema}") # 查看 MCP 返回的原始 inputSchema
|
|
|
|
# 所有参数定义
|
|
properties = schema.get("properties",{}) # 允许为空
|
|
# 必需字段
|
|
required = schema.get("required",[]) # 允许为空
|
|
|
|
# 初始空字典
|
|
fields = {}
|
|
|
|
# 类型映射表：将JSON类型映射为Python类型
|
|
type_map = {
|
|
"string":str,
|
|
"integer":int,
|
|
"number":float,
|
|
"boolean":bool,
|
|
"array":list,
|
|
"object":dict
|
|
}
|
|
|
|
for field_name,field_info in properties.items():
|
|
# 1.获取字段类型
|
|
json_type = field_info.get("type","string")
|
|
python_type = type_map.get(json_type,Any)
|
|
|
|
# 2.获取描述
|
|
description = field_info.get("description","")
|
|
|
|
# 3.是否为必需项
|
|
# 如果是必填，默认值为 ... (Ellipsis): 否则为None
|
|
if field_name in required:
|
|
default_value = ...
|
|
else:
|
|
default_value = None
|
|
|
|
# 4.构建Pydantic字段定义 —— create_model 要求的特定格式
|
|
fields[field_name] = (python_type,Field(default=default_value,description=description))
|
|
|
|
# 动态创建一个Pydantic模型类
|
|
return create_model(f"{name}Schema",**fields)
|
|
|
|
async def get_tools(self):
|
|
"""
|
|
核心方法：获取并转换工具
|
|
返回的是标准的LangChain Tool列表，可以直接喂给bind_tools
|
|
"""
|
|
# 从MCP Server 获取原始工具列表
|
|
mcp_tools = await self.client.list_tools()
|
|
langchain_tools = []
|
|
|
|
|
|
for tool_info in mcp_tools:
|
|
# 1.动态生成参数模型 -- 要处理schema为空的情况
|
|
# inputSchema一般会放好MCP各种工具/参数的介绍
|
|
raw_schema = tool_info.get("input_schema",{})
|
|
args_model = self._schema_to_pydantic(tool_info["name"],raw_schema)
|
|
# 2.定义执行函数
|
|
async def _dynamic_tool_func(tool_name=tool_info["name"],**kwargs):
|
|
# ⚠️:必须绑定 tool_name 默认参数，否则循环会覆盖 tool_name
|
|
return await self.client.call_tool(tool_name,kwargs)
|
|
|
|
# 3.包装成llm可调用的工具(注入args_schema)
|
|
tool = StructuredTool.from_function(
|
|
coroutine=_dynamic_tool_func,
|
|
name=tool_info["name"],
|
|
description=tool_info["description"],
|
|
args_schema=args_model # 把说明书传给 LangChain
|
|
)
|
|
langchain_tools.append(tool)
|
|
return langchain_tools
|
|
|
|
@classmethod
|
|
async def load_mcp_tools(cls,stack: AsyncExitStack, configs: list):
|
|
"""
|
|
负责遍历配置，批量建立连接，收集所有工具。
|
|
使用stack将连接生命周期托管给上层
|
|
"""
|
|
all_tools = []
|
|
for conf in configs:
|
|
print(f'🔌 正在连接:{conf["name"]} == ({conf.get("transport","stdio")})...')
|
|
|
|
# 根据 transport 类型创建不同的客户端
|
|
transport = conf.get("transport","stdio")
|
|
if transport == "stdio":
|
|
# 初始化 Client
|
|
client = MCPClient(
|
|
transport="stdio",
|
|
command=conf["command"],
|
|
args=conf["args"],
|
|
env=conf.get("env") # 可选参数
|
|
)
|
|
else: # http
|
|
client = MCPClient(
|
|
transport="http",
|
|
url=conf["url"]
|
|
)
|
|
|
|
# 🔥:enter_async_context 替代了async with 缩进
|
|
# 这样无论有多少个MCP，代码层级都不会变深
|
|
adapter = await stack.enter_async_context(cls(client))
|
|
# 批量获取一个MCP下的所有工具
|
|
tools = await adapter.get_tools()
|
|
print(f' ✅️ 获取工具{[t.name for t in tools]}')
|
|
all_tools.extend(tools)
|
|
|
|
return all_tools |