rag.py

import multiprocessing
import time

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

class Rag:
    def __init__(self):
        self.vectors = None
        self.raw_database = None
        self.process_data = None
        self.embedding_model = None
        self.llm_model = None

        self.data_file_name = 'train.txt'
        self.embedding_model_name = "thenlper/gte-small"
        self.model_name = "HuggingFaceH4/zephyr-7b-beta"

        multiprocessing.freeze_support()

    def build_vector_database(self):
        if self.vectors is None:
            self.load_document()
            self.generate_chunks()

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

    def load_document(self):
        logger.info('Carregando arquivo no qual será baseado o RAG.')
        with open(self.data_file_name, 'r') as f:
            data = f.read()

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

    def generate_chunks(self):
        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)
        self.process_data = splitter.split_documents([self.raw_database])
        self.process_data = self.process_data[:5]  # TODO: REMOVER DEPOIS


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

    def load_model(self):
        if self.llm_model is None:
            load_dotenv()
            login(token=os.getenv('HF_TOKEN'))
            time.sleep(2)

            logger.info(f'Carregamento do modelo de linguagem principal: {self.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(self.model_name, quantization_config=bnb_config)
            tokenizer = AutoTokenizer.from_pretrained(self.model_name)

            self.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 {self.model_name} carregado com sucesso.')

    def get_answer(self, question, use_context=True):
        self.build_vector_database()
        self.load_model()

        if use_context:
            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|>
            """

            search_results = self.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'Prompt final: \n{final_prompt}\n')
            answer = self.llm_model(final_prompt)
            logger.info(f"Resposta da IA: {answer[0]['generated_text']}")

        else:
            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|>
            ---
            Here is the question you need to answer.

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

            final_prompt = prompt.format(question)
            logger.info(f'Prompt final: \n{final_prompt}\n')
            answer = self.llm_model(final_prompt)
            logger.info(f"Resposta da IA: {answer[0]['generated_text']}")

        return answer[0]['generated_text']