Chaza

A minimal static-site search engine — Zig + WASM. Finds Korean text even when you type only initial consonants (choseong).

Live demo | [한국어]

Features

• Small. The index stores binary fuse filters, not document text. ~0.5 MB for a few hundred pages.
• Korean choseong search. Type ㄱㄴ and match 가나, 강남, 경남…
• All query work in WASM. JavaScript only passes strings and renders results.
• No toolchain for end users. The runtime WASM is embedded in the CLI binary. Building an index is just byte concatenation — no compiler needed.

vs tinysearch

Chaza is inspired by tinysearch and reads the same corpus format. Same 100-document Wikipedia corpus (Korean + English), both indexing title + body, Apple M1 Max, Node v22:

	chaza	tinysearch 0.10
Index build time	7.8 ms	4.6 s	~600× faster
Search latency	5.2 µs/query	313 µs/query	~60× faster
Output size (gzip)	25 KB	69 KB	~3× smaller
Recall@20	96.7%	87.5%
Known-item MRR@10	0.99	0.98
False positives	0.39%/doc	0.27%/doc
Korean choseong search	✅ (100% retrieval)	❌
Single binary, no toolchain	✅	❌ needs Rust + wasm toolchain

Full results (500/1,000-doc scaling, accuracy methodology, honest caveats): bench/RESULTS.md. Reproduce with bench/.

Install

npm:

npm install chaza-cli

Shell (standalone binary):

curl -fsSL https://raw.githubusercontent.com/taggon/chaza/main/scripts/install.sh | sh

Platform binaries are auto-selected for macOS and Linux (x64, arm64).

Quick start

# Build a search index from a JSON corpus
npx chaza build corpus.json -o chaza.bundle --config chaza.json

This produces two files:

File	Description
chaza.bundle	Bundle: [runtime.wasm][index][tail-meta 16 B]. Not a pure WASM module — load via the loader.
chaza.js	ESM loader shared by all sites.

Add --no-js to skip writing the loader.

In your page

<script type="module">
  import { Chaza } from "chaza-cli";
  const chaza = await Chaza.load("./chaza.bundle");
  const results = chaza.search("ㄱㄴ");
</script>

Each result contains { title, url, meta, hits } — hits is the number of query tokens that matched (0~16).

CLI

chaza build <corpus.json> [options]

Options:
  -o, --output <path>      Output bundle path (default: chaza.bundle)
  --config <path>          Path to chaza.json config file
  --stopwords <path>       Stopwords file (replaces built-in default;
                           empty file disables removal)
  --no-choseong            Disable choseong search
  --no-js                  Skip writing chaza.js loader
  -q, --quiet              Suppress progress output
  -h, --help               Show this help

Input format (tinysearch-compatible)

[
  {
    "title": "Post Title",
    "url": "https://example.com/posts/1",
    "body": "Full text to index (not stored in the output)",
    "path": "/posts/1",
    "date": "2026-01-01"
  }
]

• title — indexed + stored (shown in results)
• url — stored only (shown in results, not searched)
• body — indexed only (searched but never stored in the output)
• Any other fields listed in metadata_fields — stored only, as strings

Numbers are stringified automatically.

Config file (chaza.json)

{
  "schema": {
    "indexed_fields": ["title", "body"],
    "metadata_fields": ["path", "date"],
    "prefix_fields": ["title"],
    "url_field": "url"
  },
  "korean": {
    "choseong_search": true,
    "choseong_max_len": 3
  }
}

Field	Default	Description
indexed_fields	["title"]	Fields to tokenize and index
metadata_fields	[]	Fields to store for display (not indexed)
prefix_fields	["title"]	Fields whose words also match by prefix (2–8 chars) while typing. Must be a subset of indexed_fields; [] disables
url_field	"url"	Which field holds the URL
choseong_search	true	Enable choseong prefix tokens
choseong_max_len	3	Max choseong prefix length (1–3)

Search behavior

• Multi-token queries are OR with hit-count ranking — documents matching more tokens come first; each result carries the count as hits. On equal hits, documents whose title matches rank above body-only matches.
• The last query token also matches by prefix (2–8 chars) against prefix_fields words — progr finds a title containing "Programming".
• Choseong queries like ㅅㅈ match documents with a word starting with those initial consonants.
• max_results defaults to 20: chaza.search("hello", { maxResults: 10 }). Results are ordered by hits; for other orders (date, etc.), sort the returned array in JavaScript.

Limitations

• Input must be UTF-8 + NFC. NFD (decomposed) Hangul breaks choseong extraction.
• False positives. ~0.4% of lookups may match a document that doesn't contain the token — inherent to the filter. Practical corpus ceiling is a few thousand documents.
• Static index. No incremental updates; rebuild to add or remove documents.
• Stopwords are not searchable. A query made up entirely of stopwords returns nothing.

Documentation

• How it works — tokenization pipeline, choseong/prefix tokens, binary fuse filters, bundle format, size limits
• SPEC.md — full format and behavior specification

License

MIT