Grounded generation means the model answers only from the supplied context.
The prompt explicitly tells it not to use outside knowledge.
It must be allowed to say the context does not contain the answer.
This dramatically reduces hallucination compared to open-ended answering.
Grounding is the entire reason RAG exists.
Grounded generation means answering from the supplied context rather than from memory.
It is the entire point of RAG, and it does not happen automatically.
Given a question and documents, a model will happily blend both sources.
The prompt has to say explicitly that the answer must come from the context.
It must also permit the answer that the context does not contain it.
Without that permission the model will invent something rather than appear unhelpful.
A single sentence granting that permission removes a large share of hallucinations.
Asking for citations grounds it further, since a claim needing a source is harder to fabricate.
Asking for the supporting quote before the answer is stronger still.
Quoting first makes the model look at the text rather than reconstruct from impression.
Grounding reduces hallucination sharply and does not eliminate it.
The model can still misread, over-generalise, or merge two documents into one wrong claim.
It also cannot tell whether the retrieved document is itself correct.
Grounding on a stale or wrong source produces a confident, well-cited, wrong answer.
Faithfulness checks measure whether each claim is genuinely supported by the context.
One question on this card. Every option explains itself once you pick, right or wrong.
19 concepts in this module, then a graded test with every answer explained.