背景

ChatGLM3-6B的函数调用功能，和LangChain的Tool调用，在prompt上并没有对齐。

参考：LLM大语言模型（十二）：关于ChatGLM3-6B不兼容Langchain 的Function Call_error: valueerror: caught exception: unknown forma-CSDN博客

因此在LangChain的Agent中调用ChatGLM3-6B时，单独对prompt进行了转换。

参考：LLM大语言模型（十三）：ChatGLM3-6B兼容Langchain的Function Call的一步一步的详细转换过程记录_langchain+chatglm3-CSDN博客

今日发现，使用LangChain的不同方式定义的tool，之前的prompt转换失效了。

LangChain中不同Tool定义方式，生成的prompt不同

方式一：@tool

from langchain_core.tools import tool

@tool

def calculator(calculation:str)->str:

    "Useful for when you need to calculate math problems"

    calculation = calculation.replace("^", "**")

    if "sqrt" in calculation:

        calculation = calculation.replace("sqrt", "math.sqrt")

    elif "log" in calculation:

        calculation = calculation.replace("log", "math.log")

    return eval(calculation)

在Agent执行时生成的prompt如下，关注红色部分：

'System: Respond to the human as helpfully and accurately as possible. You have access to the following tools:\n\n

calculator: calculator(calculation: str) -> str - Useful for when you need to calculate math problems, args: {\'calculation\': {\'title\': \'Calculation\', \'type\': \'string\'}}\n\n

Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).\n\nValid "action" values: "Final Answer" or calculator\n\nProvide only ONE action per $JSON_BLOB, as shown:\n\n```\n{\n  "action": $TOOL_NAME,\n  "action_input": $INPUT\n}\n```\n\nFollow this format:\n\nQuestion: input question to answer\nThought: consider previous and subsequent steps\nAction:\n```\n$JSON_BLOB\n```\nObservation: action result\n... (repeat Thought/Action/Observation N times)\nThought: I know what to respond\nAction:\n```\n{\n  "action": "Final Answer",\n  "action_input": "Final response to human"\n}\n\nBegin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation\nHuman: 34 * 34\n\n\n (reminder to respond in a JSON blob no matter what)'

 

方式二：继承BaseTool

import os
import requests

from typing import Type, Any
from langchain.tools import BaseTool
from pydantic import BaseModel, Field

class WeatherInput(BaseModel):
    location: str = Field(description="the location need to check the weather")


class Weather(BaseTool):
    name = "weather"
    description = "Use for searching weather at a specific location"
    args_schema: Type[BaseModel] = WeatherInput

    def __init__(self):
        super().__init__()

    def _run(self, location: str) -> dict[str, Any]:
        weather = {
            "temperature": "20度",
            "description": "温度适中",
        }
        return weather

在Agent执行时生成的prompt如下，关注红色部分：

System: Respond to the human as helpfully and accurately as possible. You have access to the following tools:\n\n

Calculator: Useful for when you need to calculate math problems, args: {\'calculation\': {\'description\': \'calculation to perform\', \'title\': \'Calculation\', \'type\': \'string\'}}\n\n

Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).\n\nValid "action" values: "Final Answer" or Calculator\n\nProvide only ONE action per $JSON_BLOB, as shown:\n\n```\n{\n  "action": $TOOL_NAME,\n  "action_input": $INPUT\n}\n```\n\nFollow this format:\n\nQuestion: input question to answer\nThought: consider previous and subsequent steps\nAction:\n```\n$JSON_BLOB\n```\nObservation: action result\n... (repeat Thought/Action/Observation N times)\nThought: I know what to respond\nAction:\n```\n{\n  "action": "Final Answer",\n  "action_input": "Final response to human"\n}\n\nBegin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation\nHuman: 34 * 34\n\n\n (reminder to respond in a JSON blob no matter what)

因为生成的prompt不同了，所以之前的prompt转换也就失效了。

LangChain使用上的一个坑。