Update app.py

app.py

@@ -12,120 +12,273 @@ import chromadb
 from unidecode import unidecode
 import re
 
-# Modelos LLM disponíveis
+# Lista de modelos LLM disponíveis
 list_llm = [
-    "mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.1",
-    "google/gemma-7b-it", "google/gemma-2b-it", "HuggingFaceH4/zephyr-7b-beta", "HuggingFaceH4/zephyr-7b-gemma-v0.1",
-    "meta-llama/Llama-2-7b-chat-hf", "microsoft/phi-2", "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "mosaicml/mpt-7b-instruct",
-    "tiiuae/falcon-7b-instruct", "google/flan-t5-xxl"
+    "mistralai/Mistral-7B-Instruct-v0.2",
+    "mistralai/Mixtral-8x7B-Instruct-v0.1",
+    "mistralai/Mistral-7B-Instruct-v0.1",
+    "google/gemma-7b-it",
+    "google/gemma-2b-it",
+    "HuggingFaceH4/zephyr-7b-beta",
+    "HuggingFaceH4/zephyr-7b-gemma-v0.1",
+    "meta-llama/Llama-2-7b-chat-hf",
+    "microsoft/phi-2",
+    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+    "mosaicml/mpt-7b-instruct",
+    "tiiuae/falcon-7b-instruct",
+    "google/flan-t5-xxl"
 ]
 list_llm_simple = [os.path.basename(llm) for llm in list_llm]
 
-# Função de carregamento e divisão de documentos
-def load_and_split_documents(list_file_path, chunk_size, chunk_overlap):
+# Função para carregar documentos PDF e dividir em chunks
+def load_doc(list_file_path, chunk_size, chunk_overlap):
     loaders = [PyPDFLoader(x) for x in list_file_path]
     pages = []
     for loader in loaders:
         pages.extend(loader.load())
-    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
-    return text_splitter.split_documents(pages)
+    text_splitter = RecursiveCharacterTextSplitter(
+        chunk_size=chunk_size,
+        chunk_overlap=chunk_overlap
+    )
+    doc_splits = text_splitter.split_documents(pages)
+    return doc_splits
 
-# Função para criar banco de dados vetorial com ChromaDB
-def create_vector_db(splits, collection_name):
+# Função para criar o banco de dados vetorial
+def create_db(splits, collection_name):
     embedding = HuggingFaceEmbeddings()
+    # Usando PersistentClient para persistir o banco de dados
     new_client = chromadb.PersistentClient(path="./chroma_db")
-    return Chroma.from_documents(documents=splits, embedding=embedding, client=new_client, collection_name=collection_name)
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=new_client,
+        collection_name=collection_name,
+    )
+    return vectordb
 
-# Função para inicializar a cadeia de QA
+# Função para inicializar a cadeia de QA com o modelo LLM
 def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
-    progress(0.1, desc="Inicializando tokenizer e Hub...")
-    llm = HuggingFaceEndpoint(
-        repo_id=llm_model, temperature=temperature, max_new_tokens=max_tokens, top_k=top_k, load_in_8bit=True
+    progress(0.1, desc="Inicializando tokenizer da HF...")
+    progress(0.5, desc="Inicializando Hub da HF...")
+    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            temperature=temperature,
+            max_new_tokens=max_tokens,
+            top_k=top_k,
+            load_in_8bit=True,
+        )
+    elif llm_model in ["HuggingFaceH4/zephyr-7b-gemma-v0.1", "mosaicml/mpt-7b-instruct"]:
+        raise gr.Error("O modelo LLM é muito grande para ser carregado automaticamente no endpoint de inferência gratuito")
+    elif llm_model == "microsoft/phi-2":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            temperature=temperature,
+            max_new_tokens=max_tokens,
+            top_k=top_k,
+            trust_remote_code=True,
+            torch_dtype="auto",
+        )
+    elif llm_model == "TinyLlama/TinyLlama-1.1B-Chat-v1.0":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            temperature=temperature,
+            max_new_tokens=250,
+            top_k=top_k,
+        )
+    elif llm_model == "meta-llama/Llama-2-7b-chat-hf":
+        raise gr.Error("O modelo Llama-2-7b-chat-hf requer uma assinatura Pro...")
+    else:
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            temperature=temperature,
+            max_new_tokens=max_tokens,
+            top_k=top_k,
+        )
+
+    progress(0.75, desc="Definindo memória de buffer...")
+    memory = ConversationBufferMemory(
+        memory_key="chat_history",
+        output_key='answer',
+        return_messages=True
     )
-    progress(0.5, desc="Definindo memória de buffer e cadeia de recuperação...")
-    memory = ConversationBufferMemory(memory_key="chat_history", output_key='answer', return_messages=True)
     retriever = vector_db.as_retriever()
-    qa_chain = ConversationalRetrievalChain.from_llm(llm, retriever=retriever, chain_type="stuff", memory=memory, return_source_documents=True)
+    progress(0.8, desc="Definindo cadeia de recuperação...")
+    qa_chain = ConversationalRetrievalChain.from_llm(
+        llm,
+        retriever=retriever,
+        chain_type="stuff",
+        memory=memory,
+        return_source_documents=True,
+        verbose=False,
+    )
     progress(0.9, desc="Concluído!")
     return qa_chain
 
 # Função para gerar um nome de coleção válido
 def create_collection_name(filepath):
     collection_name = Path(filepath).stem
-    collection_name = unidecode(collection_name.replace(" ", "-"))
-    return re.sub('[^A-Za-z0-9]+', '-', collection_name)[:50]
+    collection_name = collection_name.replace(" ", "-")
+    collection_name = unidecode(collection_name)
+    collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)
+    collection_name = collection_name[:50]
+    if len(collection_name) < 3:
+        collection_name = collection_name + 'xyz'
+    if not collection_name[0].isalnum():
+        collection_name = 'A' + collection_name[1:]
+    if not collection_name[-1].isalnum():
+        collection_name = collection_name[:-1] + 'Z'
+    print('Caminho do arquivo: ', filepath)
+    print('Nome da coleção: ', collection_name)
+    return collection_name
 
-# Função para inicializar o banco de dados e o modelo LLM
-def initialize_database_and_llm(list_file_obj, chunk_size, chunk_overlap, llm_option, llm_temperature, max_tokens, top_k, progress=gr.Progress()):
+# Função para inicializar o banco de dados
+def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
     list_file_path = [x.name for x in list_file_obj if x is not None]
     progress(0.1, desc="Criando nome da coleção...")
     collection_name = create_collection_name(list_file_path[0])
-    progress(0.25, desc="Carregando e dividindo documentos...")
-    doc_splits = load_and_split_documents(list_file_path, chunk_size, chunk_overlap)
+    progress(0.25, desc="Carregando documento...")
+    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
     progress(0.5, desc="Gerando banco de dados vetorial...")
-    vector_db = create_vector_db(doc_splits, collection_name)
-    progress(0.75, desc="Inicializando modelo LLM...")
+    vector_db = create_db(doc_splits, collection_name)
+    progress(0.9, desc="Concluído!")
+    return vector_db, collection_name, "Completo!"
+
+# Função para inicializar o modelo LLM
+def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
     llm_name = list_llm[llm_option]
+    print("Nome do LLM: ", llm_name)
     qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
-    progress(0.9, desc="Concluído!")
-    return vector_db, collection_name, qa_chain
+    return qa_chain, "Completo!"
 
-# Função de interação com o chatbot
+# Função para formatar o histórico de conversa
+def format_chat_history(message, chat_history):
+    formatted_chat_history = []
+    for user_message, bot_message in chat_history:
+        formatted_chat_history.append(f"Usuário: {user_message}")
+        formatted_chat_history.append(f"Assistente: {bot_message}")
+    return formatted_chat_history
+
+# Função para realizar a conversa com o chatbot
 def conversation(qa_chain, message, history):
-    formatted_chat_history = [f"Usuário: {user_message}\nAssistente: {bot_message}" for user_message, bot_message in history]
+    formatted_chat_history = format_chat_history(message, history)
     response = qa_chain({"question": message, "chat_history": formatted_chat_history})
-    response_answer = response["answer"].split("Resposta útil:")[-1]
-    response_sources = [doc.page_content.strip() for doc in response["source_documents"]]
-    response_pages = [doc.metadata["page"] + 1 for doc in response["source_documents"]]
+    response_answer = response["answer"]
+    if response_answer.find("Resposta útil:") != -1:
+        response_answer = response_answer.split("Resposta útil:")[-1]
+    response_sources = response["source_documents"]
+    response_source1 = response_sources[0].page_content.strip()
+    response_source2 = response_sources[1].page_content.strip()
+    response_source3 = response_sources[2].page_content.strip()
+    response_source1_page = response_sources[0].metadata["page"] + 1
+    response_source2_page = response_sources[1].metadata["page"] + 1
+    response_source3_page = response_sources[2].metadata["page"] + 1
     new_history = history + [(message, response_answer)]
-    return qa_chain, gr.update(value=""), new_history, *response_sources, *response_pages
+    return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
 
-# Função de carregamento de arquivos
+# Função para carregar arquivos
 def upload_file(file_obj):
-    return [file_obj.name for file_obj in file_obj if file_obj is not None]
+    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
 
-# Interface Gradio
 def demo():
     with gr.Blocks(theme="base") as demo:
-        vector_db, qa_chain, collection_name = gr.State(), gr.State(), gr.State()
-        gr.Markdown("<center><h2>Chatbot baseado em PDF</center></h2><h3>Faça qualquer pergunta sobre seus documentos PDF</h3>")
-
+        vector_db = gr.State()
+        qa_chain = gr.State()
+        collection_name = gr.State()
+        
+        gr.Markdown(
+        """<center><h2>Chatbot baseado em PDF</center></h2>
+        <h3>Faça qualquer pergunta sobre seus documentos PDF</h3>""")
+        gr.Markdown(
+        """<b>Nota:</b> Este assistente de IA, utilizando Langchain e LLMs de código aberto, realiza geração aumentada por recuperação (RAG) a partir de seus documentos PDF. \
+        A interface do usuário mostra explicitamente várias etapas para ajudar a entender o fluxo de trabalho do RAG. 
+        Este chatbot leva em consideração perguntas anteriores ao gerar respostas (via memória conversacional), e inclui referências documentais para maior clareza.<br>
+        <br><b>Aviso:</b> Este espaço usa a CPU básica gratuita do Hugging Face. Algumas etapas e modelos LLM utilizados abaixo (pontos finais de inferência gratuitos) podem levar algum tempo para gerar uma resposta.
+        """)
+        
         with gr.Tab("Etapa 1 - Carregar PDF"):
-            document = gr.Files(height=100, file_count="multiple", file_types=["pdf"])
-
+            with gr.Row():
+                document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Carregue seus documentos PDF (único ou múltiplos)")
+                # upload_btn = gr.UploadButton("Carregando documento...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
+        
         with gr.Tab("Etapa 2 - Processar documento"):
-            db_btn = gr.Button("Gerar banco de dados vetorial")
-            slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Tamanho do bloco")
-            slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Sobreposição do bloco")
-            db_progress = gr.Textbox(label="Inicialização do banco de dados vetorial")
-
+            with gr.Row():
+                db_btn = gr.Radio(["ChromaDB"], label="Tipo de banco de dados vetorial", value = "ChromaDB", type="index", info="Escolha o banco de dados vetorial")
+            with gr.Accordion("Opções avançadas - Divisor de texto do documento", open=False):
+                with gr.Row():
+                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Tamanho do bloco", info="Tamanho do bloco", interactive=True)
+                with gr.Row():
+                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Sobreposição do bloco", info="Sobreposição do bloco", interactive=True)
+            with gr.Row():
+                db_progress = gr.Textbox(label="Inicialização do banco de dados vetorial", value="Nenhum")
+            with gr.Row():
+                db_btn = gr.Button("Gerar banco de dados vetorial")
+            
         with gr.Tab("Etapa 3 - Inicializar cadeia de QA"):
-            llm_btn = gr.Radio(list_llm_simple, label="Modelos LLM")
-            slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.7, step=0.1, label="Temperatura")
-            slider_maxtokens = gr.Slider(minimum=224, maximum=4096, value=1024, step=32, label="Máximo de Tokens")
-            slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="Amostras top-k")
-            llm_progress = gr.Textbox(value="Nenhum", label="Inicialização da cadeia QA")
-            qachain_btn = gr.Button("Inicializar cadeia de Pergunta e Resposta")
+            with gr.Row():
+                llm_btn = gr.Radio(list_llm_simple, \
+                    label="Modelos LLM", value = list_llm_simple[0], type="index", info="Escolha seu modelo LLM")
+            with gr.Accordion("Opções avançadas - Modelo LLM", open=False):
+                with gr.Row():
+                    slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.7, step=0.1, label="Temperatura", info="Temperatura do modelo", interactive=True)
+                with gr.Row():
+                    slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Máximo de Tokens", info="Máximo de tokens do modelo", interactive=True)
+                with gr.Row():
+                    slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="Amostras top-k", info="Amostras top-k do modelo", interactive=True)
+            with gr.Row():
+                llm_progress = gr.Textbox(value="Nenhum",label="Inicialização da cadeia QA")
+            with gr.Row():
+                qachain_btn = gr.Button("Inicializar cadeia de Pergunta e Resposta")
 
         with gr.Tab("Etapa 4 - Chatbot"):
             chatbot = gr.Chatbot(height=300)
-            doc_source1, doc_source2, doc_source3 = gr.Textbox(label="Referência 1"), gr.Textbox(label="Referência 2"), gr.Textbox(label="Referência 3")
-            source1_page, source2_page, source3_page = gr.Number(label="Página 1"), gr.Number(label="Página 2"), gr.Number(label="Página 3")
+            with gr.Accordion("Avançado - Referências do documento", open=False):
+                with gr.Row():
+                    doc_source1 = gr.Textbox(label="Referência 1", lines=2, container=True, scale=20)
+                    source1_page = gr.Number(label="Página", scale=1)
+                with gr.Row():
+                    doc_source2 = gr.Textbox(label="Referência 2", lines=2, container=True, scale=20)
+                    source2_page = gr.Number(label="Página", scale=1)
+                with gr.Row():
+                    doc_source3 = gr.Textbox(label="Referência 3", lines=2, container=True, scale=20)
+                    source3_page = gr.Number(label="Página", scale=1)
+            with gr.Row():
+                msg = gr.Textbox(placeholder="Digite a mensagem (exemplo: 'Sobre o que é este documento?')", container=True)
+            with gr.Row():
+                submit_btn = gr.Button("Enviar mensagem")
+                clear_btn = gr.ClearButton([msg, chatbot], value="Limpar conversa")
             
-            # Campo de texto para enviar mensagens
-            user_input = gr.Textbox(label="Sua mensagem")
-
-            # Implementação de lógica de interação de conversa
-            def send_message(message, history, qa_chain):
-                formatted_chat_history = [f"Usuário: {user_message}\nAssistente: {bot_message}" for user_message, bot_message in history]
-                response = qa_chain({"question": message, "chat_history": formatted_chat_history})
-                response_answer = response["answer"].split("Resposta útil:")[-1]
-                response_sources = [doc.page_content.strip() for doc in response["source_documents"]]
-                response_pages = [doc.metadata["page"] + 1 for doc in response["source_documents"]]
-                new_history = history + [(message, response_answer)]
-                return qa_chain, gr.update(value=""), new_history, *response_sources, *response_pages
-
-            user_input.submit(send_message, inputs=[user_input, chatbot.history, qa_chain], outputs=[qa_chain, gr.update(value=""), chatbot.history, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page])
-        
-    demo.launch()
+        # Eventos de pré-processamento
+        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
+        db_btn.click(initialize_database, \
+            inputs=[document, slider_chunk_size, slider_chunk_overlap], \
+            outputs=[vector_db, collection_name, db_progress])
+        qachain_btn.click(initialize_LLM, \
+            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
+            outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+
+        # Eventos do Chatbot
+        msg.submit(conversation, \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+        submit_btn.click(conversation, \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+        clear_btn.click(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+    demo.queue().launch(debug=True)
+
 
+if __name__ == "__main__":
+    demo()