goldmanager14147/gaia-agent-final-assignment

GAIA Benchmark Agent — HF Agents Course Final Assignment

This Space hosts the code for my agent built for the Hugging Face AI Agents Course, Unit 4 final assignment. The agent answers a subset of GAIA-benchmark questions and submits its answers to the course scoring API (https://agents-course-unit4-scoring.hf.space) for grading.

How it works

• agent.py builds a smolagents.CodeAgent with tools for web search (DuckDuckGo), webpage visiting, Wikipedia search, and reading local attachments (Excel, text/code, audio transcription).
• The agent's backend model is configurable:
  • AGENT_BACKEND=hf (default in this Space) uses Hugging Face Inference Providers (Qwen/Qwen2.5-Coder-32B-Instruct).
  • AGENT_BACKEND=ollama uses a locally-run Ollama model instead (used for development/testing without consuming API credits).
• app.py provides the Gradio UI (login via gr.LoginButton, "Run Evaluation & Submit All Answers" button) required by the course template.
• run_eval.py is a CLI alternative that fetches questions, runs the agent, and (optionally) submits results — used for local development/testing.

Files

• agent.py — agent definition (tools + model + prompt formatting).
• app.py — Gradio app that fetches questions, runs the agent, and submits answers through the scoring API.
• run_eval.py — CLI runner used during local development.
• requirements.txt — Python dependencies.