Create rag_functions.py

rag_functions.py

@@ -0,0 +1,193 @@
+# rag_functions.py
+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 HuggingFaceEndpoint
+from langchain.memory import ConversationBufferMemory
+from pathlib import Path
+import chromadb
+from unidecode import unidecode
+import re
+
+# 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"
+]
+list_llm_simple = [os.path.basename(llm) for llm in list_llm]
+
+# 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
+    )
+    doc_splits = text_splitter.split_documents(pages)
+    return doc_splits
+
+# 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")
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=new_client,
+        collection_name=collection_name,
+    )
+    return vectordb
+
+# Função para inicializar a cadeia de QA com o modelo LLM
+def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=None):
+    if 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,
+        )
+    elif llm_model in ["HuggingFaceH4/zephyr-7b-gemma-v0.1", "mosaicml/mpt-7b-instruct"]:
+        raise ValueError("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 ValueError("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,
+        )
+
+    if progress:
+        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()
+    if progress:
+        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,
+    )
+    if progress:
+        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 = 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
+def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=None):
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    if progress:
+        progress(0.1, desc="Criando nome da coleção...")
+    collection_name = create_collection_name(list_file_path[0])
+    if progress:
+        progress(0.25, desc="Carregando documento...")
+    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
+    if progress:
+        progress(0.5, desc="Gerando banco de dados vetorial...")
+    vector_db = create_db(doc_splits, collection_name)
+    if progress:
+        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=None):
+    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)
+    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:
+        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 = format_chat_history(message, history)
+    response = qa_chain({"question": message, "chat_history": formatted_chat_history})
+    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, "", 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):
+        file_path = file_obj.name
+        list_file_path.append(file_path)
+    return list_file_path