app.py

"""
Credit to Derek Thomas, [email protected]
"""
import os
import logging
from pathlib import Path
from time import perf_counter

import gradio as gr
from jinja2 import Environment, FileSystemLoader

from backend.query_llm import generate_hf, generate_openai
from backend.semantic_search import retrieve
import itertools

emb_models = ["bge", "minilm"]
splitters = ['ct', 'rct', 'nltk']
chunk_sizes = ["500", "2000"]
sub_vectors = ["8", "16", "32"]

# Create all combinations of the provided arrays
combinations = itertools.product(emb_models, splitters, chunk_sizes, sub_vectors)

TOP_K = int(os.getenv("TOP_K", 4))

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')


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, api_kind,
            num_docs, model_kind, sub_vector_size, chunk_size, splitter_type):
    query = history[-1][0]

    if not query:
        raise gr.Warning("Please submit a non-empty string as a prompt")

    logger.info('Retrieving documents...')
    # Retrieve documents relevant to query
    document_start = perf_counter()

    documents = retrieve(query, int(num_docs), model_kind, sub_vector_size, chunk_size, splitter_type)

    document_time = perf_counter() - document_start
    logger.info(f'Finished Retrieving documents in {round(document_time, 2)} seconds...')

    # Create Prompt
    prompt = template.render(documents=documents, query=query)
    prompt_html = template_html.render(documents=documents, query=query)


    if api_kind == "HuggingFace":
         generate_fn = generate_hf
    elif api_kind == "OpenAI":
         generate_fn = generate_openai
    else:
         raise gr.Error(f"API {api_kind} is not supported")
    

    history[-1][1] = ""
    # for character in generate_fn(prompt, history[:-1]):
    #     history[-1][1] = character
    #     yield history, prompt_html
    
    for model_name, doc, size, sub_vector in combinations:
        documents_i = retrieve(query, int(num_docs), model_name, sub_vector, size, doc)
        prompt_i = template.render(documents=documents_i, query=query)
        prompt_html = template_html.render(documents=documents, query=query)
        
        hist_chunk = ""
        prev_hist = history[-1][1] + f"\nmodel {model_name}, splitter {doc}, size {size}, sub vector {sub_vector}\n"
        for character in generate_fn(prompt_i, history[:-1]):
            hist_chunk = character
            history[-1][1] = prev_hist + hist_chunk
            yield history, prompt_html


with gr.Blocks() as demo:
    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 text and press enter",
                container=False,
                )
        txt_btn = gr.Button(value="Submit text", scale=1)


    with gr.Row():
        num_docs = gr.Slider(1, 20, label="number of docs", step=1, value=4)
        model_kind = gr.Radio(choices=emb_models, value="bge", label="embedding model")
        sub_vector_size = gr.Radio(choices=sub_vectors, value="32", label="sub-vector size")
    with gr.Row():
        api_kind = gr.Radio(choices=["HuggingFace", "OpenAI"], value="HuggingFace", label="Chat model engine")
        chunk_size = gr.Radio(choices=chunk_sizes, value="2000", label="chunk size")
        splitter_type = gr.Radio(choices=splitters, value="nltk", label="splitter")

    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, api_kind,
            num_docs, model_kind, sub_vector_size, chunk_size, splitter_type
            ], [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, api_kind,
            num_docs, model_kind, sub_vector_size, chunk_size, splitter_type
            ], [chatbot, prompt_html])

    # Turn it back on
    txt_msg.then(lambda: gr.Textbox(interactive=True), None, [txt], queue=False)

demo.queue()
demo.launch(debug=True)