I belive this should be considered a bug fix, as it resolves unexpected behavior and clarifies how to use undocumented capabilities - Images were not being passed to multimodal models when images were provided and raw was set to true as described in this issue.

This change introduces an optional boolean parameter prepend_images_to_raw_prompt, which if set to true will add the image placeholder fields necessary to pass the images through to the multimodal model. This parameter only has effect if raw is also set to true and images are provided. The prepend_images_to_raw_prompt parameter can be omitted, and the default behavior remains entirely unchanged to avoid breaking existing users.

It also updates the api.md documentation to describe the new functionality (including an example request/response), and also provides and example of manually passing an image placeholder in the text of a raw prompt directly, which offers more control over where the image is inserted in the raw prompt.