Grounding

Grounding is the degree to which an AI system's response is actually supported by, and traceable to, the retrieved source material it was given — as opposed to being drawn from the model's own trained-in memory, where it can't be checked and might simply be wrong.

A grounded answer can be checked, claim by claim, against a specific passage that supports it. An ungrounded answer might be entirely correct, entirely fabricated, or some invisible mix of both — from the outside, fluent text looks the same either way. Grounding is what RAG architectures exist to produce, and it's a property that has to be measured, not assumed just because a retrieval step ran.

How it works

Measuring groundedness means decomposing a response into its individual factual claims and checking each one against the retrieved passages the model was given: does this specific sentence follow from something actually in the context, or did the model add a detail, a number, or a citation that isn't there? This is usually done with an LLM-as-judge, prompted specifically to compare claim against source rather than to judge overall quality, because the check is narrow and mechanical enough to automate but too nuanced for a code grader — it involves reading comprehension, not pattern matching.

A retrieval step running is not the same as grounding succeeding. A model can be handed perfectly relevant source passages and still add an unsupported detail, misremember a number from the context, or answer a part of the question the passages didn't actually cover — retrieval quality (recall@k, precision@k) and generation groundedness are separate failure points that need separate measurement, because fixing one doesn't guarantee the other improved.

Why it matters

Groundedness rate is the metric that turns "our RAG system reduces hallucination" from a marketing claim into a measured fact, tracked the same way any other quality metric is tracked — per release, with regressions caught before they ship. It's also diagnostic: a low groundedness rate with high recall@k points at the generation step (the model isn't sticking to its sources), while a low groundedness rate with low recall@k points upstream, at retrieval never finding the right evidence in the first place.

A worked example

An answer to "what's our refund window" makes five distinct factual claims: the refund period is 30 days, it starts from delivery date, it excludes final-sale items, refunds process in 5-7 business days, and store credit is offered as an alternative. Checking each against the retrieved policy document finds four directly supported — but "processes in 5-7 business days" appears nowhere in the retrieved text; the model added a plausible-sounding detail. Groundedness rate for this response: 4/5 = 80 percent, and the specific unsupported claim is exactly the one a downstream citation check should have caught before the answer shipped.

How miatz teaches it

The grounding lab has learners take a generated answer, split it into individual claims by hand, and check each against the source documents themselves before ever trusting an automated judge to do it — so a groundedness score is understood as a claim-by-claim audit, not an opaque percentage.