What criteria would you use to evaluate the effectiveness of a recommendation system that takes tasks from a database or generates them via LLM?