app.py

from openai import AsyncAssistantEventHandler
from openai import AsyncOpenAI
import gradio as gr
import asyncio
import os

# set the keys
client = AsyncOpenAI(
  api_key=os.getenv("OPENAI_API_KEY")
)

assistantID = os.getenv("OPENAI_ASSISTANT_ID")
username = os.getenv("YOUR_ID")
password = os.getenv("YOUR_PASSWORD")

mytitle = "<h1 align=center>RTL AI News Reader : Wat war lass am Land 🇱🇺 an op der Welt 🌎 ?</h1>"

mydescription="""
<h3 align='center'>Wat fir een Thema interesséiert Iech : 🐶 🏃🏻‍♂️ 🌗 🍇 🌈 🍽️ 🏆 🚘 ✈️ 🩺 </h3>
<table width=100%>
  <tr>
    <th width=50% bgcolor="Moccasin">Stell deng Froen op Lëtzebuergesch, oder an enger anerer Sprooch :</th>
    <th bgcolor="Khaki">Äntwert vum OpenAI File-Search Assistent : </th>
  </tr>
</table>
"""

myarticle ="""
<h3>Hannergrënn :</h3>
<p>Dës HuggingFace Space Demo gouf vum <a href="https://github.com/mbarnig">Marco Barnig</a> realiséiert. 
Als kënstlech Intelligenz gëtt, mëttels API, den <a href="https://platform.openai.com/docs/models">OpenAI Modell</a> 
gpt-4o-mini-2024-07-18 benotzt, deen als Kontext bis 128.000 Tokens ka benotzen, eng Äntwert op eng Fro vu maximal 16.384 
Tokens ka ginn a bis zu 200.000 Tokens pro Minutt (TPM) ka beaarbechten. 
De ganze lëtzebuergesche Contenu vun RTL.lu vum Ufank (2000, 2012, ???) bis September 2024 gouf a 50 JSON-Dateien opgespléckt an op 
e Vector Store vum OpenAI File-Search Assistent "RTL News Reader" eropgelueden. 
All Datei huet manner wéi 5 Milliounen Token, wat eng iewescht Grenz fir den AI Modell ass. 
Et ass méiglech bis zu 10.000 Dateien op en OpenAI Assistent opzelueden.
D'Äntwerte vun de Beispiller sinn am Cache gespäichert a ginn duerfir ouni Delai ugewise.</p>
"""

myinput = gr.Textbox(lines=3, label="Wat wëllt Der wëssen ?")

myexamples = [
  "Wat war lass am Juni 2023 ?",
  "Wat ass gewosst iwwert de SREL ?",
  "Wat fir eng Katastroph war 2022 zu Lëtzebuerg ?",
  "Koumen an de leschte Jore gréisser Kriminalfäll viru Geriicht ?"
]

class EventHandler(AsyncAssistantEventHandler):
    def __init__(self) -> None:
        super().__init__()
        self.response_text = ""

    async def on_text_created(self, text) -> None:
        self.response_text += str(text)

    async def on_text_delta(self, delta, snapshot):
        self.response_text += str(delta.value)

    async def on_text_done(self, text):
        pass

    async def on_tool_call_created(self, tool_call):
        self.response_text += f"\n[Tool Call]: {str(tool_call.type)}\n"

    async def on_tool_call_delta(self, delta, snapshot):
        if snapshot.id != getattr(self, "current_tool_call", None):
            self.current_tool_call = snapshot.id
            self.response_text += f"\n[Tool Call Delta]: {str(delta.type)}\n"

        if delta.type == 'code_interpreter':
            if delta.code_interpreter.input:
                self.response_text += str(delta.code_interpreter.input)
            if delta.code_interpreter.outputs:
                self.response_text += "\n\n[Output]:\n"
                for output in delta.code_interpreter.outputs:
                    if output.type == "logs":
                        self.response_text += f"\n{str(output.logs)}"

    async def on_tool_call_done(self, text):
        pass

# Initialize session variables
session_data = {"assistant_id": assistantID, "thread_id": None}

async def initialize_thread():
    # Create a Thread
    thread = await client.beta.threads.create()
    # Store thread ID in session_data for later use
    session_data["thread_id"] = thread.id

async def generate_response(user_input):
    assistant_id = session_data["assistant_id"]
    thread_id = session_data["thread_id"]

    # Add a Message to the Thread
    oai_message = await client.beta.threads.messages.create(
        thread_id=thread_id,
        role="user",
        content=user_input
    )

    # Create and Stream a Run
    event_handler = EventHandler()

    async with client.beta.threads.runs.stream(
        thread_id=thread_id,
        assistant_id=assistant_id,
        instructions="Please assist the user with their query.",
        event_handler=event_handler,
    ) as stream:
        # Yield incremental updates
        async for _ in stream:
            await asyncio.sleep(0.1)  # Small delay to mimic streaming
            yield event_handler.response_text

# Gradio interface function (generator)
async def gradio_chat_interface(user_input):
    # Create a new event loop if none exists (or if we are in a new thread)
    try:
        loop = asyncio.get_running_loop()
    except RuntimeError:
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)

    # Initialize the thread if not already done
    if session_data["thread_id"] is None:
        await initialize_thread()

    # Generate and yield responses
    async for response in generate_response(user_input):
        yield response

# Set up Gradio interface with streaming
interface = gr.Interface(
    fn=gradio_chat_interface,
    inputs=myinput,
    outputs="markdown",
    title=mytitle,
    description=mydescription,
    article=myarticle,
    live=False,
    allow_flagging="never",
    examples=myexamples
)

# Launch the Gradio app
interface.launch(auth=(username, password), auth_message="<h1>RTL AI News Reader</h1><p>Dëse HuggingFace Space ass e Prototyp an nach net zougänglech fir jiddereen. De Projet baséiert op engem OpenAI API File-Search Assistent a benotzt de Modell GPT-4o-mini. Dir musst e Chrome Browser benotzen. Interesséiert KI-Spezialiste kënnen eng ID a Passwuert beim [email protected] ufroen.</p>")