README.md

metadata

title: Template Final Assignment
emoji: 🕵🏻‍♂️
colorFrom: indigo
colorTo: indigo
sdk: gradio
sdk_version: 5.25.2
app_file: app.py
pinned: false
hf_oauth: true
hf_oauth_expiration_minutes: 480

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

GAIA Agent

An OpenAI function-calling agent for the GAIA Unit 4 leaderboard.

Setup

In your Space, add a Secret (Settings → Variables and secrets):

• OPENAI_API_KEY — required.
• OPENAI_MODEL — optional, defaults to gpt-4o (needs vision for image tasks).

Then log in with the HF button and click Run Evaluation & Submit All Answers.

ZeroGPU

The Space runs on ZeroGPU hardware. Inference is done via the OpenAI API (remote), so no local GPU is used. ZeroGPU refuses to start without a @spaces.GPU function, so app.py registers a tiny no-op probe (_zerogpu_probe) just to pass startup detection. The real evaluation loop is intentionally not decorated — @spaces.GPU caps a call at ~60s and evaluating all 20 questions takes minutes.

Tip: since no GPU is actually used, plain CPU basic hardware also works and is simpler. ZeroGPU is only needed if you specifically want that hardware.

Structure

app.py                 # Gradio UI + fetch/run/submit loop (template, minimally edited)
agent/
  gaia_agent.py        # GaiaAgent: the OpenAI tool-calling loop + answer extraction
  prompts.py           # GAIA system prompt (exact-match formatting rules)
agent_tools/
  __init__.py          # tool registry + dispatch
  web.py               # web_search, visit_webpage
  files.py             # read_file (text/CSV/Excel) for attached task files
  media.py             # transcribe_audio (Whisper), analyze_image (vision)
  python_exec.py       # run_python (calculations)