# LangChain Record
## Custom functions with OpenAI Functions Agent
https://python.langchain.com/docs/modules/agents/how_to/custom-functions-with-openai-functions-agent

OpenAi version problems
<AttributeError: module ‘openai’ has no attribute ‘error’>
Because LangChain didn't update to fix it
(2023/11/24)
https://community.openai.com/t/attributeerror-module-openai-has-no-attribute-error/486676

pip install openai==0.28.1

## Pydantic + LangChain Tools

Abstract :
1. 先想好你要做什麼樣的功能，把這個功能寫成一個 function ，同時定義好參數
2. 之後透過 pydantic 來建立參數限制與定義 Input Class
3. LangChain 的 BaseTool 建立你要的工具方法與描述，並且引用上述的限制跟變數定義放在 args_schema , 而你已經完成實作的 function 則是包在 _run() 裏面讓 Agent 去執行。

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

# Pydantic
class YourClassNameInput(BaseModel):
    # variable : type = Field(Qequired ? , description='')
    # "variable" : Is your variable name
    # "type" : Variable data type , if the data is optional  you can use "Optional[type]"
    # "Field" : The Field function is used to customize and add metadata to fields of models.
    # "None" in the Field() also mean "Not Qequired Variable"
    # "..." in the Field() mean "Qequired Variable"

    # Examples
    city_county : Optional[str] = Field( None , description = """
            City name input by user For example, 台中->臺中市、台北->臺北市、雲林->雲林縣、 
            Note: In Chinese, "台" and "臺" have the same meaning. Therefore, we uniformly use "臺". 
            If a different input is provided, it will automatically be corrected to the standard county/city name.
            採用繁體中文輸入
        """ )

Pydantic : https://docs.pydantic.dev/2.5/
Fields : https://docs.pydantic.dev/latest/concepts/fields/

# LangChain
class YourClassNameInput(BaseTool):
    name = 'search_real_estate_by_criteria'
    description : ''
    args_schema: Type[BaseModel] = YourClassNameInput
    
    def _run(self , variables ):
        # here you can excute your function
        result = YourFunctionName(var)
        return result 
        # It will return to langchain to generate response
        
    def _arun(self , variables ):
        raise NotImplementedError("This tool does not support async")
        # We don't use that in this case
        

## Set your Agent

Agent : https://python.langchain.com/docs/modules/agents/

1. Import modules
1. Set tools
1. Create Model (ChatOpenAI)

from langchain.agents import AgentType, initialize_agent
from langchain.chat_models import ChatOpenAI

# Note this is my tool name. You should import yourself
from functions.RealEstate_Recommendation import RealEstateRecommendationTool


model = ChatOpenAI( 
    model="gpt-4-1106-preview" ,
    temperature= 0.9,
    openai_api_key ="Your API key or you can set in env" 
    )
    
tools = [RealEstateRecommendationTool()]
# If you have more than one tools you can set at this List

agent = initialize_agent(tools, model, agent=AgentType.OPENAI_FUNCTIONS, verbose=False)

ans = agent.run("What you want to ask to your chatbot")

print(ans)

## Chat Messages
這邊 Agent 無法使用要使用其他方法，但 ChatModel 可以用
https://python.langchain.com/docs/modules/memory/chat_messages/

from langchain.memory import ChatMessageHistory

history = ChatMessageHistory()

history.add_user_message("hi!")

history.add_ai_message("whats up?")

## Verbose
這邊會顯示 LangChain 執行的步驟（在 CMD 內顯示過程）
https://stackoverflow.com/questions/76405986/in-langchain-how-to-save-the-verbose-output-to-a-variable

## Callback
https://python.langchain.com/docs/modules/callbacks/

## Agent Memoery (chat history)
https://python.langchain.com/docs/modules/agents/how_to/add_memory_openai_functions