Wild · feature-extraction · WASM
The famous word2vec trick — king − man + woman ≈ queen — was vector arithmetic on
embeddings. GTE is a sentence embedder, not word2vec, so it is not built for this. But the
geometry still bends toward the answer. Pick three terms, and we compute
vec(A) − vec(B) + vec(C), then rank a vocabulary by cosine to that target. You see the
real nearest neighbours — hit or miss.
Reading it: the top row is the vector-space "answer". Because this is a sentence model, expect it to sometimes just return C or a synonym of A — that's the honest behaviour, and part of the point.