A grounded answer uses relevant source information rather than relying only on a model's general knowledge. Retrieval-augmented generation retrieves selected material and supplies it to the model for a request; retrieval does not by itself retrain the model.
Follow the retrieval pipeline.
A source is prepared for retrieval, often through indexing and division into manageable passages. At request time, search identifies relevant material, which is supplied to the model alongside the question and application instructions. The model then generates an answer using that context.
Retrieval quality depends on content structure, metadata, search behavior, and freshness. A technically available document can still be a poor source if its location, effective date, audience, or version cannot be identified. More retrieved text is not necessarily better evidence.
Enforce access before generation.
Search and retrieval should respect the identity and permissions appropriate to the task. Do not give a model unauthorized material and rely on an instruction to hide it afterward. Indexes, cached results, and connectors need access behavior consistent with the source and application design.
Permission changes and document deletion also need to propagate to the retrieval path. Understand refresh delays and supported enforcement points before using a grounded application for sensitive work.
Example: resolving an expense question.
An employee asks whether a particular travel cost is covered. Useful retrieval identifies the current policy for their location and employment type, including applicable exceptions. An older global policy may look relevant but produce the wrong recommendation.
Review both the selected passages and the answer. A citation should support the associated claim, not merely point to a related document. If evidence conflicts or is incomplete, the application should make that limitation visible instead of completing the answer with unsupported assumptions.
Measure evidence quality separately.
Test source relevance, permission handling, claim support, and the treatment of missing information. Keep examples where retrieval should return no authorized answer. Assign source owners and evaluate how content changes affect results. Grounding can improve relevance and traceability, but it does not guarantee truth, current information, or a correct interpretation of the evidence.
Further reading
For source readiness, access review, and pilot evaluation, read Copilot and data readiness.