A query filters on tenant_id then orders by vector distance. Which index strategy is most effective?
- A A single HNSW index on the vector only
- B Partial HNSW indexes per tenant, or a composite B-tree on tenant_id plus HNSW with sufficient ef_search (correct answer)
- C A B-tree on the vector column
- D No index
Shortcut: Pre-filtering interacts poorly with approximate indexes because the filter can eliminate most graph candidates. Partitioning or partial indexes per tenant keeps recall high under filtering.