Basics · feature-extraction · WASM
The fundamental capability, felt directly: GTE turns each sentence into a 384-d unit vector, and cosine similarity — the dot product of those vectors — is one number from −1 to 1 that says how close in meaning they are. Edit either line and watch it move. Then compare how three different embedders score the same pair.
cosine = –
GTE (mean pool, no prefix), MiniLM (mean pool) and BGE (CLS pool, prefix-tuned) each learned their own geometry. The absolute number differs per model — what matters is whether each one separates "same meaning" from "different". Loads MiniLM + BGE on demand.