Users shouldn't need to remind the LLM not to fabricate data when it doesn't have access to a specific dataset.