Grounding
Tying model outputs to verifiable sources. Usually retrieved documents or tool results.
Last updated: April 26, 2026
Definition
Grounding means the model's answer comes from a specific, traceable source rather than its training data. The cleanest grounding pattern: pass retrieved documents as context, instruct the model to answer ONLY from those documents, and require it to cite which document each claim came from. Grounded outputs can be verified after the fact. Ungrounded outputs cannot. For high-stakes applications (legal, medical, financial), grounding is non-negotiable.
When To Use
Required for any answer that could be wrong with consequences. Paired with RAG it virtually eliminates hallucination on the topics you have indexed.
Related Terms
Building with Grounding?
I've shipped this pattern in real production systems. If you want a second pair of eyes on your architecture, that's what I do.