LangCache stores finished answers rather than model state, so a paraphrased question can skip inference altogether.