Yup, tried all variations of that. There's the advantage that you can use a lower hallucination OCR specific model for the pre-processing, at least for clean text. But for something hard like handwritten forms, applying VLM with context is less error prone than preprocessing to text.
Also - and this is bizarre - the token cost of doing that is higher, not lower, at least in Gemini world, and by a large margin. That's very counterintuitive, but a page encoded as image tokens can be smaller than same page as text, and is not meaningfully lossy on documents that are just typed text because the models are well trained on those.
Also - and this is bizarre - the token cost of doing that is higher, not lower, at least in Gemini world, and by a large margin. That's very counterintuitive, but a page encoded as image tokens can be smaller than same page as text, and is not meaningfully lossy on documents that are just typed text because the models are well trained on those.