A good fit for
- → Mixed semantic and keyword queries
- → Metadata-filtered retrieval
- → Multi-stage candidate generation and reranking
Know the limits
- → Fusion settings still need a labeled relevance set.
- → Indexing metadata is part of performance planning.
- → A separate datastore adds backups, synchronization and operational work.
Your first useful workflow
Collect hard queries
Include exact IDs, synonyms and ambiguous phrases.
Build two retrievers
Compare lexical, dense and fused candidates on the same cases.
Measure filtered recall
Test relevance under real tenant and permission filters.
The practical take
Semantic similarity misses exact identifiers; combine lexical evidence when the task demands it.
Sources & evidence
Capabilities above are summarized from these official documents. Workflow suggestions are editorial guidance, not a claim of measured product superiority.