Grounding
Grounding means an answer is tied to specific supplied material rather than produced from model memory — and, in a usable system, that the tie is visible as a citation. It is the property that separates a system you can deploy where statements carry consequences from one you cannot. In practice it is closer to binary than to a spectrum: either an unsupported statement can be emitted or it cannot.
Why it behaves as a yes-or-no property
Because the interesting failure is not a slightly-off answer but an invented one, and inventions are produced by the same mechanism whether the model is strong or weak. A system permitted to compose freely when it finds nothing will occasionally compose something plausible; a system required to cite cannot, because there is nothing to cite.
That is why "how accurate is it" is the wrong procurement question and "what happens when the answer is not in the corpus" is the right one. The first invites a number; the second reveals the design.
What a citation has to be to be worth anything
Specific enough to check in seconds. A document name is not a citation when the document has ninety pages — the reader will not open it, and an unverified citation provides the appearance of rigour without the substance. Paragraph or page level is the threshold at which people actually verify.
The second requirement is that it points at the passage the answer was built from, not at a plausibly related one. Systems that retrieve, answer, and then search for a supporting quote produce citations that look correct and sometimes are not — a failure mode that is close to undetectable from the outside.
The part nobody configures
Permission to say nothing. Most disappointing deployments trace back to a system that must always answer, because that setting was never questioned. Turning it off converts a class of invisible errors into visible gaps, which feels like a downgrade for a week and is the opposite.
The measurable consequence is that a refusal rate above zero becomes a sign of health rather than a defect. A system that answers everything on a corpus that does not contain everything is not more capable; it is less honest.
What it is not to be confused with
Accuracy
Accuracy is a rate over a sample. Grounding is a structural property of the system. A well-grounded system with a lower headline accuracy is more deployable than an ungrounded one with a higher number, because its errors are visible.
RAG
RAG is the architecture that makes grounding practical; grounding is the property you actually want. It is possible to build RAG and still let the model answer from memory when retrieval returns nothing — which is RAG without grounding, and the combination is worse than either alone.
Frequently asked
How do we test whether a system is grounded?+
Ask a question whose answer is demonstrably not in the corpus. A grounded system says so. Then ask a question whose answer is in a single known paragraph and check that the citation points there rather than somewhere adjacent. Two questions, ten minutes, and more informative than any benchmark.
Does grounding slow the system down?+
Slightly, and the cost is in retrieval rather than generation. The larger practical effect is on coverage: a grounded system answers fewer questions, because it declines the ones its corpus cannot support. That is the trade, and it is the right one wherever an answer gets acted on.
Can a model be grounded without citations?+
It can be constrained to supplied material without surfacing the source, and some products do exactly that. The result is untestable from the outside: you cannot tell a grounded answer from a fluent one. For work with consequences, a citation you can check is the whole point.
What about answers that combine several documents?+
That is the normal case and the reason single-document search is insufficient. What matters is that each statement carries its own source rather than the answer carrying one citation at the end — a combined answer with a single reference is unverifiable in exactly the places where combination introduced the error.
The two questions that settle it
One question whose answer is not in your documents, one whose answer is in a paragraph you can name. What comes back tells you whether the system can be deployed where statements have consequences.
Request a test