What is grounding?

Grounding means making the model answer from source text supplied at reply time, not from what it absorbed in training, so that every claim in a reply traces back to a passage in your content. A grounded agent, asked something its sources do not cover, says it does not know. An ungrounded one answers anyway, from the internet's average opinion of businesses like yours.

It matters because grounding is the main defense against hallucination, and because it is a spectrum rather than a switch. Retrieval-augmented generation supplies the passages; grounding is the set of instructions and checks that keep the model inside them: answer only from the passages, repeat numbers rather than rounding them, say "I don't know" when nothing relevant was found, and prefer a handoff to a guess. Some products add citations that show which page each answer came from, which makes grounding visible and checkable.

Where it runs out: grounding limits the model to your content, so it inherits your content's mistakes. An outdated pricing page produces a grounded, confident, wrong answer. Grounding cannot settle a question your content answers twice in two ways. And a model told strictly to answer only from sources may decline questions a person would have answered from obvious context, so the fallback message and the handoff decide whether strictness feels careful or unhelpful.