Invoice processing prebuilt AI model Inconsistent date format

[NickWarner]

I'm using the Invoice processing prebuilt AI model in an accounts payable automation process and we are hitting a frustrating issue where the date of the invoice is inconsistently formatted. When it is unambiguous, eg the day is greater than 12 and/or the month is in text the format is correct. However when it is ambiguous, it inconsistently assumes American format. I've recently had 4 invoices from the same vendor in the same document format with a 1/09/2023 date where two output as 2023-09-01 and two output 2023-01-09. Does anyone know how the date format gets determined or if it can be controlled?

[Plamena]

Hi @NickWarner 

have you tried with the formatDateTime expression? 

Here there is a video on that .

https://learn.microsoft.com/en-us/ai-builder/form-processing-model-in-flow#convert-a-document-proces...

Best

Plamena 

[NickWarner]

Unfortunately formatDateTime doesn't accept the text string values from the invoice

[jk_one]

Hi @NickWarner 
Can you please send us a sample of the invoices that you encounter the issue with ? 
Thanks,
JK

[takolota]

@NickWarner 

If all else fails & all your invoices are less than 4 pages, then you may want to try a different method like this template: https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/Extract-Data-From-PDFs-and-Images-With-G...

It OCRs the PDF text to create a replica text that it can feed to a GPT prompt. When I use it I add in a line in the prompt requesting all dates to be converted to international yyyy-mm-dd format & it automatically returns the dates in that format in the JSON data extract.

[NickWarner]

As an update for this issue, I have setup in my flow a formatDateTime compose step which works sometimes on the OCR date text value. this also sometimes switches between US and International format on ambiguous dates but seems a bit more reliable in picking international. there's some conditional logic so that if this compose fails it uses the models converted datevalue if it is present, else defaults to a fixed value. I wonder if the pdf gets some locale metadata when created from the user machine that the model is reading?

[jk_one]

Thank you for the additional info, I'll forward to the team.

JK