inputx-wubi 1.0.2

Self-developed Wubi 86 encoder, dictionary, and dataset (PHF + FST, WASM-ready). Powers the Inputx IME.
Documentation

inputx-wubi

Self-developed Wubi 86 (五笔字型) encoder + dictionary for Rust. PHF + FST backed, with a built-in L0 / L1+ ranking model and per-user auto-learning. WASM-ready via the companion inputx-wubi-wasm crate.

Powers the Inputx IME on iOS and the web — this crate is the standalone, reusable Wubi engine, also publishable to crates.io for any downstream that wants a clean, permissively-licensed Wubi stack.

License: MIT OR Apache-2.0 (dual). The dictionary data derives from the public Wubi 86 standard (王永民, 1986); no GPL or LGPL data is embedded.

Read this in 简体中文 · 日本語.

What's in the box

  • 135,822 FST entries:
    • 25 一级简码, 616 二级简码, 5,173 三级简码
    • 53 hand-curated 字根 + 70,317 algorithmically-decomposed CJK chars
    • 61,205 phrases (词组)
  • Wubi 86 encoder, all four canonical rules
  • L0 / L1+ ranking — immutable layered lexicon (Auto < Phrase < Zigen < Jianma3 < Jianma2 < Jianma1 by base weight) plus a mutable per-user override layer with a 3-pick auto-promotion rule
  • Layer prefs — host-tunable multipliers per layer
  • Reproducible weight pipeline (wubi-build-weights) with CI byte-diff verify

Quick start

# Cargo.toml
[dependencies]
inputx-wubi = "1.0"

use wubi::WubiDict;

let dict = WubiDict::embedded();

// Lookup is L0 + layer + freq ranked.
let candidates = dict.lookup("khlg");
// → ["中国", "跑车", "跨国", "䟧", ...]

// Hot loop (IME use case): reuse the buffer.
let mut buf = Vec::new();
dict.lookup_into("ipbf", &mut buf);

// Tell the dict the user picked a candidate. After 3 picks of the same
// (code, word), it's auto-pinned to L0. Pin/forget/layer-prefs APIs are
// also exposed for explicit host control.
dict.record_pick("khlg", "跑车");

The crate name on crates.io is inputx-wubi, but the lib name is wubi for ergonomic imports — use wubi::... works directly.

Performance (Apple Silicon, release)

Op Latency
字根 / 一级简码 PHF lookup ~10 ns
dict.lookup (1–6 cand) 270–620 ns
dict.lookup miss ~145 ns
dict.prefix (~5K cand) ~1.3 ms
Encoder 8–15 ns

Tools (--features tools)

  • wubi-fetch-corpus — download corpora declared in data/corpus/manifest.toml, SHA-verified, cached locally
  • wubi-build-weights — scan corpora, derive data/weights/weights.tsv
    • data/weights/provenance.toml. verify mode for CI byte-diff.
cargo run --features tools --release --bin wubi-build-weights
cargo run --features tools --release --bin wubi-build-weights -- verify

License

Dual-licensed under MIT OR Apache-2.0 © 2026 GOLIA K.K., at your option.