app.py

import hashlib
import json
import os
import random
import uuid
from datetime import datetime
from pathlib import Path

import gradio as gr
from huggingface_hub import CommitScheduler, get_token, login
from openai import OpenAI

from prompts import basic_prompt, detailed_genre_description_prompt
from theme import TufteInspired

# Ensure you're logged in to Hugging Face
login(os.getenv("HF_TOKEN"))

# Define available models
MODELS = [
    "meta-llama/Meta-Llama-3-70B-Instruct",
    "mistralai/Mixtral-8x7B-Instruct-v0.1",
    "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO",
]


def get_random_model():
    return random.choice(MODELS)


def create_client(model_id):
    return OpenAI(
        base_url=f"https://api-inference.huggingface.co/models/{model_id}/v1",
        api_key=get_token(),
    )


# Set up dataset storage
dataset_folder = Path("dataset")
dataset_folder.mkdir(exist_ok=True)


# Function to get the latest dataset file
def get_latest_dataset_file():
    files = list(dataset_folder.glob("data_*.jsonl"))
    return max(files, key=os.path.getctime) if files else None


# Check for existing dataset and create or append to it
if latest_file := get_latest_dataset_file():
    dataset_file = latest_file
    print(f"Appending to existing dataset file: {dataset_file}")
else:
    dataset_file = dataset_folder / f"data_{uuid.uuid4()}.jsonl"
    print(f"Creating new dataset file: {dataset_file}")

# Set up CommitScheduler for dataset uploads
repo_id = "davanstrien/summer-reading-preferences"
scheduler = CommitScheduler(
    repo_id=repo_id,
    repo_type="dataset",
    folder_path=dataset_folder,
    path_in_repo="data",
    every=1,  # Upload every minute
)

# Global dictionary to store votes
votes = {}


def generate_prompt():
    if random.choice([True, False]):
        return detailed_genre_description_prompt()
    else:
        return basic_prompt()


def get_and_store_prompt():
    prompt = generate_prompt()
    print(prompt)  # Keep this for debugging
    return prompt


def generate_blurb(prompt):
    model_id = get_random_model()
    client = create_client(model_id)
    max_tokens = random.randint(100, 1000)
    chat_completion = client.chat.completions.create(
        model="tgi",
        messages=[
            {"role": "user", "content": prompt},
        ],
        stream=True,
        max_tokens=max_tokens,
    )
    full_text = ""
    for message in chat_completion:
        full_text += message.choices[0].delta.content
        yield full_text, model_id
    return full_text, model_id  # Return final result with model_id


def generate_vote_id(user_id, blurb):
    return hashlib.md5(f"{user_id}:{blurb}".encode()).hexdigest()


def log_blurb_and_vote(
    prompt,
    blurb,
    vote,
    model_id,
    user_info: gr.OAuthProfile | None,
):
    user_id = user_info.username if user_info is not None else str(uuid.uuid4())
    vote_id = generate_vote_id(user_id, blurb)

    if vote_id in votes:
        gr.Info("You've already voted on this blurb!")
        return None, gr.Row.update(visible=False)

    votes[vote_id] = vote

    log_entry = {
        "timestamp": datetime.now().isoformat(),
        "prompt": prompt,
        "blurb": blurb,
        "vote": vote,
        "user_id": user_id,
        "model_id": model_id,
    }
    with scheduler.lock:
        with dataset_file.open("a") as f:
            f.write(json.dumps(log_entry) + "\n")

    gr.Info("Thank you for voting! Your feedback will be synced to the dataset.")
    return f"Logged: {vote} by user {user_id}", gr.Row.update(visible=False)


# Create custom theme
tufte_theme = TufteInspired()

# Create Gradio interface
with gr.Blocks(theme=tufte_theme) as demo:
    gr.Markdown("<h1 style='text-align: center;'>Would you read this book?</h1>")
    gr.Markdown(
        """<p style='text-align: center;'>Looking for your next summer read? 
    Would you read a book based on this LLM generated blurb? <br> Your vote will be added to <a href="https://huggingface.co/datasets/your-username/your-dataset-repo">this</a> Hugging Face dataset</p>"""
    )

    with gr.Row():
        login_btn = gr.LoginButton(size="sm")
    with gr.Row():
        generate_btn = gr.Button("Create a book", variant="primary")

    prompt_state = gr.State()
    blurb_output = gr.Markdown(label="Book blurb")
    user_state = gr.State()
    model_state = gr.State()

    with gr.Row(visible=False) as voting_row:
        upvote_btn = gr.Button("👍 would read")
        downvote_btn = gr.Button("👎 wouldn't read")

    vote_output = gr.Textbox(label="Vote Status", interactive=False, visible=True)

    def generate_and_show(prompt, user_info):
        return "Generating...", gr.Row.update(visible=False), user_info, None

    def show_voting_buttons(blurb, model_id):
        return blurb, gr.Row.update(visible=True), model_id

    generate_btn.click(get_and_store_prompt, outputs=prompt_state).then(
        generate_and_show,
        inputs=[prompt_state, login_btn],
        outputs=[blurb_output, voting_row, user_state, model_state],
    ).then(
        generate_blurb, inputs=prompt_state, outputs=[blurb_output, model_state]
    ).then(
        show_voting_buttons,
        inputs=[blurb_output, model_state],
        outputs=[blurb_output, voting_row, model_state],
    )

    upvote_btn.click(
        log_blurb_and_vote,
        inputs=[
            prompt_state,
            blurb_output,
            gr.Textbox(value="upvote", visible=False),
            model_state,
            user_state,
        ],
        outputs=[vote_output, voting_row],
    )
    downvote_btn.click(
        log_blurb_and_vote,
        inputs=[
            prompt_state,
            blurb_output,
            gr.Textbox(value="downvote", visible=False),
            model_state,
            user_state,
        ],
        outputs=[vote_output, voting_row],
    )

if __name__ == "__main__":
    demo.launch(debug=True)