Matryoshka Embeddings

A training technique that concentrates the most useful information in a vector's leading dimensions, so the same embedding can be cut down to a shorter vector with limited loss of accuracy.