MrHendy/bge-m3

BGE-M3 In-Browser Retrieval Demo

A zero-compute, fully client-side demo of `BAAI/bge-m3`. The model (ONNX build `Xenova/bge-m3`) is downloaded and run entirely in your browser with 🤗 transformers.js, using WebGPU when available and WebAssembly otherwise. No server, no GPU quota.

What it does

Enter a query and a set of passages; the app embeds them with BGE-M3 and ranks the passages by dense semantic similarity (cosine). BGE-M3 is multilingual (100+ languages), so cross-lingual queries work too.

Scope: in-browser inference gives BGE-M3's dense embedding. Its other two signals — learned sparse/lexical weights and ColBERT multi-vectors — require the Python FlagEmbedding library and a compute backend, so they aren't part of this static demo.

First run downloads the quantized model weights (cached by the browser afterwards).