# ReLU.chat

> Free, browser-based, privacy-first open-source chatbots that run entirely in your browser. No servers, no LLMs, no tracking.

## Overview

ReLU.chat is an open-source platform for interactive on-device chatbots. All NLP processing happens in the browser using a quantized ONNX sentence-transformer (all-MiniLM-L6-v2, 384-dim, ~22MB), field-weighted BM25 sparse retrieval, dense-sparse ensemble ranking, and a reinforcement-learning-trained MLP policy network. No data ever leaves the device.

## Key Pages

- [Home](https://relu.chat/) — Landing page with feature overview
- [How It Works](https://relu.chat/how-it-works.html) — Full technical architecture
- [Blog](https://relu.chat/blog/) — Technical articles on on-device AI and NLP (48 posts)

## Chatbots (try them in your browser)

- [Game Theory Chat](https://relu.chat/chat/game-theory-chat/) — On-device assistant for game theory: Nash equilibrium, Shapley value, auctions, prisoner's dilemma and 55+ topics with LaTeX math.
- [Golden Age Inquiry](https://relu.chat/chat/golden-age-inquiry/) — On-device assistant for the science and philosophy of the Islamic Golden Age (8th-14th centuries): algebra, optics, astronomy, medicine.
- [Data Science Chat](https://relu.chat/chat/data-science-chat/) — On-device assistant for data science and ML: pandas, NumPy, statistics, classification, regression, clustering.

## Architecture

- **Embedding**: all-MiniLM-L6-v2 (quantized ONNX, 384-dim)
- **Retrieval**: BM25 sparse (k1=1.5, b=0.75, field-weighted + bigrams) + dense cosine ensemble
- **Policy**: MLP 25->128->64->6 action heads (~13K params), RL-trained (REINFORCE), int8-quantized
- **Runtime**: Pure browser JavaScript, WebAssembly, ONNX Runtime, Web Workers
- **Storage**: Client-side only (IndexedDB, no server state)

## Source Code

GitHub: https://github.com/yunusemrejr/relu-chat (MIT license)

## Contact

For questions, open an issue on GitHub or visit the repository.