Explaining what Mosaic AI Vector Search is made of and how it works

Mosaic AI Vector Search is Databricks' managed vector database that lets you create searchable indexes over embeddings and retrieve semantically similar records for use in RAG and other GenAI applications. It integrates natively with Delta tables and Unity Catalog, so indexes can stay synchronized with source data and inherit the same governance model.

1 · Learn the must-know

• 1A Vector Search Endpoint is the compute resource that must exist before you can create or query any index, and multiple indexes can share one endpoint.
• 2A Delta Sync Index is linked to a source Delta table and automatically keeps embeddings up to date as the table changes, using either Continuous or Triggered (scheduled/manual) sync modes.
• 3A Direct Vector Access Index lets you read and write vectors directly via API without a backing Delta table, useful when embeddings are computed or managed externally.
• 4You can either let Databricks compute embeddings automatically using a specified embedding model endpoint (Compute Embeddings) or supply pre-computed embedding vectors yourself.
• 5Indexes and endpoints are registered in Unity Catalog under a catalog.schema.index_name namespace, so standard Unity Catalog permissions govern access to them.
• 6Similarity search returns approximate nearest neighbors ranked by a distance metric (e.g., cosine or L2), and queries can combine vector similarity with metadata filters on other Delta table columns.

2 · Check your understanding