Grounding means the answer is anchored to retrieved documents that can be cited and checked.
DefinitionWhat it means
Grounding constrains a model to base its answer on specific retrieved or provided source material, rather than relying solely on knowledge baked into its weights. A grounded system typically instructs the model to answer only from the supplied context and to cite which passage supports each claim, making the output traceable back to a source document.
Why it mattersWhy you should care
Grounding is what turns a general-purpose language model into a trustworthy enterprise assistant: it enables citations, lets users verify claims, and keeps answers current with data the model was never trained on, like this quarter's internal policy or the latest pricing sheet. It is also the primary lever teams pull to reduce hallucination in customer-facing and compliance-sensitive applications.
At a glanceSee it
Attribution gate — the draft is split into atomic claims and each must be backed by a retrieved chunk to earn a citation, otherwise it is dropped rather than asserted as fact.
Iterative retrieval loop — when the fetched evidence is judged insufficient the system reformulates and re-queries, so generation begins only once the context can actually support the answer.
Where you see itIn the wild
- Enterprise search assistants that show a citation link next to every claim.
- Legal and medical copilots where an ungrounded answer is a compliance risk.
- Evaluation rubrics scoring whether a claim is supported by the retrieved context.