The ability to reduce an embedding to fewer dimensions without losing its usefulness, saving memory and computation while maintaining semantic meaning.