Text this: Rethinking visual prompting for multimodal large language models with external knowledge