I’m building an inventory review flow where a request may include several photos of different products. I’d like the model to return a label and a short note for each photo, without mixing up which observation belongs to which image.

For people using the OpenAI API this way, what output format has worked best? Do you include an image ID alongside each image in the prompt and ask for matching IDs in the response, or is there a more reliable pattern for mapping answers to image IDs? I’m especially interested in what holds up when there are a lot of images in one request.