Grounding
Tying a model's output to specific, verifiable sources — retrieved documents, structured data, tool results — instead of relying on its training memory.
Definition
Tying a model's output to specific, verifiable sources — retrieved documents, structured data, tool results — instead of relying on its training memory. Grounded answers can be cited, audited, and trusted in operational settings.
Example
An ungrounded model might confidently invent a statistic. A grounded one quotes the exact line from the source document it was given, so an operator can verify the claim before acting on it.
See it in context Learn how Grounding fits into the bigger picture of how software actually works.
Read the Guide →