"""Hugging Face Space entrypoint.
The free Space tier serves the Gradio SDK on ZeroGPU hardware. We build on
``gradio.Server`` — a FastAPI subclass with Gradio's launch engine baked in — so
our FastAPI routes and a status page are served, and ``app.launch()`` is the
Gradio-native launch the Space expects. ZeroGPU additionally refuses to start
unless the app defines at least one ``@spaces.GPU`` function (see below).
"""
import os
from fastapi.responses import HTMLResponse
from gradio import Server
from app.main import _startup, configure
FRONTEND_URL = "https://dillguill.github.io/knowledge-assistant/"
app = Server()
configure(app)
# gradio.Server manages its own lifespan, so run our startup work explicitly at
# module import (how the Space executes this file) rather than via a FastAPI
# lifespan we'd risk clobbering.
_startup()
# ZeroGPU Spaces fail at startup with "No @spaces.GPU function detected" unless
# the app defines at least one @spaces.GPU function. This backend does no GPU
# work, so this never-called stub exists only to satisfy that requirement.
# `spaces` is force-installed by HF on the Space but is absent locally / in tests.
try:
import spaces
@spaces.GPU
def _zerogpu_startup_probe() -> None:
return None
except ImportError:
pass
@app.get("/", response_class=HTMLResponse)
async def status_page() -> str:
return f"""
Knowledge Assistant API
Knowledge Assistant API
Backend for Knowledge Assistant —
open the frontend to chat.
"""
# The Space imports this module in Gradio reload mode (so __name__ != "__main__"),
# which is why launch() must run at module level — matching the canonical
# gradio.Server Space example (ysharma/text-behind-image). Gate on SPACE_ID so
# importing the module in tests/local tooling doesn't start a server.
if os.environ.get("SPACE_ID") or __name__ == "__main__":
app.launch(show_error=True)