Spaces:
Build error
Build error
| import logging | |
| from functools import partial | |
| from pathlib import Path | |
| from time import perf_counter | |
| import gradio as gr | |
| from jinja2 import Environment, FileSystemLoader | |
| from transformers import AutoTokenizer | |
| from backend.query_llm import check_endpoint_status, generate | |
| from backend.semantic_search import retriever | |
| proj_dir = Path(__file__).parent | |
| # Setting up the logging | |
| logging.basicConfig(level=logging.INFO) | |
| logger = logging.getLogger(__name__) | |
| # Set up the template environment with the templates directory | |
| env = Environment(loader=FileSystemLoader(proj_dir / 'templates')) | |
| # Load the templates directly from the environment | |
| template = env.get_template('template.j2') | |
| template_html = env.get_template('template_html.j2') | |
| # Initialize tokenizer | |
| tokenizer = AutoTokenizer.from_pretrained('derek-thomas/jais-13b-chat-hf') | |
| # Examples | |
| examples = ['ู ู ูุงู ุทุฑูู ู ุนุฑูุฉ ุงูุชููู ุงูุจุญุฑูุฉุ', | |
| 'ูู ุงูุณู ุงุก ุฒุฑูุงุกุ', | |
| "ู ู ูุงุฒ ุจูุฃุณ ุงูุนุงูู ููุฑุฌุงู ูู ุนุงู 2014ุ",] | |
| def add_text(history, text): | |
| history = [] if history is None else history | |
| history = history + [(text, None)] | |
| return history, gr.Textbox(value="", interactive=False) | |
| def bot(history, hyde=False): | |
| top_k = 5 | |
| query = history[-1][0] | |
| logger.warning('Retrieving documents...') | |
| # Retrieve documents relevant to query | |
| document_start = perf_counter() | |
| if hyde: | |
| hyde_document = generate(f"Write a wikipedia article intro paragraph to answer this query: {query}").split('### Response: [|AI|]')[-1] | |
| logger.warning(hyde_document) | |
| documents = retriever(hyde_document, top_k=top_k) | |
| else: | |
| documents = retriever(query, top_k=top_k) | |
| document_time = perf_counter() - document_start | |
| logger.warning(f'Finished Retrieving documents in {round(document_time, 2)} seconds...') | |
| # Function to count tokens | |
| def count_tokens(text): | |
| return len(tokenizer.encode(text)) | |
| # Create Prompt | |
| prompt = template.render(documents=documents, query=query) | |
| # Check if the prompt is too long | |
| token_count = count_tokens(prompt) | |
| while token_count > 2048: | |
| # Shorten your documents here. This is just a placeholder for the logic you'd use. | |
| documents.pop() # Remove the last document | |
| prompt = template.render(documents=documents, query=query) # Re-render the prompt | |
| token_count = count_tokens(prompt) # Re-count tokens | |
| prompt_html = template_html.render(documents=documents, query=query) | |
| history[-1][1] = "" | |
| response = generate(prompt) | |
| history[-1][1] = response.split('### Response: [|AI|]')[-1] | |
| return history, prompt_html | |
| intro_md = """ | |
| # Arabic RAG | |
| This is a project to demonstrate Retreiver Augmented Generation (RAG) in Arabic and English. It uses | |
| [Arabic Wikipedia](https://ar.wikipedia.org/wiki) as a base to answer questions you have. | |
| A retriever ([sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2/discussions/8)) | |
| will find the articles relevant to your query and include them in a prompt so the reader ([core42/jais-13b-chat](https://huggingface.co/core42/jais-13b-chat)) | |
| can then answer your questions on it. | |
| You can see the prompt clearly displayed below the chatbot to understand what is going to the LLM. | |
| # Read this if you get an error | |
| I'm using Inference Endpoint's Scale to Zero to save money on GPUs. If the staus shows its not "Running" send a | |
| chat to wake it up. You will get a `500 error` and it will take ~7 min to wake up. | |
| """ | |
| with gr.Blocks() as demo: | |
| gr.Markdown(intro_md) | |
| endpoint_status = gr.Textbox(check_endpoint_status, label="Inference Endpoint Status", every=1) | |
| with gr.Tab("Arabic-RAG"): | |
| chatbot = gr.Chatbot( | |
| [], | |
| elem_id="chatbot", | |
| avatar_images=('https://aui.atlassian.com/aui/8.8/docs/images/avatar-person.svg', | |
| 'https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg'), | |
| bubble_full_width=False, | |
| show_copy_button=True, | |
| show_share_button=True, | |
| ) | |
| with gr.Row(): | |
| txt = gr.Textbox( | |
| scale=3, | |
| show_label=False, | |
| placeholder="Enter query in Arabic or English and press enter", | |
| container=False, | |
| ) | |
| txt_btn = gr.Button(value="Submit text", scale=1) | |
| gr.Examples(examples, txt) | |
| prompt_html = gr.HTML() | |
| # Turn off interactivity while generating if you click | |
| txt_msg = txt_btn.click(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( | |
| bot, chatbot, [chatbot, prompt_html]) | |
| # Turn it back on | |
| txt_msg.then(lambda: gr.Textbox(interactive=True), None, [txt], queue=False) | |
| # Turn off interactivity while generating if you hit enter | |
| txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then( | |
| bot, chatbot, [chatbot, prompt_html]) | |
| # Turn it back on | |
| txt_msg.then(lambda: gr.Textbox(interactive=True), None, [txt], queue=False) | |
| with gr.Tab("Arabic-RAG + HyDE"): | |
| hyde_chatbot = gr.Chatbot( | |
| [], | |
| elem_id="chatbot", | |
| avatar_images=('https://aui.atlassian.com/aui/8.8/docs/images/avatar-person.svg', | |
| 'https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg'), | |
| bubble_full_width=False, | |
| show_copy_button=True, | |
| show_share_button=True, | |
| ) | |
| with gr.Row(): | |
| hyde_txt = gr.Textbox( | |
| scale=3, | |
| show_label=False, | |
| placeholder="Enter text and press enter", | |
| container=False, | |
| ) | |
| hyde_txt_btn = gr.Button(value="Submit text", scale=1) | |
| gr.Examples(examples, hyde_txt) | |
| hyde_prompt_html = gr.HTML() | |
| # Turn off interactivity while generating if you click | |
| hyde_txt_msg = hyde_txt_btn.click(add_text, [hyde_chatbot, hyde_txt], [hyde_chatbot, hyde_txt], | |
| queue=False).then( | |
| partial(bot, hyde=True), [hyde_chatbot], [hyde_chatbot, hyde_prompt_html]) | |
| # Turn it back on | |
| hyde_txt_msg.then(lambda: gr.Textbox(interactive=True), None, [hyde_txt], queue=False) | |
| # Turn off interactivity while generating if you hit enter | |
| hyde_txt_msg = hyde_txt.submit(add_text, [hyde_chatbot, hyde_txt], [hyde_chatbot, hyde_txt], queue=False).then( | |
| partial(bot, hyde=True), [hyde_chatbot], [hyde_chatbot, hyde_prompt_html]) | |
| # Turn it back on | |
| hyde_txt_msg.then(lambda: gr.Textbox(interactive=True), None, [hyde_txt], queue=False) | |
| demo.queue() | |
| demo.launch(debug=True) | |