← AI tools

Taxonomy snap idle

Type anything a shopper might. Get the nearest of 860 retail category labels. The encoder downloads once (33 MB) and runs entirely in your browser — no server, no API key, nothing typed here leaves the page.

Press Load to fetch the model and index the taxonomy.

Try:

Score it yourself

All 468 WANDS queries carry a gold category. This encodes every one of them here, in the page, and scores two arms against it. The measured column is what the same arms scored offline — if this page disagrees with it, believe the offline number.

How this works, and how well — sources, licence, measured numbers

What this is, and what it is not

The taxonomy is the 860 distinct product_class values in Wayfair's WANDS dataset. Every label is embedded once, in the page; your text is embedded the same way and the nearest label by cosine wins. That is the entire classifier. It never emits a category that is not in the list, because it only ever picks from the list.

It is a shortlister, not an oracle. It puts the gold category first for roughly 43% of held-out queries and in the top three for 59%. Quantising the encoder to int8 to make it a 33 MB download costs about two points of acc@1 against fp32 and almost nothing at acc@3 — the table below carries both. Queries whose wording has nothing in common with any label — gurney slade 56, a product name — have no reason to land anywhere near the right answer, and do not.

Measured

arm	acc@1	acc@3	tokens sent
this page — snap the query directly, int8	0.434	0.588	0
the same encoder at fp32, off the page	0.455	0.594	0
a cheap model writes the label first, then snap	0.571	—	6
send all 860 labels, ask for a constrained choice	0.701	0.744	5,265

468 queries, one gold label each. The second row is what a server round-trip buys, and the page can score it: the written labels are baked in, produced offline by gemini-3.5-flash-lite. The third row is the honest ceiling — when the vocabulary fits in a prompt, sending it wins by a wide margin. This approach is for the case where it does not: a provider enum cap, a five-thousand-label vocabulary, or per-call cost at volume.

Running a small model locally to write the label instead does not work. Pleias Monad (57M) and Baguettotron (321M) score 0.425 and 0.400 — below the encoder's own baseline — because they echo the query or drift into invented wording. As likelihood rerankers over this page's top ten they score 0.325 and 0.350, while the right answer sits in that top ten 82.5% of the time. What a remote model contributes here is a prior over how taxonomies name things, and that is the first thing a shrinking model loses.

One trap worth knowing about

This page pins the encoder to the WASM backend. Running the same int8 weights through transformers.js on device: "webgpu" returns a collapsed embedding space — every pair of labels about 0.995 apart, so the ranking is noise — and it does so silently, with no error and a plausible-looking result. The first version of this page shipped that way and answered Pillow with Wedding, Drains, Fabric, Flags, Candles. The weights are not at fault: the same model_quantized.onnx matches fp32 PyTorch to three decimals under onnxruntime. Hence the smoke test on load — the page holds its own gold labels, so it checks that the index it just built can rank before it will show you anything.

Credits and licence

Queries, labels and relevance judgements from WANDS, © Wayfair, MIT. Encoder Xenova/gte-small (int8 ONNX, 33 MB, MIT); runtime transformers.js. Method, arms and artifacts: oaustegard/experiments. Pattern after Doug Turnbull, “Don't classify. Hallucinate!”