Grounding Check
Also called Grounding Check
Verifying that each claim in an answer is actually supported by the sources it was given.
Think of it like
A fact-checker holding the article against the interview transcript, line by line, to confirm nothing was invented.
Example
In a RAG app, a checker confirms the answer’s statement "the warranty is two years" appears in a retrieved document, and flags it if not.
How it actually works
Retrieval is supposed to keep a model honest, but the model can still drift beyond its sources or blend in remembered guesses. A grounding check (or faithfulness check) compares the generated claims against the provided context and scores how well each is supported — often using another model as the judge. It’s the specific test of "did it stick to the documents," and a core piece of trustworthy RAG.
For product teams
The measurable answer to "can we trust that this answer came from our docs and not the model’s imagination?"
For engineers
Score each claim’s support against the provided context (often via an LLM judge); the faithfulness metric central to RAG evaluation.
Related
- Grounding — The property it verifies.
- Hallucination Detection — A broader detection effort it feeds.
- RAG — It checks the sources RAG retrieved.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome