Good point! Some models are fast enough to generate the whole response, but defining another kind for streaming responses is great.