import gradio as gr import os from langchain_community.vectorstores import FAISS from langchain_community.document_loaders import PyPDFLoader from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain.chains import ConversationalRetrievalChain from langchain_community.embeddings import HuggingFaceEmbeddings from langchain_community.llms import HuggingFacePipeline from langchain.memory import ConversationBufferMemory import torch from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline # Lista de modelos públicos e leves list_llm = ["EleutherAI/gpt-neo-125m", "distilbert/distilgpt2"] list_llm_simple = [os.path.basename(llm) for llm in list_llm] # Carregar e dividir documento PDF def load_doc(list_file_path): loaders = [PyPDFLoader(x) for x in list_file_path] pages = [] for loader in loaders: pages.extend(loader.load()) text_splitter = RecursiveCharacterTextSplitter( chunk_size=512, # Reduzido para acelerar a busca chunk_overlap=32 # Menor sobreposição para menos processamento ) doc_splits = text_splitter.split_documents(pages) return doc_splits # Criar banco de vetores def create_db(splits): embeddings = HuggingFaceEmbeddings() vectordb = FAISS.from_documents(splits, embeddings) return vectordb # Inicializar o chain LLM local def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): tokenizer = AutoTokenizer.from_pretrained(llm_model) model = AutoModelForCausalLM.from_pretrained( llm_model, device_map="auto", # Usa GPU se disponível torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, # Otimiza para GPU trust_remote_code=True ) # Pipeline otimizado pipe = pipeline( "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=max_tokens, temperature=temperature, top_k=top_k, do_sample=False, # Greedy decoding para mais velocidade repetition_penalty=1.1, return_full_text=False ) llm = HuggingFacePipeline(pipeline=pipe) memory = ConversationBufferMemory( memory_key="chat_history", output_key="answer", return_messages=True ) retriever = vector_db.as_retriever(search_kwargs={"k": 2}) # Reduzir número de documentos retornados qa_chain = ConversationalRetrievalChain.from_llm( llm, retriever=retriever, chain_type="stuff", memory=memory, return_source_documents=True, verbose=False, ) return qa_chain # Inicializar banco de dados def initialize_database(list_file_obj, progress=gr.Progress()): list_file_path = [x.name for x in list_file_obj if x is not None] doc_splits = load_doc(list_file_path) vector_db = create_db(doc_splits) return vector_db, "Database created!" # Inicializar LLM def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): llm_name = list_llm[llm_option] print("llm_name: ", llm_name) qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress) return qa_chain, "QA chain initialized. Chatbot is ready!" def format_chat_history(message, chat_history): formatted_chat_history = [] for user_message, bot_message in chat_history: formatted_chat_history.append(f"User: {user_message}") formatted_chat_history.append(f"Assistant: {bot_message}") return formatted_chat_history def conversation(qa_chain, message, history): formatted_chat_history = format_chat_history(message, history) response = qa_chain.invoke({"question": message, "chat_history": formatted_chat_history}) response_answer = response["answer"] if response_answer.find("Helpful Answer:") != -1: response_answer = response_answer.split("Helpful Answer:")[-1] response_sources = response["source_documents"] response_source1 = response_sources[0].page_content.strip() response_source2 = response_sources[1].page_content.strip() response_source3 = "" # Menos referências para acelerar response_source1_page = response_sources[0].metadata["page"] + 1 response_source2_page = response_sources[1].metadata["page"] + 1 response_source3_page = 0 new_history = history + [(message, response_answer)] return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page def upload_file(file_obj): list_file_path = [] for idx, file in enumerate(file_obj): file_path = file_obj.name list_file_path.append(file_path) return list_file_path def demo(): with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo: vector_db = gr.State() qa_chain = gr.State() gr.HTML("