src/agents/chat_llm.py

import yaml
from box import Box
from langchain_community.chat_models import ChatOpenAI
from langchain_community.llms import DeepInfra
from langchain.schema import HumanMessage, SystemMessage, AIMessage
import logging
import os


def chat_llm(messages, model, temperature, max_tokens, n, timeout=600, stop=None, return_tokens=False):
    if model.__contains__("gpt"):
        iterated_query = False
        try:
            chat = ChatOpenAI(model_name=model,
                            temperature=temperature,
                            max_tokens=max_tokens,
                            n=n,
                            request_timeout=timeout,
                            model_kwargs={"seed": 0,
                                            "top_p": 0
                                            })
        except Exception as e:
            print(f"Error in loading model: {e}")
            return None
    else:
        # deepinfra
        iterated_query = True
        chat = ChatOpenAI(model_name=model,
                          openai_api_key=None,
                          temperature=temperature,
                          max_tokens=max_tokens,
                          n=1,
                          request_timeout=timeout,
                          openai_api_base="https://api.deepinfra.com/v1/openai")

    longchain_msgs = []
    for msg in messages:
        if msg['role'] == 'system':
            longchain_msgs.append(SystemMessage(content=msg['content']))
        elif msg['role'] == 'user':
            print('human message', msg)
            longchain_msgs.append(HumanMessage(content=msg['content']))
        elif msg['role'] == 'assistant':
            print('8' * 20)
            print(msg['content'])
            longchain_msgs.append(AIMessage(content=msg['content']))
        else:
            raise NotImplementedError
    # add an empty user message to avoid no user message error
    longchain_msgs.append(HumanMessage(content=""))
    if n > 1 and iterated_query:
        response_list = []
        total_completion_tokens = 0
        total_prompt_tokens = 0
        for n in range(n):
            generations = chat.generate([longchain_msgs], stop=[
                stop] if stop is not None else None)
            responses = [
                chat_gen.message.content for chat_gen in generations.generations[0]]
            response_list.append(responses[0])
            completion_tokens = generations.llm_output['token_usage']['completion_tokens']
            prompt_tokens = generations.llm_output['token_usage']['prompt_tokens']
            total_completion_tokens += completion_tokens
            total_prompt_tokens += prompt_tokens
        responses = response_list
        completion_tokens = total_completion_tokens
        prompt_tokens = total_prompt_tokens
    else:
        generations = chat.generate([longchain_msgs], stop=[
                                    stop] if stop is not None else None)
        responses = [
            chat_gen.message.content for chat_gen in generations.generations[0]]
        completion_tokens = generations.llm_output['token_usage']['completion_tokens']
        prompt_tokens = generations.llm_output['token_usage']['prompt_tokens']

    return {
        'generations': responses,
        'completion_tokens': completion_tokens,
        'prompt_tokens': prompt_tokens
    }