Comment by wwn_se

Doing a preprocess using some pdf extraction and ocr tool and then feeding that to the big model is usually way more stable.