Sizing a Vector Search configuration for embeddings, update frequency, latency, and cost

Databricks Vector Search index and sync-mode choice should be driven by how many embeddings you have, how often they change, and your latency/cost tolerance. A Delta Sync Index (auto-managed from a Delta table) is preferred for most RAG workloads, while a Direct Vector Access Index gives full manual control over embedding writes when data isn't Delta-native or updates need custom logic.

1 · Learn the must-know

• 1Delta Sync Index automatically keeps the vector index in sync with a source Delta table and supports Continuous, Triggered, or Scheduled sync modes.
• 2Continuous sync gives near-real-time freshness but keeps compute running constantly, making it the most expensive option—use it only when low update latency truly matters.
• 3Triggered or Scheduled sync modes batch updates (on-demand or at set intervals) and cost less, making them appropriate when data changes infrequently or slight staleness is acceptable.
• 4Direct Vector Access Index requires you to manage embedding generation and index updates yourself via the API, offering flexibility for custom or non-Delta pipelines at the cost of more operational overhead.
• 5Vector Search endpoints incur cost based on provisioned compute uptime, not per query, so endpoint sizing (and thus cost) should scale with index size (number of embeddings) and expected query throughput/latency needs.
• 6Larger embedding counts or higher-dimensional embeddings increase both index storage and query latency, so right-sizing the endpoint and periodically reviewing index growth is a key configuration step.

2 · Check your understanding