Getting started

Seekite builds portable search metadata and a lexical index, then either ships precomputed corpus vectors or creates them on demand in the browser.

Install

pnpm add @seekite/core @seekite/vite @seekite/embeddings-ternlight
pnpm add -D seekite

Static vectors in CI

Use an application-owned loader and write the index below Vite's public directory:

// search.config.ts
import { seekiteEmbeddings } from "@seekite/embeddings-ternlight"
import { defineSearch, defineSource, staticVectors } from "seekite"

const docs = defineSource("docs-api", async () => {
  const response = await fetch(process.env.DOCS_SEARCH_URL!)
  return response.json()
})

export default defineSearch({
  output: "public/search",
  corpora: {
    docs: { source: docs, vectors: staticVectors(seekiteEmbeddings()) },
  },
})

Run seekite build before vite build in CI. Configure Vite with seekite({ indexing: "external" }); it validates the manifest and lets Vite copy public/search into the final public bundle.

Runtime vectors

For a small corpus, replace the vector declaration:

import { runtimeVectors } from "seekite"

docs: { source: docs, vectors: runtimeVectors(seekiteEmbeddings()) }

Use regular seekite() Vite integration. No vectors.bin is emitted; the browser creates and caches document vectors only after its first semantic or hybrid query.

Search

import { createSearch } from "@seekite/core"
import { seekiteEmbeddings } from "@seekite/embeddings-ternlight"

const search = createSearch({
  key: "search",
  embeddings: seekiteEmbeddings(),
})

const response = await search.query("canvas drawing", {
  corpora: ["docs"],
  mode: "hybrid",
  limit: 8,
})

for (const result of response.results) {
  console.log(result.title, result.score)
}

The client is plain TypeScript/JavaScript and does not require a UI framework. key: "search" resolves to <Vite BASE_URL>/search internally; because search is also the default key, it may be omitted. Use url: "https://cdn.example.com/index" when assets are hosted separately. Explicit URLs must be absolute or root-relative; relative locations are keys.

For search-as-you-type UIs, move the same client into a worker with workerSearch, or install @seekite/react for an accessible dialog and inline search box.