loading
loading
Forcing the model to base its answer on retrieved source documents instead of its own memory, so it stops making things up.
A grounded response is one tied to evidence you put in the context — the whole point of RAG. You retrieve real passages, hand them to the model, and instruct it to answer only from those, ideally citing them. This is the main defense against hallucination: when the model can quote a source it's far less likely to invent one. The catch is it's only as good as retrieval — feed it the wrong chunks and it'll confidently ground its answer in garbage.
Plainly
Think of Grounding as a named building block in a big LEGO app. Forcing the model to base its answer on retrieved source documents instead of its own memory, so it stops making things up.
In practice
Use it when you are mapping how the app is structured or explaining a feature to a teammate or agent. In practice, define the owner, input, output, and failure mode before you rely on it.