CoolFace
Apppublic

AlwinXavier19/gaia-benchmark-agent

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes
App README

GAIA Benchmark Agent

A production-quality AI agent that solves GAIA Level 1 benchmark questions and submits to the HuggingFace leaderboard.

Architecture

Gradio UI (HuggingFace Space)
    ↓
BasicAgent class
    ↓
ReAct-style LLM Loop (Google Gemini 2.0 Flash)
    ↓
Tool Router (16 tools)

Tools

ToolPurpose
web_searchDuckDuckGo search
wikipedia_searchWikipedia articles
analyze_imageGemini Vision
transcribe_audioWhisper API
execute_python_file/codeSandboxed code
read_excel_file / excel_querySpreadsheet processing
get_youtube_transcriptVideo captions
reverse_text / decode_textText utilities
fetch_webpageWeb page content
calculateSafe math evaluation

Setup

  1. 1.Set the following HuggingFace Space secrets:
  2. 2.GOOGLE_API_KEY — Get free at aistudio.google.com
  3. 3.HF_TOKEN — Your HuggingFace token
  4. 4.OPENAI_API_KEY (optional) — For Whisper audio transcription
  1. 1.Click Run Agent & Submit to Leaderboard

Projected Score

CategoryExpected
Pure reasoning~100%
Web search / Wikipedia~70%
Audio transcription~80%
Excel / Code execution~90%
Image analysis~60%
YouTube video~50%
Overall~65-75%