We just had that same bug when we used a conventional ML engine to OCR a document. Thats very common on quantities. The new LLMs allow for prompt engineering, so we were able to solve it.