Definition
Grounding means the model's claims are supported by retrieved context, tool results, or structured data — often with citations users can check.
RAG is a common grounding pattern; browsing tools and SQL agents are others.
In simple terms
Speaking with footnotes — every strong claim points to a source on the desk, not just memory.
Where you see it
- Doc chatbots that cite page numbers.
- Customer support answers limited to policy PDFs.
How it works
1.Retrieve evidence
Pull relevant chunks or tool outputs.
2.Condition generation
Prompt the model to use only that evidence.
3.Attribute
Show citations or quotes alongside the answer.
Why it matters
- Grounding is the practical antidote to unconstrained hallucination in enterprise AI.
Often confused
If context is provided, the model cannot hallucinate.
Models may still ignore or contradict context — evaluate faithfulness explicitly.