Retrieval inherits your permissions problem

A retrieval system that ignores document permissions is a data leak with a search box. If a salary review, a client contract and a board paper sit in the same index as the staff handbook, then anyone who can ask a question can reach all of them. Access control has to propagate from the source systems into retrieval, so results are filtered by who is asking before an answer is composed. That constraint shapes the architecture from day one and is close to impossible to retrofit onto an index that was built without it.

Citations, and knowing when to say nothing

The valuable behaviour in a knowledge system is not answering — it is declining to answer when the evidence is not there. An assistant that produces a plausible paragraph with no source is worse than a search box, because it is confidently wrong and hard to check. Every answer carries its citations, and the no-evidence path is a first-class outcome that gets logged and reviewed, since a recurring unanswerable question is a content gap the business should close. Scanned documents and dense tables need dedicated handling; treating them like prose is the most common reason retrieval quality disappoints.

Evaluate a permission-aware knowledge assistant pilot

Choose a small set of approved, current documents and representative questions with expected source passages, including questions the system must decline. Sync source permissions and test retrieval as users with different access before allowing answer generation; each answer should cite the document version and passage, while stale, conflicting or missing evidence returns a clear no-answer and review route. Measure citation support, answer correctness, abstention on unanswerable questions, retrieval latency and permission-test failures on the same evaluation set before and after changes. Name a content owner to resolve stale sources, and retain a rollback path to the previous accepted index if a sync or permission change degrades results.

How to measure the pilot

A pilot is accepted against the current process, not an industry promise. Agree the sample, baseline and owner before launch, then review the same measures after the workflow has handled real cases.