The final stage: the model reads your retrieved context and writes an answer. Grounding means constraining it to use only that context. This is achieved through explicit instructions, attribution markers, an easy path to saying "I don't know", and verification of the output. Get this wrong and all your retrieval work is spent on an answer the model half-ignores.

This is the most under-appreciated part of grounded generation. A model asked a question will strongly tend to produce an answer — that is what it was trained to do. Unless refusing is made explicitly acceptable and concrete, it will fill the gap with plausible invention.

Verify citations programmatically. Models sometimes cite [3] when the claim came from [1], or cite a source number that does not exist. Since you control the context, you can check every emitted marker against the actual chunk list — and you should. A wrong citation is worse than no citation, because it manufactures unearned confidence.

Temperature 0 is the right default for RAG and surprises people who expect it to make answers "robotic". It does not — the model still writes fluent prose. It simply picks the most likely wording rather than sampling among alternatives, which is exactly what you want when the content must mirror the source.

If you run output verification — citation checks, safety filters, groundedness scoring — streaming creates a genuine tension: you have shown the user text you have not yet validated. In high-stakes domains, verify before displaying and accept the slower first token.

None of these eliminate hallucination — they reduce its rate and make it detectable. A production system in a high-stakes domain should assume some answers will be wrong and design for that. Show sources prominently, make checking easy, and keep a human in the loop where the cost of being wrong is high.