- Feature: Introduce
useOpenAICompatibleVisionProvider for proper support of OpenAI API compatible providers
- Feature: Add
coordinateSpace ("pixels" default | "normalized") and coordinateScale (default 1000) options to request the coordinate convention a model family is trained on; pixels mode states the image dimensions in the prompt, and the confidence-1.0 anchor is removed so models report real confidence
- Feature: vision-evals: Automated test suite for vision models, with a standard set of test images and expected outputs, to benchmark models and detect regressions
- Improve confidence output