使用ChatGLM作为Langchain的后端模型使用Agent

langchain默认对接的OpenAI的chatgpt，如果要使用自己的模型的话，首先需要把自己的模型服务化，也就是提供一个web api，然后通过LLMs接入，比如我自己做了一个接口和ChatGPT类似的接口。

格式差不多是这样：

curl http://www.baifachuan.com:8004/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer token1" \
  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

{
    "id": "chatcmpl-77PZm95TtxE0oYLRx3cxa6HtIDI7s",
    "object": "chat.completion",
    "created": 1682000966,
    "model": "gpt-3.5-turbo-0301",
    "usage": {
        "prompt_tokens": 10,
        "completion_tokens": 10,
        "total_tokens": 20
    },
    "choices": [
        {
            "message": {
                "role": "assistant",
                "content": "Hello! How can I help you today?"
            },
            "finish_reason": "stop",
            "index": 0
        }
    ]
}

然后重载一下LLM，整个代码非常简单：

import time
import logging
import requests
from typing import Optional, List, Dict, Mapping, Any

import langchain
from langchain.llms.base import LLM
from langchain.cache import InMemoryCache

logging.basicConfig(level=logging.INFO)
# 启动llm的缓存
langchain.llm_cache = InMemoryCache()

class ChatGLM(LLM):
    # 模型服务url
    url = "http://www.baifachuan.com:8004/v1/chat/completions"
    chat_history = []

    @property
    def _llm_type(self) -> str:
        return "chatglm"

    def _construct_query(self, prompt: str) -> Dict:
        """构造请求体
        """
        self.chat_history.append({"role": "user", "content": prompt})
        query = {
            "model": "gpt-3.5-turbo",
            "messages": self.chat_history
        }
        return query

    @classmethod
    def _post(cls, url: str,
              query: Dict) -> Any:
        """POST请求
        """
        _headers = {"Content_Type": "application/json", "Authorization": "Bearer token1"}
        with requests.session() as sess:
            resp = sess.post(url,
                             json=query,
                             headers=_headers,
                             timeout=60)
        return resp

    def _call(self, prompt: str,
              stop: Optional[List[str]] = None) -> str:
        """_call
        """
        # construct query
        query = self._construct_query(prompt=prompt)

        # post
        resp = self._post(url=self.url,
                          query=query)

        if resp.status_code == 200:
            resp_json = resp.json()
            self.chat_history.append(resp_json['choices'][0]['message'])
            predictions = resp_json['choices'][0]['message']['content']
            return predictions
        else:
            return "请求模型"

    @property
    def _identifying_params(self) -> Mapping[str, Any]:
        """Get the identifying parameters.
        """
        _param_dict = {
            "url": self.url
        }
        return _param_dict


if __name__ == "__main__":
    llm = ChatGLM()
    while True:
        human_input = input("Human: ")

        begin_time = time.time() * 1000
        # 请求模型
        response = llm(human_input, stop=["you"])
        end_time = time.time() * 1000
        used_time = round(end_time - begin_time, 3)
        logging.info(f"chatGLM process time: {used_time}ms")

        print(f"ChatGLM: {response}")

由于已经有了一个继承LLM的ChatGLM，那么就可以无缝的和langchain对应起来了：

from langchain.tools import ShellTool
from langchain.agents import initialize_agent
from langchain.agents import AgentType
from langchain.agents import create_sql_agent
from langchain.agents.agent_toolkits import SQLDatabaseToolkit
from langchain.sql_database import SQLDatabase
from langchain.llms.openai import OpenAI
from langchain.agents import AgentExecutor

from chat_glm import ChatGLM

shell_tool = ShellTool()

llm = ChatGLM()

shell_tool.description = shell_tool.description + f"args {shell_tool.args}".replace("{", "{{").replace("}", "}}")
self_ask_with_search = initialize_agent([shell_tool], llm, agent=AgentType.CHAT_ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
self_ask_with_search.run("执行echo 1，返回执行结果。")

ChatGLM可以完成对话能力，但是对于action的能力，相对薄弱很多，这部分只能做微调了。

---

扫码手机观看或分享：