app.py

import gradio as gr
import os
import torch

from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain.embeddings import HuggingFaceEmbeddings
from langchain_huggingface import HuggingFacePipeline
from langchain.memory import ConversationBufferMemory

from transformers import AutoTokenizer, pipeline

# ===================================================================
# CONFIGURAÇÃO RADICAL DE HARDWARE
# ===================================================================
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
TORCH_DTYPE = torch.bfloat16 if DEVICE == "cuda" else torch.float32
MAX_MEMORY = "16GB" if DEVICE == "cpu" else None

# ===================================================================
# LISTA DE MODELOS OTIMIZADOS
# ===================================================================
LLM_MODELS = {
    "TinyLlama-1.1B-Chat": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    "Phi-2": "microsoft/phi-2",
    "Mistral-7B-Instruct": "mistralai/Mistral-7B-Instruct-v0.2",
    "Zephyr-7B-Beta": "HuggingFaceH4/zephyr-7b-beta"
}

# ===================================================================
# NÚCLEO DO SISTEMA
# ===================================================================
class DocumentProcessor:
    @staticmethod
    def load_and_split(files, chunk_size=512, chunk_overlap=64):
        """Carrega e processa documentos com paralelismo extremo"""
        try:
            loaders = [PyPDFLoader(file.name) for file in files]
            return [page for loader in loaders for page in loader.load_and_split(
                RecursiveCharacterTextSplitter(
                    chunk_size=chunk_size,
                    chunk_overlap=chunk_overlap,
                    separators=["\n\n", "\n", " ", ""]
                )
            )]
        except Exception as e:
            raise RuntimeError(f"FALHA CRÍTICA NO PROCESSAMENTO: {str(e)}")

class VectorDBManager:
    @staticmethod
    def create(splits):
        """Cria vetorização com aceleração de hardware"""
        return Chroma.from_documents(
            documents=splits,
            embedding=HuggingFaceEmbeddings(),
            persist_directory="./chroma_db"
        )

class LLMEngine:
    @staticmethod
    def initialize(model_name, temp=0.7, max_tokens=512):
        """Inicialização agressiva do modelo com otimizações de baixo nível"""
        try:
            tokenizer = AutoTokenizer.from_pretrained(LLM_MODELS[model_name])
            
            pipe = pipeline(
                "text-generation",
                model=LLM_MODELS[model_name],
                tokenizer=tokenizer,
                device=DEVICE,
                torch_dtype=TORCH_DTYPE,
                max_new_tokens=max_tokens,
                do_sample=True,
                top_k=50,
                temperature=temp,
                model_kwargs={"load_in_4bit": True} if "cuda" in DEVICE else {}
            )
            
            return HuggingFacePipeline(pipeline=pipe)
        except KeyError:
            raise ValueError("MODELO NÃO SUPORTADO!")
        except Exception as e:
            raise RuntimeError(f"FALHA NUCLEAR NO MODELO: {str(e)}")

# ===================================================================
# INTERFACE DE COMBATE
# ===================================================================
def create_war_interface():
    with gr.Blocks(theme=gr.themes.Soft(), title="⚔️ PDF Assault v1.0") as warzone:
        state = gr.State({"db": None, "llm": None})
        
        # Zona de Upload
        with gr.Row(variant="panel"):
            file_upload = gr.Files(label="DOCUMENTOS ALVO", file_types=[".pdf"])
            process_btn = gr.Button("ATAQUE!", variant="stop")
        
        # Controles Táticos
        with gr.Row(variant="compact"):
            model_selector = gr.Dropdown(list(LLM_MODELS.keys()), label="ARMA PRINCIPAL", value="TinyLlama-1.1B-Chat")
            temp_slider = gr.Slider(0, 1, 0.7, label="POTÊNCIA DE FOGO")
            deploy_btn = gr.Button("DEPLOY MODELO", variant="primary")
        
        # Campo de Batalha
        chatbot = gr.Chatbot(height=600, label="ZONA DE OPERAÇÕES")
        msg_input = gr.Textbox(label="COMANDO DE ATAQUE", placeholder="Insira o alvo...")
        
        # Sistema de Logs
        combat_log = gr.Textbox(label="RELATÓRIO DE COMBATE", interactive=False)

        # ===== Operações Militares =====
        @process_btn.click(inputs=[file_upload], outputs=[state, combat_log])
        def assault_documents(files):
            try:
                splits = DocumentProcessor.load_and_split(files)
                db = VectorDBManager.create(splits)
                return {"db": db, "llm": None}, "✅ DOCUMENTOS CAPTURADOS!"
            except Exception as e:
                return state.value, f"☠️ FALHA CATACLÍSMICA: {str(e)}"

        @deploy_btn.click(inputs=[model_selector, temp_slider, state], outputs=[state, combat_log])
        def deploy_model(model, temp, current_state):
            try:
                llm = LLMEngine.initialize(model, temp)
                current_state["llm"] = ConversationalRetrievalChain.from_llm(
                    llm=llm,
                    retriever=current_state["db"].as_retriever(),
                    memory=ConversationBufferMemory(memory_key="chat_history", return_messages=True),
                    return_source_documents=True
                )
                return current_state, f"🚀 {model} PRONTO PARA COMBATE!"
            except Exception as e:
                return current_state, f"💥 FALHA NO DEPLOY: {str(e)}"

        @msg_input.submit(inputs=[msg_input, chatbot, state], outputs=[msg_input, chatbot])
        def execute_combat(command, history, state):
            if not state["llm"]:
                return command, history + [(command, "⚠️ MODELO NÃO DEPLOYADO!")]
            
            try:
                result = state["llm"]({"question": command, "chat_history": history})
                response = f"🎯 RESPOSTA:\n{result['answer']}\n\n"
                response += "📌 INTEL:\n" + "\n".join(
                    f"Página {doc.metadata['page']+1}: {doc.page_content[:75]}..." 
                    for doc in result["source_documents"][:3]
                )
                return "", history + [(command, response)]
            except Exception as e:
                return command, history + [(command, f"☢️ FALHA CRÍTICA: {str(e)}")]

    return warzone

# ===================================================================
# INICIALIZAÇÃO DO SISTEMA
# ===================================================================
if __name__ == "__main__":
    interface = create_war_interface()
    interface.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        auth=("admin", "combat123"),
        show_error=True
    )