This tests security architecture. The interviewer wants to know about data isolation strategies — physical separation, logical separation, metadata filtering, and vector DB-level controls.
Askgenai In Answered question
-
Metadata Filtering: Store
tenant_idas metadata in vector DB. Every query filters bytenant_idbefore similarity search. Supported by Pinecone, Weaviate, pgvector. - Namespace/Partitioning: Use separate namespaces (Pinecone) or collections (Milvus) per tenant. Best for strict isolation.
- Row-Level Security: In pgvector, use PostgreSQL RLS policies so tenants can only query their own rows.
- Embedding Isolation: Never share embedding spaces across tenants. Even with metadata filters, shared indexes risk side-channel attacks.
- API Gateway: Enforce tenant scoping at the API layer. Validate JWT tenant claims before any retrieval call.
Askgenai In Answered question