Non-commercial Rule-based Chat Assistant (offline, Android)
A compact offline chat assistant built on rules — lightweight and designed to run on low-end devices (currently Android). The system is split into two complementary parts that produce synergy:
- Engine (runtime) — the app/handler that runs dialogue logic.
- Models (template sets) — collections of rules and patterns that define the bot’s behavior.
- Limited contextual understanding — supports synonyms, simple paraphrases, basic topic tracking, and respects punctuation and case.
- Partial memory & adaptive behavior — stores user parameters (preferences, interests) and adapts responses based on past interactions.
- Typo and simple error handling — understands inputs like
how you are? hello!as intent-equivalent tohello, how are you?. - Model file separation for efficiency — structured model files reduce memory and speed up lookups.
- TTS (text-to-speech) — local speech synthesis for answer playback.
- Response variability — multiple phrasings per scenario for naturalness.
- Flexible model customization — easily create models for different domains (e.g., mental health support, school tutoring). You can keep unlimited distinct models to optimize behavior per domain.
- Unlimited characters / mascots — switchable personas.
- Privacy-focused design — sensitive data stays local. Note: if a device is rooted, system security depends on the device/firmware.
- Integration potential — theoretical support for syncing with other apps.
- Language-agnostic engine — works with rule models in any language.
…and other features depending on the model content.
Rule-based assistants provide:
- predictable, controllable behavior;
- full offline operation on limited hardware;
- simple, focused configuration for narrow domains;
- much lower compute/memory requirements vs LLMs.
This project does not aim to replace LLMs — large models are inherently more capable. The goal is to deliver a safe, compact, and usable assistant where internet or offline LLMs are unavailable, or where deterministic replies and resource constraints are important.
Tip: combine this assistant with an LLM for broader coverage and higher quality.
We aim to provide a safe, private environment: most data is stored locally. Do not use the bot to create illegal or harmful instructions. Authors are not responsible for misuse, third-party modifications, device damage, or other consequences arising from use of this software. Use at your own risk.
Project is distributed under Apache-2.0.
The assistant’s effectiveness depends strongly on the quality of the models (the “content”). A well-designed model makes a rule-based bot practical and reliable in its target (or universal) domain.
The app is signed with a test signature due to technical limitations. If you are concerned about that, feel free to rebuild the APK manually. Since the signature itself does not protect against tampering, a test one was used for simplicity.
It is recommended to download the APK from syndes rep or from the Releases section of this repository. Even though it uses a test signature, the APK is a release build with debugging disabled.
Я буду рад любым улучшениям и дополнением. Как и остальные проекты, этот — для меня всего лишь хобби. Я надеюсь, кому-нибудь также будет интересен этот проект. Приветствую пользователей и разработчиков не только из России, но и из других стран тоже. Всего хорошего, друзья!