README.md

---
title: GAIA Agent - smolagents
emoji: 🧠
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: false
---

# GAIA Level-1 Agent (smolagents)

Hands-on assignment for **Unit 4** of the [Hugging Face Agents Course](https://huggingface.co/learn/agents-course/unit4/hands-on).

This Space runs an agent (built with **smolagents**) over the 20 GAIA level-1
evaluation questions served by the course scoring API, and can submit the
answers to the leaderboard.

## Framework / rationale
See `REPORT.md` (in the workspace root of this repository) for the full solution
report. Short version: a single `CodeAgent` writes its own Python, so it computes the exact answers itself while the provided tools cover web search, page reading, file download, whisper transcription, vision analysis and python execution.

## Setup (Space secrets)
| Secret | Required | Purpose |
|---|---|---|
| `LLM_API_KEY` | ✅ | key for an OpenAI-compatible endpoint |
| `LLM_BASE_URL` | optional | default `https://api.openai.com/v1` |
| `LLM_MODEL` | optional | default `gpt-4o-mini` |
| `HF_TOKEN` | optional | GAIA-dataset file fallback + local self-scoring |
| `VISION_MODEL` | optional | vision-capable model for image questions |
| `WHISPER_MODEL` | optional | `small` (default) / `base` / `tiny` |

## Usage
1. Log in with your Hugging Face account.
2. Click **Run all 20 questions AND submit to leaderboard**.
3. The submission uses your username and this Space's public code link.

> ⚠️ Running all 20 questions calls the LLM API many times and may take several
> minutes (YouTube/audio transcription is the slowest part). Keep the Space
> public so the leaderboard can verify your submission.