rag_test_bkp.py

import multiprocessing
from langchain.docstore.document import Document as LangChainDocument
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from huggingface_hub import login
from loguru import logger
from transformers import pipeline
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import os
from dotenv import load_dotenv


def main():
    load_dotenv()
    logger.info('Carregando arquivo no qual será baseado o RAG.')
    with open('train.txt', 'r') as f:
        data = f.read()

    logger.info('Representando o documento utilizando o LangChainDocument.')
    raw_database = LangChainDocument(page_content=data)

    MARKDOWN_SEPARATORS = [
        "\n#{1,6} ",
        "```\n",
        "\n\\*\\*\\*+\n",
        "\n---+\n",
        "\n___+\n",
        "\n\n",
        "\n",
        " ",
        "",
    ]

    logger.info('Quebrando o documento para a criação dos chunks.')
    splitter = RecursiveCharacterTextSplitter(separators=MARKDOWN_SEPARATORS, chunk_size=1000, chunk_overlap=100)
    process_data = splitter.split_documents([raw_database])
    process_data = process_data[:5]  # TODO: REMOVER DEPOIS

    embedding_model_name = "thenlper/gte-small"
    logger.info(f'Definição do modelo de embeddings: {embedding_model_name}.')
    embedding_model = HuggingFaceEmbeddings(
        model_name=embedding_model_name,
        multi_process=True,
        model_kwargs={"device": "cpu"},  # TODO: AJUSTAR DEPOIS
        encode_kwargs={"normalize_embeddings": True},  # Set `True` for cosine similarity
    )

    logger.info('Criação da base de dados vetorial (em memória).')
    vectors = FAISS.from_documents(process_data, embedding_model)

    # model_name = "meta-llama/Llama-3.2-1B"
    model_name = "HuggingFaceH4/zephyr-7b-beta"
    # model_name = "mistralai/Mistral-7B-Instruct-v0.3"
    # model_name = "meta-llama/Llama-3.2-3B-Instruct"
    logger.info(f'Carregamento do modelo de linguagem principal: {model_name}')

    # bnb_config = BitsAndBytesConfig(
    #     load_in_4bit=True,
    #     bnb_4bit_use_double_quant=True,
    #     bnb_4bit_quant_type="nf4",
    #     bnb_4bit_compute_dtype=torch.bfloat16,
    # )
    # model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=bnb_config)
    model = AutoModelForCausalLM.from_pretrained(model_name)
    tokenizer = AutoTokenizer.from_pretrained(model_name)

    llm_model = pipeline(
        model=model,
        tokenizer=tokenizer,
        task="text-generation",
        do_sample=True,
        temperature=0.4,
        repetition_penalty=1.1,
        return_full_text=False,
        max_new_tokens=500
    )
    logger.info(f'Modelo {model_name} carregado com sucesso.')

    prompt = """
    <|system|>
    You are a helpful assistant that answers on medical questions based on the real information provided from different sources and in the context.
    Give the rational and well written response. If you don't have proper info in the context, answer "I don't know"
    Respond only to the question asked.

    <|user|>
    Context:
    {}
    ---
    Here is the question you need to answer.

    Question: {}
    ---
    <|assistant|>
    """

    question = "What is Cardiogenic shock?"
    search_results = vectors.similarity_search(question, k=3)

    logger.info('Contexto: ')
    for i, search_result in enumerate(search_results):
        logger.info(f"{i + 1}) {search_result.page_content}")

    context = " ".join([search_result.page_content for search_result in search_results])
    final_prompt = prompt.format(context, question)
    logger.info(f'\n{final_prompt}\n')

    answer = llm_model(final_prompt)

    logger.info("AI response: ", answer[0]['generated_text'])


if __name__ == '__main__':
    multiprocessing.freeze_support()
    access_token = os.getenv("ACCESS_TOKEN")
    login(token=access_token)
    logger.info('Login realizado com sucesso.')
    main()