<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> LIME is an AI-driven pinyin input method using LLMs to enhance text prediction. It targets users seeking intelligent typing, aiming for traditional engine speed with LLM context. The project integrates a small LLM with user pinyin to guide generation, offering a novel input method approach.

How It Works

This project employs a compact LLM (Qwen3-0.6B-IQ4_XS) for pinyin input. Unlike standard LLMs, LIME uses user pinyin to influence the model's sampling strategy, enabling interactive text generation. Built with TypeScript and Deno.js, it integrates as a custom input method within the RIME framework, balancing efficiency and predictive text.

Quick Start & Requirements

• Installation: Clone repo (git clone https://github.com/xushengfeng/lime). Install Deno runtime. Download Qwen3-0.6B-GGUF model (git clone https://www.modelscope.cn/unsloth/Qwen3-0.6B-GGUF.git) to project sibling directory. On Linux, install luasocket (sudo luarocks install luasocket LIBFLAG="-shared -llua").
• Setup: Generate key (deno run -A key.ts), embed in llm_pinyin.lua. Copy project's rime folder contents to RIME config.
• Running: Start server (deno serve -A --port 5000 server.ts). Switch to "llm pinyin" in RIME.
• Prerequisites: Deno.js, RIME, luasocket, Qwen3-0.6B-GGUF model.
• Links: Deno: https://deno.land/, Model: https://www.modelscope.cn/unsloth/Qwen3-0.6B-GGUF.git

Highlighted Details

• AI-optimized pinyin input using LLMs.
• Small LLM (Qwen3-0.6B-IQ4_XS) balances speed and suggestion quality.
• User pinyin guides LLM sampling for contextual relevance.
• Supports fuzzy pinyin and Natural Code Shuangpin.

Maintenance & Community

No specific details on maintainers, community channels (e.g., Discord, Slack), or project roadmaps were found in the provided README.

Licensing & Compatibility

The provided README does not specify a software license. This omission may affect commercial use or integration into closed-source projects.

Limitations & Caveats

Does not support direct long-sentence input (phrase selection possible mid-input). Rapid typing may miss characters. Lacks data persistence; context is lost on server restart. Candidate rankings may be suboptimal. Functions as a standalone RIME input method, not combinable with others.