app.py

import os
from typing import Optional, Tuple
import gradio as gr
from langchain_community.document_loaders import PyPDFLoader, DirectoryLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_community.llms import HuggingFacePipeline
from langchain.chains import RetrievalQA
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
import torch
import tempfile
import time

# Configurações
EMBEDDING_MODEL = "sentence-transformers/all-mpnet-base-v2"
LLM_MODEL = "google/flan-t5-large"
DOCS_DIR = "documents"

class RAGSystem:
    def __init__(self):
        self.tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
        self.model = AutoModelForSeq2SeqLM.from_pretrained(
            LLM_MODEL,
            device_map="auto",
            torch_dtype=torch.float32
        )
        
        pipe = pipeline(
            "text2text-generation",
            model=self.model,
            tokenizer=self.tokenizer,
            max_length=512,
            temperature=0.7,
            top_p=0.95
        )
        
        self.llm = HuggingFacePipeline(pipeline=pipe)
        self.embeddings = HuggingFaceEmbeddings(
            model_name=EMBEDDING_MODEL,
            model_kwargs={'device': 'cpu'}
        )
        self.base_db = self.load_base_knowledge()

    def load_base_knowledge(self) -> Optional[FAISS]:
        try:
            if not os.path.exists(DOCS_DIR):
                os.makedirs(DOCS_DIR)
                return None

            loader = DirectoryLoader(
                DOCS_DIR,
                glob="**/*.pdf",
                loader_cls=PyPDFLoader
            )
            documents = loader.load()
            
            if not documents:
                return None
            
            text_splitter = RecursiveCharacterTextSplitter(
                chunk_size=500,
                chunk_overlap=100,
                length_function=len,
                separators=["\n\n", "\n", ".", " ", ""]
            )
            texts = text_splitter.split_documents(documents)
            
            return FAISS.from_documents(texts, self.embeddings)
            
        except Exception as e:
            print(f"Erro ao carregar base de conhecimento: {str(e)}")
            return None

    def process_pdf(self, file_content: bytes) -> Optional[FAISS]:
        try:
            with tempfile.NamedTemporaryFile(delete=False, suffix='.pdf') as tmp_file:
                tmp_file.write(file_content)
                tmp_path = tmp_file.name

            loader = PyPDFLoader(tmp_path)
            documents = loader.load()
            os.unlink(tmp_path)
            
            if not documents:
                return None
            
            text_splitter = RecursiveCharacterTextSplitter(
                chunk_size=500,
                chunk_overlap=100,
                length_function=len,
                separators=["\n\n", "\n", ".", " ", ""]
            )
            texts = text_splitter.split_documents(documents)
            
            db = FAISS.from_documents(texts, self.embeddings)
            
            if self.base_db is not None:
                db.merge_from(self.base_db)
            
            return db
            
        except Exception as e:
            print(f"Erro ao processar PDF: {str(e)}")
            return None

    def generate_response(self, file_obj, query: str, progress=gr.Progress()) -> Tuple[str, str, str]:
        """Retorna (resposta, status, tempo_decorrido)"""
        if not query.strip():
            return "Por favor, insira uma pergunta.", "⚠️ Aguardando pergunta", "0s"
        
        start_time = time.time()
        try:
            progress(0, desc="Iniciando processamento...")
            
            # Processa documento
            progress(0.2, desc="Processando documento...")
            if file_obj is not None:
                db = self.process_pdf(file_obj)
                if db is None:
                    return "Não foi possível processar o PDF.", "❌ Erro no processamento", "0s"
            elif self.base_db is not None:
                db = self.base_db
            else:
                return "Nenhuma base de conhecimento disponível.", "❌ Sem documentos", "0s"
            
            progress(0.4, desc="Buscando informações relevantes...")
            qa_chain = RetrievalQA.from_chain_type(
                llm=self.llm,
                chain_type="stuff",
                retriever=db.as_retriever(
                    search_kwargs={"k": 4, "fetch_k": 6}
                ),
                return_source_documents=True
            )
            
            progress(0.6, desc="Gerando resposta...")
            prompt = f"""Baseado nos documentos fornecidos, responda em português à seguinte pergunta:
            {query}
            
            Se a resposta vier da base de documentos permanente, indique isso no início.
            Se a resposta vier do PDF enviado, indique isso no início.
            Se não encontrar informações suficientes, indique isso claramente."""
            
            result = qa_chain({"query": prompt})
            elapsed_time = f"{time.time() - start_time:.1f}s"
            
            progress(1.0, desc="Concluído!")
            return result["result"], "✅ Sucesso", elapsed_time
            
        except Exception as e:
            elapsed_time = f"{time.time() - start_time:.1f}s"
            return f"Erro ao gerar resposta: {str(e)}", "❌ Erro", elapsed_time

def create_demo():
    rag = RAGSystem()
    
    with gr.Blocks(theme=gr.themes.Soft()) as demo:
        with gr.Column(elem_id="container"):
            # Cabeçalho
            gr.Markdown(
                """
                # 🤖 Assistente de Documentos Inteligente
                
                Este sistema usa tecnologia RAG (Retrieval-Augmented Generation) para responder perguntas sobre seus documentos.
                """
            )
            
            # Área principal
            with gr.Row():
                # Coluna de entrada
                with gr.Column():
                    with gr.Group():
                        gr.Markdown("### 📄 Documentos")
                        file_input = gr.File(
                            label="Upload de PDF (opcional)",
                            type="binary",
                            file_types=[".pdf"],
                            height=100,
                        )
                        info = gr.Markdown(
                            f"""
                            ℹ️ Além do upload, o sistema também consulta a pasta `{DOCS_DIR}`
                            """
                        )
                    
                    with gr.Group():
                        gr.Markdown("### ❓ Sua Pergunta")
                        query_input = gr.Textbox(
                            placeholder="Digite sua pergunta aqui...",
                            lines=3,
                            max_lines=6,
                            show_label=False,
                        )
                    
                    with gr.Row():
                        clear_btn = gr.Button("🗑️ Limpar", variant="secondary")
                        submit_btn = gr.Button("🔍 Enviar Pergunta", variant="primary")
                
                # Coluna de saída
                with gr.Column():
                    with gr.Group():
                        gr.Markdown("### 📝 Resposta")
                        with gr.Row():
                            status_output = gr.Textbox(
                                label="Status",
                                value="⏳ Aguardando...",
                                interactive=False,
                                show_label=False,
                            )
                            time_output = gr.Textbox(
                                label="Tempo",
                                value="0s",
                                interactive=False,
                                show_label=False,
                            )
                        
                        response_output = gr.Textbox(
                            label="Resposta",
                            placeholder="A resposta aparecerá aqui...",
                            interactive=False,
                            lines=12,
                            show_label=False,
                        )
            
            # Exemplos
            with gr.Accordion("📚 Exemplos de Perguntas", open=False):
                gr.Examples(
                    examples=[
                        [None, "Qual é o tema principal dos documentos?"],
                        [None, "Pode resumir os pontos principais?"],
                        [None, "Quais são as principais conclusões?"],
                        [None, "Explique o contexto deste documento."],
                    ],
                    inputs=[file_input, query_input],
                )
            
            # Rodapé
            gr.Markdown(
                """
                ---
                ### 🔧 Sobre o Sistema
                * Usa modelo T5 para geração de respostas
                * Processamento de documentos com tecnologia RAG
                * Suporte a múltiplos documentos PDF
                * Respostas baseadas apenas no conteúdo dos documentos
                """
            )
        
        # Eventos
        submit_btn.click(
            fn=rag.generate_response,
            inputs=[file_input, query_input],
            outputs=[response_output, status_output, time_output],
        )
        
        clear_btn.click(
            lambda: (None, "", "⏳ Aguardando...", "0s"),
            outputs=[file_input, query_input, status_output, time_output],
        )
        
        # Limpa a resposta quando a pergunta muda
        query_input.change(
            lambda: ("", "⏳ Aguardando...", "0s"),
            outputs=[response_output, status_output, time_output],
        )
    
    return demo

if __name__ == "__main__":
    demo = create_demo()
    demo.launch()