retrieval grounding

Grounding means the model's answer is tied to the retrieved passages rather than spun from memory. A well-grounded answer says only what the source text actually supports, and if the passages do not cover the question, a grounded system says so plainly instead of guessing. It is the discipline of staying within the evidence you were handed.

You encourage grounding by instructing the model to answer only from the provided context and to admit when the answer simply is not there. Grounding is the whole point of RAG — it is the mechanism that cuts hallucination — but it is never automatic: a model can still ignore the passages, blend them with half-remembered training facts, or quietly fill a gap. So grounding is something you have to test for and design toward, not something you can assume just because retrieval happened.