Needle is a research project for training an evidence-seeking agentic RAG system. It studies how reinforcement learning can help an agent search for relevant evidence, use that evidence in its answer, and avoid unnecessary retrieval steps.

Early research prototype. The current milestone is a deterministic HotpotQA vertical slice; model inference and reinforcement learning are not implemented yet.

Requirements:

• uv

Set up the environment:

uv sync --dev

Run the quality gates:

uv run ruff check .
uv run ruff format --check .
uv run pytest