Send each prompt to the model best suited to its topic, refuse whole topics, and answer reworded repeat questions from a semantic cache.