Privacy and self-hosting

What leaves the browser

• Queries: nothing. Search runs entirely on the visitor's device. There is no search API and no tracking.
• The index is loaded from your own site.
• The model is downloaded from huggingface.co the first time the search field gets focus, then served from the browser cache. This request reveals the visitor's IP address to Hugging Face, like any third-party asset.

If that request is a problem under the GDPR or the Swiss FADP for your site, serve the model yourself.

Serving the model from your own domain

Copy the model files (config.json, tokenizer.json, tokenizer_config.json, onnx/model_quantized.onnx) from the model page to e.g. public/models/Xenova/all-MiniLM-L6-v2/, then point transformers.js there:

import { Cosine, loadIndex, transformersEmbedder } from "@sweberdev/cosine";

const embedder = transformersEmbedder({
  model: "english",
  load: async () => {
    const transformers = await import("@huggingface/transformers");
    transformers.env.allowRemoteModels = false;
    transformers.env.localModelPath = "/models/";
    return transformers;
  },
});
const cosine = await loadIndex("/cosine/cosine-index.json", { embedder });
document.querySelector("cosine-search").cosine = cosine;

The ONNX runtime itself (WebAssembly) is loaded from a CDN by transformers.js; set transformers.env.backends.onnx.wasm.wasmPaths to serve it yourself as well.

Keyword search only

<cosine-search load-model="never"> or an index built with --lexical-only never contacts a third party.