Grounding
Constraining a model to answer from provided source material rather than its own training, to keep answers factual and checkable.
Grounding means tying a model's answer to a specific source — the retrieved documents, the pasted context — rather than letting it draw on whatever it absorbed in training. A grounded answer can be traced back to the material it came from, which is what makes it trustworthy in an application.
In practice grounding is enforced by the prompt: instruct the model to use only the provided context, to cite the passage it relied on, and to say plainly when the answer is not present. Without those instructions, a model will happily fill gaps from memory, which is where ungrounded hallucinations come from.