import gradio as gr import os api_token = os.getenv("HF_TOKEN") from langchain_community.vectorstores import FAISS 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.llms import HuggingFaceEndpoint import torch 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] def load_doc(list_file_path): loaders = [PyPDFLoader(x) for x in list_file_path] pages = [] for loader in loaders: pages.extend(loader.load()) text_splitter = RecursiveCharacterTextSplitter( chunk_size = 1024, chunk_overlap = 64 ) doc_splits = text_splitter.split_documents(pages) return doc_splits def create_db(splits): embeddings = HuggingFaceEmbeddings() vectordb = FAISS.from_documents(splits, embeddings) return vectordb def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): if llm_model == "meta-llama/Meta-Llama-3-8B-Instruct": llm = HuggingFaceEndpoint( repo_id=llm_model, huggingfacehub_api_token = api_token, temperature = temperature, max_new_tokens = max_tokens, top_k = top_k, ) else: llm = HuggingFaceEndpoint( huggingfacehub_api_token = api_token, repo_id=llm_model, temperature = temperature, max_new_tokens = max_tokens, top_k = top_k, ) memory = ConversationBufferMemory( memory_key="chat_history", output_key='answer', return_messages=True ) retriever=vector_db.as_retriever() qa_chain = ConversationalRetrievalChain.from_llm( llm, retriever=retriever, chain_type="stuff", memory=memory, return_source_documents=True, verbose=False, ) return qa_chain 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, "Database created!" def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): 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, "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"] if response_answer.find("Helpful Answer:") != -1: response_answer = response_answer.split("Helpful Answer:")[-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, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_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) return list_file_path def demo(): custom_css = """ #column-container { display: flex !important; flex-direction: row !important; flex-wrap: wrap !important; width: 100% !important; max-width: 100% !important; } #column-left { min-width: 250px !important; flex: 1 1 30% !important; max-width: 100% !important; box-sizing: border-box !important; } #column-right { min-width: 250px !important; flex: 1 1 65% !important; max-width: 100% !important; box-sizing: border-box !important; overflow-x: visible !important; } .chatbot { max-width: 100% !important; } .textbox { max-width: 100% !important; } @media (max-width: 800px) { #column-left, #column-right { flex: 1 1 100% !important; } } """ with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky"), css=custom_css) as demo: vector_db = gr.State() qa_chain = gr.State() gr.HTML("

RAG PDF chatbot

") gr.Markdown("""Query your PDF documents! This AI agent is designed to perform retrieval augmented generation (RAG) on PDF documents.""") with gr.Row(elem_id="column-container"): with gr.Column(elem_id="column-left"): gr.Markdown("Step 1 - Upload PDF documents and Initialize RAG pipeline") document = gr.Files(height=300, file_count="multiple", file_types=[".pdf"], interactive=True, label="Upload PDF documents") db_btn = gr.Button("Create vector database") db_progress = gr.Textbox(value="Not initialized", show_label=False) gr.Markdown("Select Large Language Model (LLM) and input parameters") llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index") with gr.Accordion("LLM input parameters", open=False): slider_temperature = gr.Slider(0.01, 1.0, value=0.5, step=0.1, label="Temperature") slider_maxtokens = gr.Slider(128, 9192, value=4096, step=128, label="Max New Tokens") slider_topk = gr.Slider(1, 10, value=3, step=1, label="top-k") qachain_btn = gr.Button("Initialize Question Answering Chatbot") llm_progress = gr.Textbox(value="Not initialized", show_label=False) with gr.Column(elem_id="column-right"): gr.Markdown("Step 2 - Chat with your Document") chatbot = gr.Chatbot(height=505) with gr.Accordion("Relevant context from the source document", open=False): with gr.Row(): doc_source1 = gr.Textbox(label="Reference 1", lines=2, scale=20) source1_page = gr.Number(label="Page", scale=1) with gr.Row(): doc_source2 = gr.Textbox(label="Reference 2", lines=2, scale=20) source2_page = gr.Number(label="Page", scale=1) with gr.Row(): doc_source3 = gr.Textbox(label="Reference 3", lines=2, scale=20) source3_page = gr.Number(label="Page", scale=1) msg = gr.Textbox(placeholder="Ask a question") with gr.Row(): submit_btn = gr.Button("Submit") clear_btn = gr.ClearButton([msg, chatbot], value="Clear") 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], inputs=None, outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], queue=False) 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__": demo()