lcel_with_functinal_call.py

import os
import sys
os.environ["OPENAI_API_KEY"] = open(f"{os.environ['HOMEPATH']}\\.secret\\openai.txt", "r").read()
os.environ["GOOGLE_API_KEY"] = open(f"{os.environ['HOMEPATH']}\\.secret\\gemini.txt", "r").read()

from backend import weather_tool
from backend import short_talk_tool

from langchain.memory import ConversationBufferMemory
from langchain.schema.agent import AgentFinish
from langchain.tools.render import format_tool_to_openai_function
from langchain.agents.format_scratchpad import format_to_openai_functions
from langchain.agents.output_parsers import OpenAIFunctionsAgentOutputParser
from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.runnables import RunnableLambda, RunnablePassthrough
from langchain_openai import ChatOpenAI
from operator import itemgetter

# 出力フォーマットを定義
from langchain_core.output_parsers import JsonOutputParser
from langchain_core.pydantic_v1 import BaseModel, Field

class Reply(BaseModel):
    current_emotion: str = Field(description="maxe")
    character_reply: str = Field(description="れん's reply to User")

parser = JsonOutputParser(pydantic_object=Reply)

prompt_system = open("C:\\Users\\koduki\\git\\ai-tuber\\src\\backend\\prompt_system.txt", "r", encoding='utf-8').read()

prompt_for_chat = ChatPromptTemplate.from_messages([
    ("system", prompt_system),
    ("user", "{input}"),
    MessagesPlaceholder(variable_name="chat_history"),
    MessagesPlaceholder(variable_name="scratchpad"),
]).partial(format_instructions=parser.get_format_instructions())

prompt_for_tools = ChatPromptTemplate.from_messages([
    ("system", "You are agentai"),
    ("user", "{input}"),
])

tools = [weather_tool.weather_api, short_talk_tool.talk]

llm_for_chat   = ChatOpenAI(temperature=0, model='gpt-4-0613')
llm_with_tools = ChatOpenAI(temperature=0, model='gpt-3.5-turbo').bind(functions=[format_tool_to_openai_function(t) for t in tools])

def call_func(log):
    if isinstance(log, AgentFinish):
        return [(log, [])]
    else:
        tool = next(x for x in tools if x.name == log.tool)
        observation = tool.run(log.tool_input)
        return [(log, observation)]

def store_memory(response):
    input = {"input":response["input"]}
    output = {"output": response["return_values"].return_values['output']}
    memory.save_context(input, output)
    return output

memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
router = (
    RunnablePassthrough().assign(
        chat_history=RunnableLambda(memory.load_memory_variables) | itemgetter("chat_history"),
        scratchpad=prompt_for_tools | llm_with_tools | OpenAIFunctionsAgentOutputParser() | call_func | format_to_openai_functions
    )| RunnablePassthrough().assign(
        return_values=prompt_for_chat | llm_for_chat | OpenAIFunctionsAgentOutputParser(),
    )| store_memory
)

router.invoke({
    "input": "今日の東京の天気は？",
})