404robbie/evidence-research-agent

Evidence Research Agent

A transparent LangGraph agent built for the Hugging Face Agents Course final assignment and its 20-question GAIA Level 1 subset.

What it can do

• —Search the web
• —Read web pages
• —Perform safe arithmetic
• —Inspect PDF, DOCX, spreadsheet, CSV, JSON, and text attachments
• —Return concise answers for exact-match scoring

Hugging Face Space setup

1. 1.Duplicate the official Final Assignment Template.
2. 2.Replace its files with the files in this project.
3. 3.In Settings → Variables and secrets, add HF_TOKEN as a secret.
4. 4.Optionally set AGENT_MODEL as a variable. The default is Qwen/Qwen3-32B.
5. 5.Keep the Space public so its code link can be submitted.

The Space uses the Hugging Face OpenAI-compatible router by default.

Local Ollama setup

Set these environment variables before running the app:

export MODEL_PROVIDER=ollama
export AGENT_MODEL=qwen2:7b
export OLLAMA_BASE_URL=http://localhost:11434/v1
python app.py

When Ollama runs on Windows and the app runs inside WSL, use the Windows host address in OLLAMA_BASE_URL rather than localhost if needed.

Design

This project constructs its own LangGraph state graph:

question → model reasoning → tool execution → model reasoning → exact answer

The loop is bounded with AGENT_RECURSION_LIMIT (default: 14) to prevent runaway tool calls.