◇ Could this help me?
Yes on the pattern, with two conditions. Routing images, video, and PDFs to a model that handles them well, in the background, and returning only the relevant summary is a genuinely good architecture, because it fixes a real weakness without bloating the main context window. The conditions: verify the third-party plugin and its publisher before installing, since a plugin holding an API key and reading local files has meaningful access, and be deliberate about which files get sent to a second provider. This is a general practice.
I want a coding assistant to be able to work with images, videos, and PDFs by delegating them to a multimodal model rather than trying to read them itself. Explain the options for wiring this up, including doing it directly with an API call I control rather than through a third-party plugin. For each option, tell me what data leaves my machine, where it goes, what credentials are involved, and what the failure modes are. Do not install anything yet. Give me the comparison and I will choose.