Creating and querying a Vector Search index

Databricks Vector Search lets you create a self-updating vector index over data stored in a Delta table, enabling similarity search for RAG and other GenAI applications. You create an index via the Vector Search UI, SDK, or API by pointing it at a source Delta table and an embedding model (or precomputed embedding column), then query it using similarity_search to retrieve the most relevant chunks/documents for a given query.

1 · Learn the must-know

• 1A Vector Search endpoint must exist first (compute layer) before you can create an index on top of it.
• 2There are two index types: Delta Sync Index, which automatically syncs with changes to the source Delta table, and Direct Vector Access Index, which is managed manually via API without a source table sync.
• 3The source Delta table must have Change Data Feed enabled for Delta Sync indexes to track updates, inserts, and deletes.
• 4You can either provide precomputed embeddings in a column or specify a Databricks-hosted embedding model endpoint so the index computes embeddings automatically from a text column.
• 5Querying is done via the similarity_search API (or SDK/REST), specifying the query text or vector, number of results (num_results), and optionally metadata filters.
• 6Vector Search indexes and endpoints are governed by Unity Catalog permissions, so access control follows standard UC grant/revoke on the underlying catalog/schema/table objects.

2 · Check your understanding