LangChain Agent 实验过程剖析 OpenAI

什么是LangChain Agent

简单来说，用户像LangChain输入的内容未知。此时可以有一套工具聚集(也可以自界说工具)，将这套自界说工具托管给LLM,让其自己决定利用工具中的某一个(假如存在的话)
例子

首先，这里自界说了两个简单的工具

1. from langchain.tools import BaseTool
3. # 天气查询工具 ，无论查询什么都返回Sunny
4. class WeatherTool(BaseTool):
5.     name = "Weather"
6.     description = "useful for When you want to know about the weather"
8.     def _run(self, query: str) -> str:
9.         return "Sunny^_^"
11.     async def _arun(self, query: str) -> str:
12.         """Use the tool asynchronously."""
13.         raise NotImplementedError("BingSearchRun does not support async")
15. # 计算工具，暂且写死返回3
16. class CustomCalculatorTool(BaseTool):
17.     name = "Calculator"
18.     description = "useful for when you need to answer questions about math."
20.     def _run(self, query: str) -> str:
21.         return "3"
23.     async def _arun(self, query: str) -> str:
24.         raise NotImplementedError("BingSearchRun does not support async")

复制代码

接下来是针对于工具的简单调用：留意，这里利用OpenAI temperature=0需要限定为0

1. from langchain.agents import initialize_agent
2. from langchain.llms import OpenAI
3. from CustomTools import WeatherTool
4. from CustomTools import CustomCalculatorTool
6. llm = OpenAI(temperature=0)
8. tools = [WeatherTool(), CustomCalculatorTool()]
10. agent = initialize_agent(tools, llm, agent="zero-shot-react-description", verbose=True)
12. agent.run("Query the weather of this week,And How old will I be in ten years? This year I am 28")

复制代码

看一下完备的响应过程：

1. I need to use two different tools to answer this question
2. Action: Weather
3. Action Input: This week
4. Observation: Sunny^_^
5. Thought: I need to use a calculator to answer the second part of the question
6. Action: Calculator
7. Action Input: 28 + 10
8. Observation: 3
9. Thought: I now know the final answer
10. Final Answer: This week will be sunny and in ten years I will be 38.

复制代码

可以看到LangChain Agent 详细分析了每一个步调，而且准确的调用了每一个可用的方法，拿到了相应的返回值，甚至在末了还修复了28+10=3这个错误。
下面看看LangChain Agent是怎样做到这点的
工作原理

首先看看我输入的问题是什么：
Query the weather of this week,And How old will I be in ten years? This year I am 28
查询本周气候，以及十年后我多少岁，今年我28
LangChain Agent中，有一套模板可以套用：

1. PREFIX = """Answer the following questions as best you can. You have access to the following tools:"""
2. FORMAT_INSTRUCTIONS = """Use the following format:
4. Question: the input question you must answer
5. Thought: you should always think about what to do
6. Action: the action to take, should be one of [{tool_names}]
7. Action Input: the input to the action
8. Observation: the result of the action
9. ... (this Thought/Action/Action Input/Observation can repeat N times)
10. Thought: I now know the final answer
11. Final Answer: the final answer to the original input question"""
12. SUFFIX = """Begin!
14. Question: {input}
15. Thought:{agent_scratchpad}"""

复制代码

通过这个模板，加上我们的问题以及自界说的工具，会酿成下面这个样子,而且附带表明：

1. Answer the following questions as best you can.  You have access to the following tools: #  尽可能的去回答以下问题，你可以使用以下的工具：
4. Calculator: Useful for when you need to answer questions about math.
5. # 计算器：当你需要回答数学计算的时候可以用到
6. Weather: useful for When you want to know about the weather #  天气：当你想知道天气相关的问题时可以用到
7. Use the following format: # 请使用以下格式(回答)
10. Question: the input question you must answer #  你必须回答输入的问题
11. Thought: you should always think about what to do
12. # 你应该一直保持思考，思考要怎么解决问题
13. Action: the action to take, should be one of [Calculator, Weather] #  你应该采取[计算器,天气]之一
14. Action Input: the input to the action #  动作的输入
15. Observation: the result of the action # 动作的结果
16. ...  (this Thought/Action/Action Input/Observation can repeat N times) # 思考-行动-输入-输出 的循环可以重复N次
17. T
18. hought: I now know the final answer # 最后，你应该知道最终结果了
19. Final Answer: the final answer to the original input question # 针对于原始问题，输出最终结果
22. Begin! # 开始
23. Question: Query the weather of this week,And How old will I be in ten years?  This year I am 28 #  问输入的问题
24. Thought:

复制代码

通过这个模板向openai规定了一系列的规范，包罗如今现有哪些工具集，你需要思考答复什么问题，你需要用到哪些工具，你对工具需要输入什么内容,等等。
假如仅仅是如许，openAI会完全补完你的答复，中心无法插入任何内容。因此LangChain利用OpenAI的stop参数，截断了AI当前对话。"stop": ["\\nObservation: ", "\\n\\tObservation: "]
做了以上设定以后，OpenAI仅仅会给到Action和 Action Input两个内容就被stop早停了。
以下是OpenAI的响应内容：

1. I need to use the weather tool to answer the first part of the question, and the calculator to answer the second part.
2. Action: Weather
3. Action Input: This week

复制代码

到这里是OpenAI的响应效果，可见，很简单就拿到了Action和Action Input。
这里从Tools中找到name=Weather的工具，然后再将This Week传入方法。详细业务处理看详细环境。这里仅返回Sunny。
由于当前找到了Action和Action Input。 代表OpenAI认定当前使命链并没有竣事。因此像请求体后拼接效果：Observation: Sunny 而且让他再次思考Thought:
开启第二轮思考：
下面是再次请求的完备请求体:

1. Answer the following questions as best you can. You have access to the following tools:Calculator: Useful for when you need to answer questions about math.Weather: useful for When you want to know about the weatherUse the following format:Question: the input question you must answerThought: you should always think about what to doAction: the action to take, should be one of [Calculator, Weather]Action Input: the input to the actionObservation: the result of the action... (this Thought/Action/Action Input/Observation can repeat N times)Thought: I now know the final answerFinal Answer: the final answer to the original input questionBegin!Question: Query the weather of this week,And How old will I be in ten years? This year I am 28Thought: I need to use the weather tool to answer the first part of the question, and the calculator to answer the second part.
2. Action: Weather
3. Action Input: This week
4. Observation: Sunny^_^Thought:

复制代码

同第一轮一样，OpenAI再次举行思考，而且返回Action 和 Action Input 后，再次被早停。

1. I need to calculate my age in ten years
2. Action: Calculator
3. Action Input: 28 + 10

复制代码

由于计算器工具只会返回3，效果会拼接出一个错误的效果，构造成了一个新的请求体
举行第三轮请求：

1. Answer the following questions as best you can. You have access to the following tools:Calculator: Useful for when you need to answer questions about math.Weather: useful for When you want to know about the weatherUse the following format:Question: the input question you must answerThought: you should always think about what to doAction: the action to take, should be one of [Calculator, Weather]Action Input: the input to the actionObservation: the result of the action... (this Thought/Action/Action Input/Observation can repeat N times)Thought: I now know the final answerFinal Answer: the final answer to the original input questionBegin!Question: Query the weather of this week,And How old will I be in ten years? This year I am 28Thought: I need to use the weather tool to answer the first part of the question, and the calculator to answer the second part.
2. Action: Weather
3. Action Input: This week
4. Observation: Sunny^_^Thought:I need to calculate my age in ten years
5. Action: Calculator
6. Action Input: 28 + 10
7. Observation: 3Thought:

复制代码

此时两个问题全都拿到了效果，根据开头的限定，OpenAi在完全拿到效果以后会返回I now know the final answer。而且根据完备上下文。把多个效果举行归纳总结：下面是完备的相应效果：

1. I now know the final answer
2. Final Answer: I will be 38 in ten years and the weather this week is sunny.

复制代码

可以看到。ai严格的按照设定返回想要的内容，而且还以外的把28+10=3这个数学错误给改正了
以上，就是LangChain Agent的完备工作流程

来源：https://blog.csdn.net/qq_35361412/article/details/129797199
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！