But indexing *is* training. It's just not using end-to-end gradient descent.