import gradio as gr import os api_token = os.getenv("HF_TOKEN") from langchain_community.vectorstores import FAISS from langchain_community.document_loaders import TextLoader from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain.chains import ConversationalRetrievalChain from langchain_community.embeddings import HuggingFaceEmbeddings from langchain_community.llms import HuggingFaceEndpoint from langchain.memory import ConversationBufferMemory from langchain.prompts import PromptTemplate list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"] list_llm_simple = [os.path.basename(llm) for llm in list_llm] # Custom prompt template CUSTOM_PROMPT_TEMPLATE = """ **Response Instructions:** - Write a detailed, coherent, and insightful article that fully addresses the query based on the provided context. - Adhere to the following principles: 1. **Define the Core Subject**: Introduce and build the discussion logically around the main topic. 2. **Establish Connections**: Highlight relationships between ideas and concepts with reasoning and examples. 3. **Elaborate on Key Points**: Provide in-depth explanations and emphasize the significance of concepts. 4. **Maintain Objectivity**: Use only the context provided, avoiding speculation or external knowledge. 5. **Ensure Structure and Clarity**: Present information sequentially for a smooth narrative flow. 6. **Engage with Content**: Explore implicit meanings, resolve doubts, and address counterpoints logically. 7. **Provide Examples and Insights**: Use examples to clarify abstract ideas and offer actionable steps if applicable. 8. **Logical Depth**: Draw inferences, explain purposes, and refute opposing ideas when necessary. Context: {context} Question: {question} Chat History: {chat_history} Craft the response as a seamless, thorough, and authoritative explanation that naturally integrates all aspects of the query. """ # Load and split text documents def load_doc(list_file_path): pages = [] for file_path in list_file_path: if file_path.endswith('.txt'): loader = TextLoader(file_path) pages.extend(loader.load()) text_splitter = RecursiveCharacterTextSplitter( chunk_size=1024, chunk_overlap=64 ) doc_splits = text_splitter.split_documents(pages) return doc_splits # Create vector database def create_db(splits): embeddings = HuggingFaceEmbeddings() vectordb = FAISS.from_documents(splits, embeddings) return vectordb # Initialize langchain LLM chain with custom prompt def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): llm = HuggingFaceEndpoint( repo_id=llm_model, 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 ) # Create custom prompt custom_prompt = PromptTemplate( template=CUSTOM_PROMPT_TEMPLATE, input_variables=["context", "question", "chat_history"] ) retriever = vector_db.as_retriever() qa_chain = ConversationalRetrievalChain.from_llm( llm, retriever=retriever, chain_type="stuff", memory=memory, return_source_documents=True, verbose=False, combine_docs_chain_kwargs={"prompt": custom_prompt} ) return qa_chain # Initialize database def initialize_database(list_file_obj, progress=gr.Progress()): list_file_path = [x.name for x in list_file_obj if x is not None] doc_splits = load_doc(list_file_path) vector_db = create_db(doc_splits) return vector_db, "Text database created!" # Initialize LLM def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): llm_name = list_llm[llm_option] qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress) return qa_chain, "QA chain initialized. Chatbot is ready!" 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) response = qa_chain.invoke({"question": message, "chat_history": formatted_chat_history}) response_answer = response["answer"] response_sources = response["source_documents"] # Get sources (with fallback for when there are fewer than 3 sources) sources_content = [] sources_pages = [] for i in range(3): if i < len(response_sources): sources_content.append(response_sources[i].page_content.strip()) sources_pages.append(0) # For text files, we don't have page numbers else: sources_content.append("") sources_pages.append(0) new_history = history + [(message, response_answer)] return (qa_chain, gr.update(value=""), new_history, sources_content[0], sources_pages[0], sources_content[1], sources_pages[1], sources_content[2], sources_pages[2]) 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("

RAG Text Document Chatbot

") gr.Markdown("""Query your text documents! This AI agent performs retrieval augmented generation (RAG) on TXT documents. Please do not upload confidential documents. """) with gr.Row(): with gr.Column(scale=86): gr.Markdown("Step 1 - Upload Text Files and Initialize RAG pipeline") with gr.Row(): document = gr.Files(height=300, file_count="multiple", file_types=["txt"], interactive=True, label="Upload TXT documents") with gr.Row(): db_btn = gr.Button("Create text database") with gr.Row(): db_progress = gr.Textbox(value="Not initialized", show_label=False) gr.Markdown("Select Large Language Model (LLM) and input parameters") with gr.Row(): llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index") with gr.Row(): with gr.Accordion("LLM input parameters", open=False): with gr.Row(): slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.5, step=0.1, label="Temperature") with gr.Row(): slider_maxtokens = gr.Slider(minimum=128, maximum=9192, value=4096, step=128, label="Max New Tokens") with gr.Row(): slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k") with gr.Row(): qachain_btn = gr.Button("Initialize Question Answering Chatbot") with gr.Row(): llm_progress = gr.Textbox(value="Not initialized", show_label=False) with gr.Column(scale=200): gr.Markdown("Step 2 - Chat with your Document") chatbot = gr.Chatbot(height=505) with gr.Accordion("Relevant context from the source document", open=False): for i in range(1, 4): with gr.Row(): doc_source = gr.Textbox(label=f"Reference {i}", lines=2, container=True, scale=20) source_page = gr.Number(label="Line Range", scale=1, visible=False) with gr.Row(): msg = gr.Textbox(placeholder="Ask a question", container=True) with gr.Row(): submit_btn = gr.Button("Submit") clear_btn = gr.ClearButton([msg, chatbot], value="Clear") # Event handlers db_btn.click(initialize_database, inputs=[document], outputs=[vector_db, 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], outputs=[chatbot] + [doc for i in range(1,4) for doc in [globals()[f"doc_source{i}"], globals()[f"source_page{i}"]]], queue=False) msg.submit(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot] + [globals()[f"doc_source{i}"] for i in range(1,4)] + [globals()[f"source_page{i}"] for i in range(1,4)]], queue=False) submit_btn.click(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot] + [globals()[f"doc_source{i}"] for i in range(1,4)] + [globals()[f"source_page{i}"] for i in range(1,4)]], queue=False) clear_btn.click(lambda: [None, "", 0, "", 0, "", 0], outputs=[chatbot] + [doc for i in range(1,4) for doc in [globals()[f"doc_source{i}"], globals()[f"source_page{i}"]]], queue=False) demo.queue().launch(debug=True) if __name__ == "__main__": demo()