Restrict AI access before generating an answer

A knowledge assistant broadens access to information. Preserving existing permissions requires source selection to account for the user's identity and authorized data scope. An instruction to the model alone does not replace access control.

Verify the entire data path

• Restrict documents during search and retrieval.
• Account separately for tenant, role and document permissions.
• Prevent cached results from being shared between users without authorization.
• Apply revoked access to indexes and stored responses as well.

A useful test asks the same question through two accounts with different permissions. The less privileged person must receive neither confidential source text nor details derived from it. Check source references too.

Keep authorization decisions traceable outside the language model. The model can then help formulate an answer while the application determines which information may enter its context.