app.py

import os
import requests
import threading
from typing import Optional, List, Tuple

import gradio as gr


ENDPOINT_URL = "https://austrian-code-wizard--metarlaif-web-dev.modal.run"


def get_feedback_options() -> List[Tuple[str, str]]:
    args = {
        "C3PO_API_KEY": os.environ.get("C3PO_API_KEY"),
    }
    response = requests.post(f"{ENDPOINT_URL}/list_adapters", json=args)
    data = response.json()["adapters"]
    return [
        (adapter["feedback_name"], adapter["feedback_id"])
    for adapter in data]


def get_completion(prompt: str, adapter: Optional[str], enable_feedback: bool) -> str:
    args = {
        "C3PO_API_KEY": os.environ.get("C3PO_API_KEY"),
        "prompt": prompt if enable_feedback else None,
        "adapter": adapter,
    }
    response = requests.post(f"{ENDPOINT_URL}/completion", json=args)
    data = response.json()
    return data["response"]


def warmup(*args):
    args = {
        "C3PO_API_KEY": os.environ.get("C3PO_API_KEY"),
    }
    # Warmup the server but don't wait for the response
    threading.Thread(target=requests.post, args=(f"{ENDPOINT_URL}/warmup"), kwargs={"json": args}, daemon=True).start()

dropdown_options = get_feedback_options()

demo = gr.Interface(
    get_completion,
    [
        gr.Markdown(
        """
        # C3PO Demo

        This is a demo of Contextualized Critiques with Constrained Preference Optimization (C3PO). See the project website [here](<insert link>), repo [here](<insert link>), and the paper [here](<insert link>).

        Selecting a feedback in the dropdown and enabling the "Use Feedback Adapter" checkbox will add the respective adapter to the model. The model will then use the feedback to generate the completion.
        """
        ),
        gr.Textbox(
            placeholder="Enter a prompt...", label="Prompt"
        ),
        gr.Dropdown(
            choices=dropdown_options, label="Feedback", info="Will add the adapter for the respective feedback to the model."
        ),
        gr.Checkbox(
            label="Use Feedback Adapter", default=True, info="Enables the feedback adapter."
        )
    ],
    "text",
    concurrency_limit=8
)

if __name__ == "__main__":
    demo.queue(max_size=32)
    demo.launch()