togatoga / karukan

项目摘要

Karukan is a Rust-based Japanese Input Method System for Linux and macOS that uses a GPT-2-based neural model running on llama.cpp for Kana-to-Kanji conversion. It integrates with fcitx5 on Linux and InputMethodKit on macOS, offering live conversion, context-aware prediction, user learning, emoji input, and rewrites ported from Mozc. The project is structured as a multi-crate workspace covering the neural engine, shared IM state machine, and platform-specific frontends.

为什么值得关注

It is one of the few open-source Japanese IMEs to integrate a genuine neural language model (GPT-2 via llama.cpp) for kana-kanji conversion, a task traditionally handled by n-gram or rule-based systems like Mozc. The combination of a modern Rust codebase, cross-platform support (Linux fcitx5 + macOS InputMethodKit), live conversion, and on-device neural inference makes it a standout project in the Japanese NLP tooling space.

优势

• Genuinely innovative use of a neural language model (GPT-2/llama.cpp) for kana-kanji conversion, going beyond traditional n-gram approaches
• Cross-platform support for both Linux (fcitx5 plugin via C FFI) and macOS (Swift/InputMethodKit) from a shared Rust core engine
• Rich feature set including live conversion, context-aware prediction, user learning, emoji input (both kana-triggered and Slack-style), and Mozc-derived candidate rewrites
• Well-organized multi-crate Rust workspace with clear separation of concerns (engine, IM logic, platform frontends, CLI tools)
• Multiple CI pipelines covering engine, IM, fcitx5, and macOS builds with tests

局限性

• Single contributor raises sustainability and bus-factor concerns for a project of this complexity
• Documentation is primarily in Japanese, limiting accessibility for international contributors and users
• Relatively young project (created Aug 2024) with only 7 open issues tracked, suggesting early-stage maturity; no Docker support or extensive examples

使用场景

• Japanese-language developers on Linux or macOS seeking a modern, neural-powered IME alternative to Mozc or Anthy
• Users who want on-device, privacy-preserving kana-kanji conversion powered by a local neural model
• NLP researchers studying neural approaches to Japanese text input and conversion evaluation (AJIMEE-Bench integration)