Search

Every Kura site ships with search built in. It comes in two flavors, and you choose per project.

Lexical (default)

Out of the box, search is lexical: a zero-dependency keyword scan over your content. It needs no model, no native modules, and no services — so a new site installs clean and deploys to Cloudflare Workers (or anywhere) with nothing to provision. For most docs sites, this is enough.

Semantic (opt-in)

Add an embedder to upgrade the same search box to semantic retrieval — matching by meaning, not just keywords. Kura's retrieval engine is pure JavaScript with zero dependencies and runs the same code on Node, Bun, Deno, and Cloudflare Workers.

import { Kb } from "@kurajs/core";
import { transformers } from "@kurajs/transformers";

const kb = new Kb({ embedder: transformers() });
await kb.addText([{ id: "intro", text: "..." }]);
const hits = await kb.searchText("How do I get started?", { topK: 5 });

At index time, documents are embedded into 1024-dim vectors with the bge-m3 model; at query time the question is embedded too, compared by similarity, and the most relevant passages are returned. On the edge, the same model runs through Cloudflare Workers AI (@cf/baai/bge-m3).

Tiered strategy

• Small corpora (≤ 10k): exact f32 brute force, 100% recall.
• Large corpora: a binary prefilter + f32 rerank — 1/32 the memory, with recall still near perfect.

Performance

Query latency is roughly 0.1 ms per 1k vectors; on a real 200k-embedding corpus, binary + rerank still holds 100% recall.