Capabilities

• Embedded or application-managed LanceDB connections.
• Dense vector search with cosine, L2, or dot distance configuration.
• search through the common vector index; agent tool integration via createVectorSearchTool() from @anvia/core/vector-store.
• Multiple physical embeddings collapsed to one logical document result.
• All metadata serialized into a single __anvia_metadata JSON column.
• Post-retrieval filtering with eq, gt, lt, and, and or via core's matchesVectorFilter.

The adapter stores reserved columns for the row ID, document ID, serialized document, metadata, and vector. Metadata keys beginning with __anvia_ are rejected.

It does not create ANN indexes, compact tables, or delete stale versions; those responsibilities remain with the application's LanceDB workflow. It does validate dimensions: validate() throws when the persisted __anvia_vector column's list size differs from the configured dimensions.