Skip to main content

Grounding (model outputs)

AI concepts

Anchoring generated content to verifiable sources or tools so answers cite reality—overlaps with retrieval patterns but emphasizes the reliability claim, not the pipeline name.

Grounding asks: can this statement be tied to evidence? RAG is one mechanism; tool calls that fetch fresh records are another. Grounding reduces but does not eliminate hallucination: retrievers can fetch wrong passages, and models can misread them.

Product language

Prefer “grounded in our docs as of DATE” over “the model knows.”