app.py

import gradio as gr
import pandas as pd
from thirdai import bolt, licensing
import os
import time


licensing.set_path("license.serialized")
max_posts = 5
df = pd.read_csv("processed_recipes_3.csv")
model = bolt.UniversalDeepTransformer.load("1bn_name_ctg_keywords_4gram.bolt")

recipe_id_to_row_num = {}

for i in range(df.shape[0]):
    recipe_id_to_row_num[df.iloc[i,0]] = i


INTRO_MARKDOWN = (
    """# A billion parameter model, trained on a single CPU, in just 90 mins, on 522K recipes from food.com !!
    """
)

LIKE_TEXT = "👍 update LLM"
FEEDBACK_RECEIVED_TEXT = "👌 Click search for updated results"
SHOW_MORE = "Show more"
SHOW_LESS = "Show less"


def retrain(query, doc_id):
    query = query.lower()
    query.replace('\n', ' ')
    query = ' '.join([query[i:i+4] for i in range(len(query)-3)])
    df = pd.DataFrame({
        "Name": [query],
        "RecipeId": [str(doc_id)]
    })

    filename = f"temptrain{hash(query)}{hash(doc_id)}{time.time()}.csv"

    df.to_csv(filename)

    prediction = None

    while prediction != doc_id:
        model.train(filename, epochs=1)
        prediction = model.predict(
            {"Name": query.replace('\n', ' ')},
            return_predicted_class=True)

    os.remove(filename)

    # sample = {"query": query.replace('\n', ' '), "id": str(doc_id)}
    # batch = [sample]

    # prediction = None

    # while prediction != doc_id:
    #     model.train_batch(batch, metrics=["categorical_accuracy"])
    #     prediction = model.predict(sample, return_predicted_class=True)


def search(query):
    query = query.lower()
    query = ' '.join([query[i:i+4] for i in range(len(query)-3)])
    scores = model.predict({"Name": query})
    ####
    sorted_ids = scores.argsort()[-max_posts:][::-1]
    relevant_posts = [
        df.iloc[pid] for pid in sorted_ids
    ]
    ####
    # K = min(2*max_posts, len(scores) - 1)
    # sorted_post_ids = scores.argsort()[-K:][::-1]
    # print(sorted_post_ids)
    # sorted_ids = []
    # relevant_posts = []
    # count = 0
    # for pid in sorted_post_ids:
    #     if pid in recipe_id_to_row_num:
    #         relevant_posts.append(df.iloc[recipe_id_to_row_num[pid]])
    #         sorted_ids.append(pid)
    #     count += 1
    #     if count==max_posts:
    #         break
    ####
    header = [gr.Markdown.update(visible=True)]
    boxes = [
        gr.Box.update(visible=True)
        for _ in relevant_posts
    ]
    titles = [
        gr.Markdown.update(f"## {post['Name']}")
        for post in relevant_posts
    ]
    toggles = [
        gr.Button.update(
            visible=True,
            value=SHOW_MORE,
            interactive=True,
        )
        for _ in relevant_posts
    ]
    matches = [
        gr.Button.update(
            value=LIKE_TEXT,
            interactive=True,
        )
        for _ in relevant_posts
    ]
    bodies = [
        gr.HTML.update(
            visible=False,
            value=f"<br/>"
            f"<h2>Description:</h2>\n{post['Description']}\n\n"
            "<hr class='solid'>"
            f"<h2>Ingredients:</h2>\n{post['RecipeIngredientParts']}\n\n"
            "<br/>"
            f"<h2>Instructions:</h2>\n{post['RecipeInstructions']}\n\n"
            "<br/>")
        for post in relevant_posts
    ]

    return (
        header +
        boxes +
        titles +
        toggles +
        matches +
        bodies +
        [sorted_ids]
    )


def handle_toggle(toggle):
    if toggle == SHOW_MORE:
        new_toggle_text = SHOW_LESS
        visible = True
    if toggle == SHOW_LESS:
        new_toggle_text = SHOW_MORE
        visible = False
    return [
        gr.Button.update(new_toggle_text),
        gr.HTML.update(visible=visible),
    ]


def handle_feedback(button_id: int):
    def register_feedback(doc_ids, query):
        retrain(
            query=query,
            doc_id=doc_ids[button_id]
        )
        return gr.Button.update(
            value=FEEDBACK_RECEIVED_TEXT,
            interactive=False,
        )

    return register_feedback


default_query = (
    "biryani lamb spicy contains cloves and red chili powder, made with ghee and hard boiled eggs, made by grinding coconut and cashew"
)


with gr.Blocks() as demo:
    gr.Markdown(INTRO_MARKDOWN)
    query = gr.Textbox(value=default_query, label="Query", lines=10)
    submit = gr.Button(value="Search")

    header = [gr.Markdown("# Relevant Recipes", visible=False)]
    post_boxes = []
    post_titles = []
    toggle_buttons = []
    match_buttons = []
    post_bodies = []
    post_ids = gr.State([])

    for i in range(max_posts):
        with gr.Box(visible=False) as box:
            post_boxes.append(box)

            with gr.Row():
                with gr.Column(scale=5):
                    title = gr.Markdown("")
                    post_titles.append(title)
                with gr.Column(scale=1, min_width=370):
                    with gr.Row():
                        with gr.Column(scale=3, min_width=170):
                            toggle = gr.Button(SHOW_MORE)
                            toggle_buttons.append(toggle)
                        with gr.Column(scale=1, min_width=170):
                            match = gr.Button(LIKE_TEXT)
                            match.click(
                                fn=handle_feedback(button_id=i),
                                inputs=[post_ids, query],
                                outputs=[match],
                            )
                            match_buttons.append(match)

            body = gr.HTML("")
            post_bodies.append(body)

            toggle.click(
                fn=handle_toggle,
                inputs=[toggle],
                outputs=[toggle, body],
            )

    allblocks = (
        header +
        post_boxes +
        post_titles +
        toggle_buttons +
        match_buttons +
        post_bodies +
        [post_ids]
    )

    query.submit(
        fn=search,
        inputs=[query],
        outputs=allblocks)
    submit.click(
        fn=search,
        inputs=[query],
        outputs=allblocks)


demo.launch()