FreshUser/gaia-agent

GAIA Agent — HF AI Agents Course Unit 4 Final Assignment

A LangGraph ReAct agent (skeleton from Unit 2.3) that answers the course's 20-question GAIA Level-1 validation subset, scoring ≥30% for the certificate.

Architecture

question ──► assistant (GLM-5.3 + bind_tools) ⇄ ToolNode ──► submit_answer
                          │                        │
                          ▼                        ▼
              web_search / wikipedia_search / visit_webpage
              file_reader (VLM Qwen3-VL / faster-whisper / pandas / pypdf)
              python_repl (subprocess, 30s timeout)

• Agentic RAG pattern (Unit 3): the agent decides when/what to retrieve
• Exact-match discipline: a submit_answer tool records the bare answer
• Salvage fallback: if the loop exhausts its step budget, one final extraction LLM call summarizes the research transcript into an answer
• Tool errors are non-fatal (handle_tool_errors=True) — the agent retries/redirects

Files

File	Role
app.py	LangGraph ReAct agent + answer salvage + format cleanup
tools.py	6 tools: DDG search, Wikipedia API, webpage reader, file reader (image VLM / audio STT / Excel / PDF), python REPL, submit_answer
runner.py	Fetch questions → download attachments (scoring API → GAIA dataset fallback) → run agent → answers.json cache
submit.py	POST answers to the course scoring API

Notes

• Attachments for file questions come from the public gaia-benchmark/GAIA validation set (the scoring API's /files endpoint was empty at build time).
• Secrets (LLM endpoint/key) are environment-driven — never in code.