Grounding

A grounded answer is one built from evidence you put in front of the model, not from what it happens to remember. You retrieve the relevant help articles, database rows, or tool results, include them in the request, and instruct the model to answer from that material and say when it falls short.

Grounding is the main defense against stale or invented facts, and it is what makes an AI feature specific to your product. A generic model knows nothing about your customer's order or your refund policy; a grounded one can answer correctly about both, as long as retrieval found the right material.

Grounding is only as good as two things you control. Retrieval has to surface the right sources, so a perfect prompt can't save a search that missed. And the model has to stick to them, which you check with faithfulness evals and enforce with citations that point to specific passages. Asking for citations without verifying them only moves the hallucination into the footnotes.

In the interface, grounding is something you can show: link each claim to its source, and mark answers that drew on no source differently from those that did.