Update app.py

app.py

@@ -1,60 +1,53 @@
 import gradio as gr
 import os
-
 from langchain_community.document_loaders import PyPDFLoader
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import Chroma
 from langchain.chains import ConversationalRetrievalChain
-from langchain_community.embeddings import HuggingFaceEmbeddings 
-from langchain_community.llms import HuggingFacePipeline
-from langchain.chains import ConversationChain
-from langchain.memory import ConversationBufferMemory
+from langchain_community.embeddings import HuggingFaceEmbeddings
 from langchain_community.llms import HuggingFaceEndpoint
-
+from langchain.memory import ConversationBufferMemory
 from pathlib import Path
 import chromadb
 from unidecode import unidecode
-
-from transformers import AutoTokenizer
-import transformers
-import torch
-import tqdm 
-import accelerate
 import re
 
-# default_persist_directory = './chroma_HF/'
+# 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]
 
-# Load PDF document and create doc splits
+# 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)
+        chunk_size=chunk_size,
+        chunk_overlap=chunk_overlap
+    )
     doc_splits = text_splitter.split_documents(pages)
     return doc_splits
 
-# Create vector database
+# Função para criar o banco de dados vetorial
 def create_db(splits, collection_name):
     embedding = HuggingFaceEmbeddings()
-    new_client = chromadb.EphemeralClient()
+    # Usando PersistentClient para persistir o banco de dados
+    new_client = chromadb.PersistentClient(path="./chroma_db")
     vectordb = Chroma.from_documents(
         documents=splits,
         embedding=embedding,
@@ -63,65 +56,58 @@ def create_db(splits, collection_name):
     )
     return vectordb
 
-# Load vector database
-def load_db():
-    embedding = HuggingFaceEmbeddings()
-    vectordb = Chroma(
-        embedding_function=embedding)
-    return vectordb
-
-# Initialize langchain LLM chain
+# 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 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,
+            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"]:
+    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",
+            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,
+            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,
+            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
     )
-    retriever=vector_db.as_retriever()
+    retriever = vector_db.as_retriever()
     progress(0.8, desc="Definindo cadeia de recuperação...")
     qa_chain = ConversationalRetrievalChain.from_llm(
         llm,
         retriever=retriever,
-        chain_type="stuff", 
+        chain_type="stuff",
         memory=memory,
         return_source_documents=True,
         verbose=False,
@@ -129,10 +115,10 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
     progress(0.9, desc="Concluído!")
     return qa_chain
 
-# Generate collection name for vector database
+# Função para gerar um nome de coleção válido
 def create_collection_name(filepath):
     collection_name = Path(filepath).stem
-    collection_name = collection_name.replace(" ","-") 
+    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]
@@ -146,7 +132,7 @@ def create_collection_name(filepath):
     print('Nome da coleção: ', collection_name)
     return collection_name
 
-# Initialize database
+# 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...")
@@ -158,12 +144,14 @@ def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Pr
     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)
+    print("Nome do LLM: ", llm_name)
     qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
     return qa_chain, "Completo!"
 
+# 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:
@@ -171,6 +159,7 @@ def format_chat_history(message, chat_history):
         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 = format_chat_history(message, history)
     response = qa_chain({"question": message, "chat_history": formatted_chat_history})
@@ -187,6 +176,7 @@ def conversation(qa_chain, message, history):
     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
 
+# Função para carregar arquivos
 def upload_file(file_obj):
     list_file_path = []
     for idx, file in enumerate(file_obj):
@@ -194,52 +184,54 @@ def upload_file(file_obj):
         list_file_path.append(file_path)
     return list_file_path
 
+# Interface Gradio
 def demo():
     with gr.Blocks(theme=gr.themes.Default(primary_hue="blue", secondary_hue="gray")) as demo:
         vector_db = gr.State()
         qa_chain = gr.State()
         collection_name = gr.State()
-        
+
         gr.Markdown(
-        """<center><h2>Chatbot baseado em PDF</center></h2>
-        <h3>Faça perguntas sobre seus documentos PDF</h3>""")
+            """<center><h2>Chatbot baseado em PDF</center></h2>
+            <h3>Faça perguntas sobre seus documentos PDF</h3>"""
+        )
         gr.Markdown(
-        """<b>Nota:</b> Este assistente AI, usando Langchain e LLMs de código aberto, realiza geração aumentada por recuperação (RAG) a partir de seus documentos PDF. \
-        A interface do usuário explicitamente mostra múltiplos passos 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 de documentos para maior clareza.<br>
-        <br><b>Aviso:</b> Este espaço usa o hardware básico gratuito da Hugging Face. Alguns passos e modelos LLM usados abaixo (endpoints de inferência gratuitos) podem levar algum tempo para gerar uma resposta.
-        """)
-        
+            """<b>Nota:</b> Este assistente AI, usando Langchain e LLMs de código aberto, realiza geração aumentada por recuperação (RAG) a partir de seus documentos PDF. \
+            A interface do usuário explicitamente mostra múltiplos passos 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 de documentos para maior clareza.<br>
+            <br><b>Aviso:</b> Este espaço usa o hardware básico gratuito da Hugging Face. Alguns passos e modelos LLM usados abaixo (endpoints de inferência gratuitos) podem levar algum tempo para gerar uma resposta.
+            """
+        )
+
         with gr.Tab("Passo 1 - Carregar 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)")
-        
+
         with gr.Tab("Passo 2 - Processar documento"):
             with gr.Row():
-                db_btn = gr.Radio(["ChromaDB"], label="Tipo de banco de dados vetorial", value = "ChromaDB", type="index", info="Escolha seu banco de dados vetorial")
+                db_btn = gr.Radio(["ChromaDB"], label="Tipo de banco de dados vetorial", value="ChromaDB", type="index", info="Escolha seu 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 chunk", info="Tamanho do chunk", interactive=True)
+                    slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Tamanho do chunk", info="Tamanho do chunk", interactive=True)
                 with gr.Row():
-                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Sobreposição do chunk", info="Sobreposição do chunk", interactive=True)
+                    slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Sobreposição do chunk", info="Sobreposição do chunk", 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("Passo 3 - Inicializar cadeia de QA"):
             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")
+                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)
+                    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)
+                    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)
+                    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 de QA")
+                llm_progress = gr.Textbox(value="Nenhum", label="Inicialização da cadeia de QA")
             with gr.Row():
                 qachain_btn = gr.Button("Inicializar cadeia de Perguntas e Respostas")
 
@@ -260,31 +252,17 @@ def demo():
             with gr.Row():
                 submit_btn = gr.Button("Enviar mensagem")
                 clear_btn = gr.ClearButton([msg, chatbot], value="Limpar conversa")
-            
-        # Preprocessing events
-        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)
 
-        # Chatbot events
-        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)
+        # Eventos de pré-processamento
+        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__":