We are using the new machine learning invoice processing stuff for template-less invoice processing right now. The product is great and works awesome for most cases, but I have found some examples where the dollar amount text pulls in correctly, but the actual value is way off. In this case, a credit memo shows -13.94 for taxes, but the value shows -1394.00. The total shows -162.19 and the value shows -16219.00. Obviously these discrepancies could be a major problem if the person validating was not paying attention.
I am wondering if this is some type of bug in the machine learning extractor that does not deal well with negative values. I have only seen this on credit memos so far. Could someone look into this?