What the term really means
Basing a model response on specified data, documents or search results rather than model knowledge alone.
How it works in practice
A grounded answer relies on identified data such as a current return policy or order record instead of guessing from model memory.
The decision to make before implementation
Sources must be current, authorised for that user and supplied at response time; stale documents can produce precise but wrong output.
How to verify that it works
Check claim-to-source consistency, citation completeness, document freshness and behaviour when evidence is missing. Compare results with an agreed baseline and review routine cases, difficult exceptions and human hand-offs separately. A practical Grounding test should have an owner, a review date and a recorded example of an outcome the team will not accept.
The PAR HOUSE Agency approach
We approach Grounding from the workflow rather than a tool demonstration. Sources must be current, authorised for that user and supplied at response time; stale documents can produce precise but wrong output. We then build a small measurable scope, record assumptions and expand only after quality review.
