Practical · feature-extraction · WASM
Once text is a vector, whole product features become a few lines of arithmetic. Here are three, running entirely on your device with GTE: semantic search, clustering, and classification by nearest label — the last one needs no training data at all.
Everything below runs on these lines. Edit them, or keep the mixed support/topic sample.
Rank every line by meaning against a query — no keyword overlap required.
Group the lines into k unlabeled clusters by embedding proximity. The support tickets, refund requests, bug reports and praise should fall out on their own.
Embed a handful of label names once, then tag every line by whichever label vector it sits closest to. Change the labels and everything re-tags instantly — no retraining.