Documentation gets written so someone can look something up someday. Reading it is optional, so a mistake in it can sit unnoticed for a long time. Agents read our knowledge base before they act, every day, and nobody reviews each answer before it counts. Mix normative rules, client information, and financial data under one roof, and a broad read can surface something that was never supposed to cross that line, without anyone doing anything wrong.
That is why we split it into three repositories by who is allowed to read what: one normative, one holding each client's information, one at a higher sensitivity level for partners only. The boundary is set by where a file lives, decided before the question ever gets asked, not by how well whatever answers it happens to behave.
Before building it this way, we tested a relevance-based search tool against data from two fictional clients. A specific question kept them apart. A broad one surfaced the wrong client, as a weak match. That settled it: confidentiality cannot depend on how well a search behaves.