Search — Quickstart

Go from "I have a table" to "I have semantic search" in four moves. No queue, no embedding service, no pipeline — the lens config is the whole story.

searcharticles

articles: {
  source: 'main',
  table: 'articles',
  fields: {
    id:    { type: 'number', primaryKey: true },
    title: { type: 'text',   searchable: { weight: 2 } },
    body:  { type: 'text',   searchable: true },
    tags:  { type: 'text' },   // stored, not embedded
  },
  grants: { search: 'public' },
}

ℹ️

Note the searchable flag. title and body opt in to embedding — title with a weight of 2 contributes twice as strongly to the ranking vector. tags is stored as metadata (you get it back in results) but isn't vectorized. Forget searchable and the field is invisible to search.

Four moves

Declare the lens

Add the block above to your semilayer.config.ts. searchable: true (or { weight: N }) opts a field into embedding; omitting the flag leaves the field stored-but-not-vectorized. Only fields with searchable participate in ranking.

Push

semilayer push

This validates the config, provisions the vector partition, and creates a paused lens. Nothing is embedded yet — that happens when you resume ingest.

Ingest

semilayer resume --lens articles

The worker streams rows from your source via the configured bridge, embeds the searchable fields, and writes vectors to the index. Watch progress with semilayer status --lens articles.

Generate Beam and call it

semilayer generate

This writes a typed client to ./beam.ts. Then in your app:

import { beam } from './beam'

const { results } = await beam.articles.search({
  query: 'how do I rotate my api keys',
  limit: 3,
})

Done. results is typed SearchResult<ArticlesMetadata>[] — fully inferred from your config. Ranking is cosine similarity across all searchable fields, weighted by the weight you declared.

What you skipped

If you've built semantic search by hand, here's what the four moves above replaced:

An embedding service + retry/backoff/idempotency
A vector store migration + index tuning
A changefeed listener (for incremental sync)
A ranking endpoint with auth, rate limits, and pagination
A typed client

Next up: Fields & weights — the searchable flag in depth.