Vector Search: Tenant Isolation
Question Variations
- "Why is prompt-only access control insufficient?"
- "When would separate indexes be preferable to filtered namespaces?"
- "How would you prevent cross-tenant retrieval leakage in a shared vector index?"
Why This Is Asked
This evaluates whether a candidate treats retrieval as a data-access boundary, not merely a search problem. Interviewers look for authorization enforced before context reaches the model, trustworthy metadata, defense in depth, and adversarial testing.
Key Concepts
- Authorization filter: Enforce tenant and document permissions before returning context.
- Defense in depth: Validate access in the application even if the store supports filters.
- Metadata integrity: Index trusted access attributes with every vector.
- Testing: Include cross-tenant adversarial queries in evaluation.
Question Variations
- “Why is prompt-only access control insufficient?”
- “When would separate indexes be preferable to filtered namespaces?”
- “How would you prevent cross-tenant retrieval leakage in a shared vector index?”