Insights

Enterprise AI · Aug 20, 2026

Knowledge Systems Fail at Permissions, Not Retrieval

Retrieval quality is a solved-enough problem for many corpora. Keeping access control intact through the retrieval step is where enterprise knowledge systems break.

SkyNet AI

The demo works

Retrieval-augmented systems demo well. Documents are embedded, a query is matched, and a model answers with a citation. For a small corpus with uniform permissions, this is genuinely close to production.

Then permissions arrive

Real organizations have documents that different people may see. The naive implementation ranks across everything and filters afterward. That leaks: result counts, similarity scores, and the existence of matches reveal restricted material even when content is withheld.

Filter before ranking

The robust approach applies permission constraints before scoring. Partition-level indexes make this practical: each partition carries a policy, and queries are routed only to partitions the requester may access. Latency stays manageable because the search space shrinks.

Provenance is not optional

Every retrieved passage should carry its source, version, and retrieval time. In regulated settings, an answer without provenance is not usable, regardless of its accuracy.

The lesson is that governed knowledge systems are a data architecture problem. Prompt engineering cannot repair a permission model that was applied in the wrong order.