BriefsWorkflowDatabricksAvailable now

Databricks publishes sizing guidance for Lakebase Search, its AI vector search

Tests from an internal slide-search app and a 4 million passage corpus show how cache size, candidate counts and compute size affect search speed and cost.

Announced by DatabricksPublished Checked

Summary

Databricks used its internal Slide Hub app and a public passage corpus to test Lakebase Search, which runs vector and keyword search inside the app's Postgres database. The post gives sizing rules: match the cache to the vector index, keep the candidate pool small, and scale the database first, then the app. It also recommends a small always-on compute for production to avoid cold starts.

What changed

  • Lakebase Search runs vector and BM25 keyword search as Postgres extensions next to the app tables.
  • A cache about as large as the vector index serves 95% to 98% of pages without going to storage.
  • Cutting candidates from 1,000 to 200 keeps 90.5% of the top 10 results and reads about a third of the pages.
  • Below 8 CU, add database compute; from 8 CU, give the app more workers.
  • For production, the post suggests a 0.5 CU minimum instead of scaling to zero.

Availability

Available now

“Lakebase Search became generally available in September.”

Pricing

PlanPrice, as statedChecked
Always-on 0.5 CU minimum for production, per month at list priceabout $31 to $40 a month“At list price that costs about $31 to $40 a month, and searches no longer wait about 1 s for the compute to wake or start with an empty cache.” source

Prices as stated by the company; we don’t convert or estimate.

Founder takeaway

If you build an app on Lakebase Postgres, measure your vector index size and pick a compute size so the cache roughly matches it. Start with 200 candidates, and consider a 0.5 CU minimum in production to avoid cold starts.

Sources

  1. Databricks — Sizing Lakebase Search: lessons from a real app and 4 million passagesPublished · Checked

Researched and drafted with AI from the official sources above, then checked automatically before publishing. Spotted a mistake? Tell us.

Recent

More briefs

All briefs

  1. Published Tool

    Beta

    Databricks launches Predictive AI Functions in Beta for SQL, Genie and dashboards

    Databricks adds four new SQL AI functions and an upgraded ai_forecast so analysts can forecast, find drivers and flag anomalies without building models.

    Databricks · Announced

  2. Published Tool

    Available now

    Amazon SageMaker Studio adds social sign-in for new AI/ML builders

    New AI/ML builders can sign up to SageMaker Studio with Google, GitHub or Apple credentials, without a credit card or a traditional AWS account setup.

    AWS · Announced

  3. Published Tool

    Available now

    Databricks makes Document Intelligence AI REST APIs generally available

    Databricks says its AI REST APIs for parsing, classifying and extracting documents are now generally available, with a routing API in beta.

    Databricks · Announced

Briefs are published here every weekday. Monday Heist is the weekly email: one useful AI workflow, the files to use it, and the catch. Free, and you can leave in one click.

One email. Every Monday.