RAG systems gather every document in the organisation into one search pool. Without correct authorisation that pool turns into a leak channel.
01. The Most Common Mistake
Throwing all documents into a single index and letting everyone query it. HR files, salary data and board notes surface on the first search in a setup like that.
02. Where Filtering Belongs
Permission checks must be applied in the search layer, while the query runs. Filtering after results come back is both slow and unsafe; the model has already seen them.
03. Labelling
Every chunk needs an access label: department, role, project. That label must come from the source system, not be assigned by hand; hand-assigned labels stop reflecting reality over time.
04. Syncing with the Source
When a document's permissions change at the source, they must change in the search layer too. Delayed synchronisation means a departed employee's documents stay reachable for weeks.
05. Personal Data
If the document set contains personal data, data protection obligations apply in full. Retention, deletion requests and privacy notices must be honoured here too; a deleted record must also leave the search index.
06. Logging and Audit
Who asked what and which documents were returned must be recorded — both for audit and to catch misconfiguration early.