Knowledge · Guide
Protect knowledge before it reaches the model
A polite refusal prompt is not a substitute for permission-aware retrieval.
Authorization is an application decision
If an employee cannot open a document, the assistant should not receive its contents on that employee’s behalf. Microsoft describes document-level access controls and security filters for retrieval. The practical design principle is to establish identity and allowed scope before assembling model context. Telling a model not to reveal restricted text is a weaker boundary than never retrieving that text for the unauthorized request.
Build negative permission tests
For an illustrative consultancy, create ordinary staff, project manager, and finance test accounts. Ask each the same question using both direct terms and paraphrases. Verify that unauthorized documents, snippets, and citations never appear. Test a user whose project access was just removed. A successful answer for an administrator says little about whether the system preserves the boundaries ordinary staff depend on.
Workflow at a glance
Check every retrieval route
The same authorization rule must cover search, document preview, export, background tasks, and support tools. A protected chat interface can still leak through an unrestricted source URL. Track how permission changes propagate to indexes and caches, and choose a failure mode when current permissions cannot be confirmed. Do not assume the platform’s preview feature has the same guarantees as a generally available integration.
Keep ownership operational
Assign an owner for group membership, document classification, and access-removal incidents. During handover, demonstrate one allowed query and one denied query with the customer. Keep the test repeatable without using live sensitive records. Permission-aware knowledge is not a checkbox added after the assistant works; it is part of defining what working correctly means for every user and every retrieval path.
An access-removal rehearsal
Use synthetic project documents to remove a staff account from a group and measure when every retrieval route reflects the change. Test previously opened conversations as well as new searches. Document any delay and the operational response rather than promising immediate revocation without evidence.
Sources and further reading
- Document-level access control in Azure AI SearchDocument-level access control in Azure AI SearchMicrosoft · Documentation ·
- Security best practices for Azure AI SearchSecurity best practices for Azure AI SearchMicrosoft · Documentation ·
Editorial approach
AI-assisted original writing, checked against linked public sources. Practical recommendations and illustrative scenarios are editorial analysis, not reported client results. Source dates and product details may differ from this collection’s preparation date.