Isn't it super expensive to run an LLM on each result?