hojin-rs

Compact, read-only search dictionaries for Japanese corporate names and corporate numbers.

hojin provides exact normalized-name lookup, address-aware ranking, corporate-number lookup, and prefix suggestions. Native indexes are memory-mapped, while variable-size records use independently compressed zstd frames. The same compact suggestion bundle can be searched from WebAssembly in browsers and Node.js.

Features

• standalone Rust crate with no application-framework dependency
• Unicode, legal-form, legacy-character, and Japanese-address normalization
• exact corporate-name lookup with address-aware ranking
• direct 13-digit corporate-number lookup
• alternate name, English name, and furigana indexes
• mmap-backed FST, compact fixed-width indexes, and delta-varint postings
• block-addressable zstd record storage
• compact prefix-suggestion bundle for WASM/npm
• optional adapter for NTA headerless Unicode CSV files

Native use

[dependencies]
hojin = "0.1"

use hojin::CorporateDictionary;

let dictionary = CorporateDictionary::open("/opt/dictionaries/corporations-2026-06")?;
let candidates = dictionary.resolve_exact(
    "（株）ＡＢＣ",
    Some("東京都千代田区丸の内１－１－１"),
)?;

let corporation = dictionary.get_by_corporate_number(7_000_012_050_002)?;
# Ok::<(), anyhow::Error>(())

Registered addresses are never overwritten by query text. Exact name-and-address matches rank first, followed by compatible address containment and name-only matches. Active corporations win ties.

Build from NTA CSV

Download and extract the Unicode CSV files from the NTA full-data download page, then run:

cargo run --release --features cli --bin build-hojin-dictionary -- \
  --output /tmp/corporations-2026-06 \
  /path/to/00_zenkoku_all_*.csv

See DATA-LICENSE.md before redistributing NTA-derived artifacts. Source data and generated nationwide dictionaries are deliberately excluded from the Rust crate, npm package, and repository.

WASM/npm suggestions

Build a browser package:

wasm-pack build crates/hojin-wasm \
  --release --target web --scope bokuweb --out-dir pkg-web

Build a Node.js package:

wasm-pack build crates/hojin-wasm \
  --release --target nodejs --scope bokuweb --out-dir pkg-node

import init, { CorporateSuggester } from "@bokuweb/hojin-wasm";

await init();
const response = await fetch("/dictionary/2026-06/suggestions.bin");
const suggester = new CorporateSuggester(
  new Uint8Array(await response.arrayBuffer()),
);

const candidates = suggester.suggest("鳥取簡易", {
  limit: 10,
  includeClosed: false,
});

Corporate numbers are returned to JavaScript as 13-digit strings. The npm package contains code only; versioned dictionary bundles should be served separately, preferably with Content-Encoding: br and optional prefecture or leading-character shards.

Artifact format

File	Purpose	Access
manifest.json	schema, source/terms/files, processor version, counts, SHA-256	startup read
names.fst	normalized name to posting offset	mmap
numbers.idx	48-bit corporate number to record ID	mmap + binary search
postings.bin	delta-varint record IDs	mmap
records.idx	compact record-to-block location	mmap
blocks.idx	zstd frame offsets and lengths	mmap
records.zst	independently compressed record blocks	mmap + selective decode
suggestions.bin	lookup FST, postings, front-coded display/address tables	WASM memory or native bytes

Do not overwrite dictionary files while readers have them mapped. Build a new versioned directory and atomically switch a symlink or configuration pointer.

Size reference

Measured with the NTA Tottori Unicode snapshot used during development: 20,153 corporations and 29,422 suggestion entries.

Artifact	Size
complete dictionary directory	3,007,399 bytes
suggestions.bin	1,355,591 bytes
suggestions.bin with Brotli level 11	586,915 bytes
release WASM	126,763 bytes
npm tarball, without dictionary data	61,650 bytes

These measurements are reference points, not guarantees for other source snapshots.

License

Source code is licensed under the MIT License. NTA source data and artifacts derived from it are not relicensed under MIT; see DATA-LICENSE.md.