Paper under https://arxiv.org/abs/2606.23050

(As a side note, I do OCR locally as a small RAG for citations I read in books and also chunk input, but merely to save RAM - interesting this natural approach also work in a streaming model)