indeed, and in practical terms, this is more often than never, and particularly with large knowledge bases. also makes super sense for VLMs and ViT models.