groundedrag

A hard gate that stops your RAG from answering without evidence.

Most RAG pipelines always call the LLM, even when retrieval finds nothing relevant — the model fills the gap with invented information, confidently. groundedrag is a small, opinionated Python library that adds two things on top of whatever retrieval stack you already have:

1. Documentary gate — before the LLM is ever called, checks whether there's real coverage for the question's topic, using an auditable topic → documents map a non-engineer can read and edit. No coverage means the LLM is never called.
2. Verified citations — every factual claim must cite [n]; a verifier runs after generation and checks each citation against what was actually retrieved.

See the GitHub repository for the full pitch, installation, and source. This site holds the deeper technical documentation:

• Quickstart — try it in two minutes, no API key.
• Architecture — how the pieces fit together and the project's roadmap.
• Benchmark — the hallucination-rate benchmark and what it does (and doesn't) measure.
• Decision records — every non-obvious design decision, its rejected alternatives, and what would invalidate it.
• Releasing — how versions get published to PyPI.