Grounding

Constraining an answer to a specific, checkable source rather than the model's trained-in memory, typically by supplying retrieved passages and requiring the answer to cite them. A grounded system can still be wrong, but the error is now traceable to a document rather than buried in the weights.

Why exams ask this

Tested as the difference between a model answering from memory and a model answering from a supplied source it can be checked against. The correct answer usually names retrieval plus citation, not a prompt instruction asking the model to be accurate.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.