Seol mar théacs é seo: Rethinking visual prompting for multimodal large language models with external knowledge