Docs

How this works, what it can't claim, and what would prove the privacy claim false.

How it works

A pasted note is split into sentence-bounded chunks (never more than 3 sentences or 80 subword tokens, whichever comes first), each chunk is embedded into a 384-dimension vector entirely on-device, and the whole current set of chunk embeddings is projected onto its own top-2 principal components to draw the map. That last step — project — reruns from scratch on every add, edit, or remove (the amber edge above), which is why the map's own caption says a note's position is relative to what you've pasted and will shift, not a fixed measurement. Search ranks over the finer-grained chunks; the map, clustering, and outlier detection operate on each note's centroid (the mean of its own chunks' embeddings) — two different granularities, used for two different questions, never conflated.

Limitations

The single most important limitation is demonstrated live, not just described: /limits shows that this technique measures how similar two sentences' wording is, not whether they mean the same thing — a flat contradiction can score as similar as, or more similar than, a genuine paraphrase. Nothing on this site should be read as the model detecting agreement, confirmation, or factual consistency.

• Language. The default model is English-tuned; results in other languages will be worse, sometimes much worse. See coverage for exactly which languages the optional multilingual model was tuned on, and which of those this project has actually re-verified versus simply cites from the model's own published list.
• The map is 2D, the model isn't. Clustering and outlier detection run in the full embedding space; the map is a 2-axis projection of that space for display only. Two points landing close together on the map is a real, local signal; two points landing far apart is not proof they're unrelated — projection can compress or separate points in ways the caption under the map explicitly disclaims.
• Session-only. Notes live in this browser tab and are gone on reload — nothing is saved to a server, by design (see Privacy below). There is no account, no sync, no history.
• Caps. A session holds at most 200 notes or 200,000 characters, whichever comes first, after which new notes are refused with a clear message rather than silently dropped or truncated.
• No analytics. This build ships no page-view or usage analytics of any kind — not scoped, not anonymized, not present at all. If that ever changes, it will be named here and scoped to route changes only, never to anything typed or pasted.

Privacy, and how to prove it wrong

The claim: nothing you paste, type, search, or cluster is ever sent anywhere. Chunking, embedding, projection, search ranking, clustering, and outlier detection all run in this tab, most of them inside a Web Worker, using a model that was fetched once from a public CDN and cached by the browser afterward.

The condition that would make this claim false: any network request whose URL, query string, or request body encodes a fragment of text you actually typed or pasted. That's it — that is the one thing that isn't supposed to happen, ever, at any point in this product.

Don't take the live request counter's word for it. Open your browser's dev tools, go to the Network tab, and paste something — the only requests a cold visit makes are for this page's own static assets, the model runtime from cdn.jsdelivr.net, and the model weights themselves from huggingface.co (once, cached after). Typing, pasting, editing, searching, and clustering should never add a single row to that list. If you ever see one that does, that is a real bug — the source is public and every one of these claims is checkable against it, not just this page's word.