@kurajs/search

Portable, zero-dependency retrieval primitives for Kura: BM25 keyword search, Reciprocal Rank Fusion (to blend keyword + semantic), and a pluggable, per-locale tokenizer layer. Runs identically on Node, Bun, Deno, Cloudflare Workers, and the browser.

BM25

import { Bm25 } from "@kurajs/search";

const bm = Bm25.from([
  { id: "a", text: "vector search engine", data: { url: "/a" } },
  { id: "b", text: "database indexing and query planning", data: { url: "/b" } },
]);

bm.search("search", { topK: 5 }); // → [{ id, score, data }, ...]

Hybrid (RRF)

rrf / rrfScored fuse several ranked lists by rank, so a BM25 score and a cosine similarity don't need to be comparable:

import { rrfScored } from "@kurajs/search";

const fused = rrfScored(
  [{ hits: keywordHits }, { hits: semanticHits }],
  (h) => h.slug,
  { topK: 8 },
); // → [{ item, score }, ...] in fused order

Per-locale tokenizers

Bm25 accepts a single tokenize function or a resolveTokenizer that picks one by language, so one index can tokenize each document by its own locale (and each query by the query locale). byLocale() builds the resolver (case-insensitive, with primary-subtag fallback). CJK tokenizers live in @kurajs/tokenizers:

import { Bm25, byLocale, latinTokenizer } from "@kurajs/search";
import { cjkSegmenter } from "@kurajs/tokenizers";

const bm = Bm25.from(records, {
  resolveTokenizer: byLocale({
    default: latinTokenizer,
    zh: cjkSegmenter("zh"),
    ja: cjkSegmenter("ja"),
  }),
});

Analyzer pipeline + 繁/簡 normalization

pipeline() composes char filters → a segmenter → token filters into a Tokenizer. Char filters are just (text) => string, so anything of that shape drops in — including OpenCC's converter, which already has that signature. That's all you need to fold Traditional/Simplified Chinese (script and regional vocabulary, e.g. 软件↔軟體) so a keyword query matches across variants — no extra Kura package required:

import * as OpenCC from "opencc-js"; // npm i opencc-js  (MIT; bundled dictionary data is Apache-2.0)
import { pipeline } from "@kurajs/search";
import { cjkSegmenter } from "@kurajs/tokenizers";

// Fold everything to Traditional-Taiwan (idempotent on text already in that variant),
// then segment. Run the SAME pipeline at index and query time so terms line up.
const zhTW = pipeline({
  pre: [OpenCC.Converter({ from: "cn", to: "twp" })],
  segment: cjkSegmenter("zh-TW"),
});

// In a Kura docs app: defineKura({ tokenizer: byLocale({ "zh-TW": zhTW }) })

OpenCC ships ~5 MB of dictionaries, so only sites that need cross-variant keyword matching pull it in. Most sites don't: a single-variant corpus is already consistent, and the hybrid vector half bridges 繁/簡 semantically on its own.

Exports

• Bm25, rrf, rrfScored
• Tokenizer toolkit: latinTokenizer, byLocale, pipeline, lowercase, minLength, stopwords
• Types: Tokenizer, TokenizerResolver, CharFilter, TokenFilter, Bm25Record, Bm25Hit, …