Caio-Fis/gaia-final-assignment

GAIA Final Assignment Agent (HF Agents Course — Unit 4)

Agent built on Gemini 2.5 (google-genai SDK) that answers the 20 GAIA level-1 questions from the course scoring API.

Architecture

• —agent.py — GaiaAgent: Gemini 2.5 Pro (fallback Flash) with Google Search grounding + URL context for research questions; native multimodal input for attachments (PNG/MP3 as bytes, YouTube links as file_data, XLSX converted to a markdown table, .py inlined as text). Answers follow GAIA exact-match formatting rules and are extracted from a final ANSWER: line.
• —run_local.py — runs the agent over all questions from GET /questions, caching results incrementally to answers.json (retries only on technical failure).
• —app.py — the original Gradio submission app from the course template (kept for reference; Gradio hosting on free Spaces is no longer available, so answers were generated locally with run_local.py and submitted to POST /submit).
• —answers.json — the agent's cached answers, exactly as submitted.

Note: the course API GET /files/{task_id} endpoint was returning 404 for all attachments at the time of the run; the 5 attachment files were fetched from a public mirror instead and fed to the same pipeline.