Spaces:
Sleeping
Sleeping
| import gradio as gr | |
| 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 | |
| list_llm = [ | |
| "mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-Instruct-v0.1", | |
| "google/gemma-7b-it", "google/gemma-2b-it", | |
| "HuggingFaceH4/zephyr-7b-beta", "HuggingFaceH4/zephyr-7b-gemma-v0.1", | |
| "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "tiiuae/falcon-7b-instruct" | |
| ] | |
| list_llm_simple = [os.path.basename(llm) for llm in list_llm] | |
| 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) | |
| return text_splitter.split_documents(pages) | |
| def create_db(splits, collection_name): | |
| embedding = HuggingFaceEmbeddings() | |
| new_client = chromadb.EphemeralClient() | |
| return Chroma.from_documents(documents=splits, embedding=embedding, client=new_client, collection_name=collection_name) | |
| def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()): | |
| progress(0.5, desc="Initializing HF Hub...") | |
| llm = HuggingFaceEndpoint( | |
| 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() | |
| return ConversationalRetrievalChain.from_llm( | |
| llm, | |
| retriever=retriever, | |
| chain_type="stuff", | |
| memory=memory, | |
| return_source_documents=True, | |
| verbose=False, | |
| ) | |
| def create_collection_name(filepath): | |
| collection_name = Path(filepath).stem | |
| collection_name = unidecode(collection_name.replace(" ", "-")) | |
| collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)[:50] | |
| if len(collection_name) < 3: | |
| 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' | |
| return collection_name | |
| def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()): | |
| list_file_path = [x.name for x in list_file_obj if x is not None] | |
| collection_name = create_collection_name(list_file_path[0]) | |
| doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap) | |
| vector_db = create_db(doc_splits, collection_name) | |
| return vector_db, collection_name, "Completed!" | |
| 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, "Completed!" | |
| def format_chat_history(message, chat_history): | |
| return [f"User: {user_message}\nAssistant: {bot_message}" for user_message, bot_message in chat_history] | |
| 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"].split("Helpful Answer:")[-1] | |
| response_sources = response["source_documents"] | |
| sources = [(source.page_content.strip(), source.metadata["page"] + 1) for source in response_sources[:3]] | |
| new_history = history + [(message, response_answer)] | |
| return qa_chain, gr.update(value=""), new_history, *[item for source in sources for item in source] | |
| def demo(): | |
| with gr.Blocks(theme="base") as demo: | |
| vector_db = gr.State() | |
| qa_chain = gr.State() | |
| collection_name = gr.State() | |
| gr.Markdown("# PDF-based Chatbot Creator") | |
| with gr.Tab("Step 1 - Upload PDFs"): | |
| document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents") | |
| with gr.Tab("Step 2 - Process Documents"): | |
| db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value="ChromaDB", type="index") | |
| with gr.Accordion("Advanced Options - Document text splitter", open=False): | |
| slider_chunk_size = gr.Slider(100, 1000, 1000, step=20, label="Chunk size") | |
| slider_chunk_overlap = gr.Slider(10, 200, 100, step=10, label="Chunk overlap") | |
| db_progress = gr.Textbox(label="Vector database initialization", value="None") | |
| db_btn = gr.Button("Generate vector database") | |
| with gr.Tab("Step 3 - Initialize QA chain"): | |
| llm_btn = gr.Radio(list_llm_simple, label="LLM models", value=list_llm_simple[5], type="index") | |
| with gr.Accordion("Advanced options - LLM model", open=False): | |
| slider_temperature = gr.Slider(0.01, 1.0, 0.3, step=0.1, label="Temperature") | |
| slider_maxtokens = gr.Slider(224, 4096, 1024, step=32, label="Max Tokens") | |
| slider_topk = gr.Slider(1, 10, 3, step=1, label="top-k samples") | |
| language_btn = gr.Radio(["Italian", "English"], label="Language", value="Italian", type="index") | |
| llm_progress = gr.Textbox(value="None", label="QA chain initialization") | |
| qachain_btn = gr.Button("Initialize Question Answering chain") | |
| with gr.Tab("Step 4 - Chatbot"): | |
| chatbot = gr.Chatbot(height=300) | |
| with gr.Accordion("Advanced options - Document references", open=False): | |
| doc_sources = [gr.Textbox(label=f"Reference {i+1}", lines=2, container=True, scale=20) for i in range(3)] | |
| source_pages = [gr.Number(label="Page", scale=1) for _ in range(3)] | |
| msg = gr.Textbox(placeholder="Enter message (e.g., 'What is this document about?')", container=True) | |
| submit_btn = gr.Button("Send message") | |
| clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation") | |
| db_btn.click(initialize_database, inputs=[document, slider_chunk_size, slider_chunk_overlap], outputs=[vector_db, collection_name, db_progress]) | |
| qachain_btn.click(initialize_LLM, inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], outputs=[qa_chain, llm_progress]) | |
| submit_btn.click(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot] + doc_sources + source_pages) | |
| msg.submit(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot] + doc_sources + source_pages) | |
| demo.queue().launch(debug=True) | |
| if __name__ == "__main__": | |
| demo() |