Wisdom-Weasel enhances the Rime/Weasel Chinese input method by integrating Large Language Model (LLM) powered intelligent prediction. It targets users seeking advanced, context-aware text completion beyond traditional input methods, offering a smarter typing experience while preserving Rime's flexibility and extensive ecosystem.

How It Works

This project augments the Rime input method framework by incorporating LLM-based candidate word generation. It maintains Rime's full scheme and dictionary support, using LLMs to predict candidate words based on current input and historical context. The system supports multiple LLM backends, including OpenAI-compatible APIs (like Ollama), local llama.cpp inference with GGUF models, and a Python backend for pinyin-constrained generation. It actively manages user input sequences as context and offers optional asynchronous memory compression for extended histories.

Quick Start & Requirements

• Primary install: Download the installer from the Releases page and install it. Configuration is done via weasel.yaml in the Rime user directory.
• Prerequisites:
  • Rime/Weasel must be installed and deployed.
  • OS: Windows 8.1 ~ Windows 11.
  • LLM Backends:
    • OpenAI-compatible: Requires an accessible API endpoint or local service (e.g., Ollama).
    • llamacpp: Requires a local GGUF model file (4GB+ VRAM or sufficient RAM recommended).
    • hf_constraint: Requires a Python environment set up using hf_backend/requirements.txt.
• Links: Rime Project Homepage: https://rime.im

Highlighted Details

• Supports multiple LLM prediction backends: OpenAI-compatible, llama.cpp (local GGUF), and hf_constraint (Python).
• Maintains user input history for context and offers optional LLM-based memory compression.
• Predictions are displayed alongside Rime's native candidates, ensuring seamless integration.
• Enables local LLM inference for privacy and offline use.

Maintenance & Community

Issues specific to Wisdom-Weasel's LLM features, build, or configuration should be reported in the project's GitHub Issues. General Rime input method issues should be directed to Rime Home. Pull Requests are welcomed.

Licensing & Compatibility

• License: GPLv3.
• Compatibility: The GPLv3 license is a strong copyleft license, requiring derivative works to also be licensed under GPLv3. This may impose restrictions on linking with closed-source applications.

Limitations & Caveats

The project is Windows-only, supporting Windows 8.1 through Windows 11. It has a hard dependency on a pre-installed and configured Rime/Weasel input method. Setting up the LLM backends requires specific user configuration, such as providing API keys, local model paths, or Python environments.