RAG applied to visually rich documents, allowing models to retrieve and reason over images and multi-page visual content.