AI vs AI

This example demonstrates a multi-model conversation where two LLMs talk to each other using LangChain, each maintaining its own message history via RunnableWithMessageHistory.

import uuid
from random import sample
from time import sleep

from langchain_classic.prompts.chat import ChatPromptTemplate, MessagesPlaceholder
from langchain_community.chat_message_histories.in_memory import ChatMessageHistory
from langchain_core.chat_history import BaseChatMessageHistory
from langchain_core.messages import HumanMessage
from langchain_core.runnables.history import RunnableWithMessageHistory

from gen_ai_hub.proxy.langchain.amazon import ChatBedrock
from gen_ai_hub.proxy.langchain.openai import ChatOpenAI

store = {}

available_models = [
    {
        "model_class": ChatBedrock,
        "model_kwargs": {"model_name": "amazon--titan-text-express"},
        "speaking_name": "Amazon Titan Text Express",
    },
    {
        "model_class": ChatBedrock,
        "model_kwargs": {"model_name": "anthropic--claude-3-haiku"},
        "speaking_name": "Anthropic Claude 3 Haiku",
    },
    {
        "model_class": ChatOpenAI,
        "model_kwargs": {"proxy_model_name": "gpt-4o-mini"},
        "speaking_name": "GPT 4o mini",
    },
]


def get_session_history(session_id: str) -> BaseChatMessageHistory:
    if session_id not in store:
        store[session_id] = ChatMessageHistory()
    return store[session_id]


def create_session_history() -> str:
    session_id = str(uuid.uuid4())
    store[session_id] = ChatMessageHistory()
    return session_id


def create_instructed_model(base_model, speaking_name, session_id):
    prompt = ChatPromptTemplate.from_messages([
        (
            "system",
            "You are the large language model {llmname}. Your goal is to answer questions nicely and keep a "
            "conversation going. Add a creative question unrelated to the previous conversation at the end of "
            "your answer. Also do everything with a pirate accent!",
        ),
        MessagesPlaceholder(variable_name="messages"),
    ])
    return {
        "model": prompt | base_model,
        "speaking_name": speaking_name,
        "session_id": session_id,
    }


def converse_ai(model_dict, prompt) -> str:
    config = {"configurable": {"session_id": model_dict["session_id"]}}
    model_with_message_history = RunnableWithMessageHistory(
        model_dict["model"],
        get_session_history,
        input_messages_key="messages",
    )
    response = model_with_message_history.invoke(
        {"messages": [HumanMessage(content=prompt)], "llmname": model_dict["speaking_name"]},
        config=config,
    )
    return response.content


def main():
    session_id_model_1 = create_session_history()
    session_id_model_2 = create_session_history()

    model_1, model_2 = sample(available_models, 2)

    model_1 = create_instructed_model(
        model_1["model_class"](**model_1["model_kwargs"]),
        speaking_name=model_1["speaking_name"],
        session_id=session_id_model_1,
    )
    model_2 = create_instructed_model(
        model_2["model_class"](**model_2["model_kwargs"]),
        speaking_name=model_2["speaking_name"],
        session_id=session_id_model_2,
    )

    model_of_next_turn = model_2
    model_of_last_turn = model_1
    last_response = "Introduce yourself."

    for _ in range(1, 11):
        last_response = converse_ai(model_of_next_turn, last_response)
        print("#########################################################")
        print(model_of_next_turn["speaking_name"] + ": " + last_response)
        print("#########################################################")
        model_of_next_turn, model_of_last_turn = model_of_last_turn, model_of_next_turn
        sleep(3)


if __name__ == "__main__":
    main()