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_community.embeddings import HuggingFaceEmbeddings from langchain_community.llms import HuggingFaceEndpoint from langchain.chains import ConversationalRetrievalChain from langchain.memory import ConversationBufferMemory api_token = os.getenv("HF_TOKEN") DEFAULT_LLM = "meta-llama/Meta-Llama-3-8B-Instruct" def load_and_create_db(files): if not files: return None, None try: # Create a list of documents list_file_paths = [] for file in files: # Save uploaded file temporarily file_name = file.name if file_name.lower().endswith('.pdf'): list_file_paths.append(file_name) else: raise ValueError(f"Unsupported file format: {file_name}. Please upload PDF files only.") if not list_file_paths: return None, None # Load documents loaders = [PyPDFLoader(path) for path in list_file_paths] pages = [] for loader in loaders: pages.extend(loader.load()) # Split documents text_splitter = RecursiveCharacterTextSplitter( chunk_size=1024, chunk_overlap=64 ) doc_splits = text_splitter.split_documents(pages) # Create vector database embeddings = HuggingFaceEmbeddings() vectordb = FAISS.from_documents(doc_splits, embeddings) # Initialize QA chain qa_chain = initialize_llmchain(vectordb) return vectordb, qa_chain except Exception as e: print(f"Error processing files: {str(e)}") return None, None def initialize_llmchain(vector_db, temperature=0.5, max_tokens=4096, top_k=3): llm = HuggingFaceEndpoint( repo_id=DEFAULT_LLM, huggingfacehub_api_token=api_token, temperature=temperature, max_new_tokens=max_tokens, top_k=top_k, ) memory = ConversationBufferMemory( memory_key="chat_history", output_key='answer', return_messages=True ) qa_chain = ConversationalRetrievalChain.from_llm( llm, retriever=vector_db.as_retriever(), chain_type="stuff", memory=memory, return_source_documents=True, verbose=False, ) return qa_chain def format_citation(source_doc): content = source_doc.page_content.strip() page = source_doc.metadata["page"] + 1 return content, page def conversation(qa_chain, message, history): if not qa_chain: return (None, gr.update(value=""), history, "", 0, "", 0, "", 0, "Please upload a document first.") formatted_history = [] for user_msg, bot_msg in history: formatted_history.append(f"User: {user_msg}") formatted_history.append(f"Assistant: {bot_msg}") response = qa_chain.invoke({ "question": message, "chat_history": formatted_history }) answer = response["answer"] if "Helpful Answer:" in answer: answer = answer.split("Helpful Answer:")[-1] # Format answer with citation numbers sources = response["source_documents"][:3] modified_answer = answer for i in range(len(sources)): modified_answer = modified_answer + f" [{i+1}]" # Get citation contents and page numbers citations = [format_citation(source) for source in sources] source1_content, page1 = citations[0] if len(citations) > 0 else ("", 0) source2_content, page2 = citations[1] if len(citations) > 1 else ("", 0) source3_content, page3 = citations[2] if len(citations) > 2 else ("", 0) new_history = history + [(message, modified_answer)] return (qa_chain, gr.update(value=""), new_history, source1_content, page1, source2_content, page2, source3_content, page3, "") 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("