Grounding
Tying a model's answer to a trusted source, such as a specific document, record, or database, so it is based on real information rather than generated from memory. Grounding is the main defense against hallucination: instead of letting the model produce a plausible-sounding answer on its own, you make it draw from and point to source material, often using retrieval-augmented generation. For clinicians and the people who buy these tools, a grounded answer that cites where it came from can be checked, which is what makes it safer to act on than an ungrounded one.
A hospital's guideline assistant answers a nurse's question about when the adult DKA protocol says to hold insulin and links to the paragraph it used, so she can check it before acting. A week later a pediatric nurse asks the same thing, but the pediatric protocol was never loaded, and the assistant answers fluently from adult practice with no link. Before go-live, ask it questions the indexed documents cannot answer, confirm it says it found nothing, and open a sample of citations to check the cited text says what the answer claims.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →