Capabilities ​
| Mode | Input | Retrieval |
|---|---|---|
| Dense | Dense embeddings | Qdrant universal query, with legacy search fallback for custom clients |
| Hybrid | Dense and sparse embeddings | Prefetched named-vector results fused with RRF or DBSF |
Both modes support deterministic document replacement, logical document deletion and retrieval, paginated inspection, metadata filters, multiple embeddings per logical document, and search, with retrieveDocuments and createVectorSearchTool integration from @anvia/core/vector-store. Mutations accept Qdrant wait, ordering, and timeout controls.
Dense-only and hybrid collection/index modes cannot be mixed. Hybrid defaults to named vectors dense and sparse; configure matching names across creation, ingestion, and query.
Constructing a store performs no I/O; ensure() and validate() check existing collections against the configured vector size, distance, and dense/hybrid shape. The adapter does not manage aliases, snapshots, replicas, shards, payload indexes, or client lifecycle. See the API reference.