import gradio as gr import os from langchain_community.document_loaders import PyPDFLoader from langchain_community.vectorstores import Chroma from langchain_community.embeddings import HuggingFaceEmbeddings from langchain_community.llms import HuggingFacePipeline from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain.chains import ConversationalRetrievalChain from langchain.chains import ConversationChain from langchain.memory import ConversationBufferMemory from langchain_community.llms import HuggingFaceHub from pathlib import Path import chromadb from transformers import AutoTokenizer import transformers import torch import tqdm import accelerate default_persist_directory = './ChromaDB' llm_name0 = "mistralai/Mixtral-8x7B-Instruct-v0.1" list_llm = [llm_name0] list_llm_simple = [os.path.basename(llm) for llm in list_llm] # Load PDF document and create doc splits def load_doc(list_file_path, chunk_size, chunk_overlap): # Processing for one document only # loader = PyPDFLoader(file_path) # pages = loader.load() loaders = [PyPDFLoader(x) for x in list_file_path] pages = [] for loader in loaders: pages.extend(loader.load()) # text_splitter = RecursiveCharacterTextSplitter(chunk_size = 600, chunk_overlap = 50) text_splitter = RecursiveCharacterTextSplitter( chunk_size = chunk_size, chunk_overlap = chunk_overlap) doc_splits = text_splitter.split_documents(pages) return doc_splits # Create vector database def create_db(splits, collection_name): embedding = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2") new_client = chromadb.EphemeralClient() vectordb = Chroma.from_documents(documents=splits, embedding=embedding, persist_directory="./chroma_db", client=new_client, collection_name=collection_name) return vectordb # Load vector database def load_db(): embedding = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2") vectordb = Chroma(persist_directory=default_persist_directory, embedding_function=embedding) return vectordb # Initialize langchain LLM chain def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): progress(0.1, desc="Initializing HF tokenizer...") # HuggingFaceHub uses HF inference endpoints progress(0.5, desc="Initializing HF Hub...") if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1": llm = HuggingFaceHub( repo_id=llm_model, model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True} ) progress(0.75, desc="Defining buffer memory...") memory = ConversationBufferMemory( memory_key="chat_history", output_key='answer', return_messages=True ) # retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3}) retriever=vector_db.as_retriever() progress(0.8, desc="Defining retrieval chain...") qa_chain = ConversationalRetrievalChain.from_llm( llm, retriever=retriever, chain_type="stuff", memory=memory, # combine_docs_chain_kwargs={"prompt": your_prompt}) return_source_documents=True, # return_generated_question=True, # verbose=True, ) progress(0.9, desc="Done!") return qa_chain # Initialize database def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()): CreaDB = True if CreaDB: list_file_path = [x.name for x in list_file_obj if x is not None] collection_name = Path(list_file_path[0]).stem print('list_file_path: ', list_file_path) print('Collection name: ', collection_name) progress(0.25, desc="Loading document...") # Load document and create splits doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap) # Create or load Vector database progress(0.5, desc="Generating vector database...") # global vector_db vector_db = create_db(doc_splits, collection_name) progress(0.9, desc="Done!") return vector_db, collection_name, "Complete!" else: collection_name = 'Documenti' vector_db = load_db() return vector_db, collection_name, "Complete!" def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): # print("llm_option",llm_option) 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, "Complete!" 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) #print("formatted_chat_history",formatted_chat_history) formatted_chat_history = "" # Generate response using QA chain response = qa_chain({"question": message, "chat_history": formatted_chat_history}) response_answer = response["answer"] response_sources = response["source_documents"] response_source1 = response_sources[0].page_content.strip() response_source2 = response_sources[1].page_content.strip() # Langchain sources are zero-based response_source1_page = response_sources[0].metadata["page"] + 1 response_source2_page = response_sources[1].metadata["page"] + 1 print ('Response: ', response) # Append user message and response to chat history new_history = history + [(message, response_answer)] # return gr.update(value=""), new_history, response_sources[0], response_sources[1] return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_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) # print(file_path) # initialize_database(file_path, progress) return list_file_path def demo(): with gr.Blocks(theme="base") as demo: vector_db = gr.State() qa_chain = gr.State() collection_name = gr.State() gr.Markdown( """