Permission-aware retrieval checks for custom and packaged enterprise knowledge systems.
Retrieval-augmented generation, or RAG, supplies source material to a model before it answers. If retrieval includes material a user should not see, the answer may expose restricted information. This is a design risk to evaluate in both custom and packaged systems, not a claim that every product leaks data.
Follow a document from its source through ingestion, indexing, retrieval, and answer generation. Establish how source identities and access rules are represented at each stage. Test group membership changes, deleted documents, and shared access.
Try questions that refer indirectly to restricted material as well as direct requests. Check citations, previews, cached answers, logs, and exports. Where permission information is unavailable or stale, define a safe fallback rather than silently widening access.
Run regression tests when connectors, models, identity integrations, or source permissions change. Keep an incident process for withdrawing incorrect or exposed content. Document residual limitations so administrators understand what the system does and does not protect.
Book a free AI audit to discover how an AI-native services firm can transform your operations.
Book a free AI Audit