Imagine trying to describe a complex image to someone who can't see it, and asking them questions about specific details. This is the challenge of Knowledge-Based Visual Question Answering (KB-VQA), where AI models must interpret images and draw on external knowledge to provide accurate answers. Traditional methods often fall short when dealing with intricate, multi-hop questions that require more than just identifying objects in an image. Researchers are now exploring innovative techniques to overcome these limitations. Instead of presenting the AI with a single, complex question, they decompose it into a series of simpler queries. This allows the AI to focus on individual visual aspects, leading to a richer understanding of the image. Moreover, researchers are using type-checking to determine if the question is primarily knowledge-based or visual. This allows them to tailor their approach, using specialized models for extracting visual information from images while using large language models (LLMs) as a general knowledge source for non-visual questions. This method significantly boosts performance on KB-VQA tasks, improving accuracy on established datasets like OK-VQA, A-OKVQA, and KRVQA. Decomposing questions enables the AI to extract more detailed visual information and integrate it with external knowledge more effectively. For example, if asked "What country is the company that made the device in her hand from?", the AI could first identify the device as a Nintendo Wii and then leverage external knowledge to pinpoint the company's origin, Japan. Adding Optical Character Recognition (OCR) further enhances the AI's understanding by extracting textual information within the image. Although these advances are promising, challenges remain. Developing robust captioning models that can perceive text within images and creating more effective chat-based captioners are crucial for further progress. As research continues, we can expect even more sophisticated AI systems capable of unraveling the mysteries hidden within complex images and answering questions that require deep reasoning and knowledge integration.