GLOSSARY
Grounding
Grounding connects AI outputs to supplied or retrieved information so answers rest on verifiable foundations; source errors and incorrect conclusions remain possible.
The source must support the specific claim
When preparing for a customer meeting, a system might connect its summary to an annual report. That is useful when readers can see which passage supports each statement. A link to the report does not substantiate every inference drawn from it. An announced investment, for example, does not establish demand for a particular product. Grounding should preserve the distinction between supported content and subsequent interpretation. With long documents, that mapping matters more than a lengthy list of links.
Check the complete evidence chain
RAG can supply relevant information; whether the answer uses it correctly is a separate question. Evidence also depends on freshness, scope and source reliability. Tests should include incorrect, outdated and only partly relevant supporting material. Examine whether the system recognises uncertainty, retains qualifications and avoids unsupported conclusions. This can make sales review more focused, but does not justify a blanket claim of error-free answers. The useful proof is a set of traceable responses to the tasks the application is actually intended to handle.