What do you use for OCR, I wonder

if it's LLM isn't that too expensive (or flaky if cheap) to be scalable?