[BUG] [DocumentIntelligence] - invoice rows suddenly returned from model incorrectly

[maxneovici]

Library name and version

Azure.AI.DocumentIntelligence 1.0.0-beta

Describe the bug

In the pre-built invoice model we've suddenly seen examples where the model is returning only a subset of rows concatenated.

I've provided an example from the studio where it's re-producible across API-versions and service resources (but anonymized the data):

The underlying table has the characters correctly identified and this invoice template from the same supplier used to work without any issues.

To clarify: there are 17 invoice lines, but we only get two of them from the model with extremely long concatenated descriptions etc from several/the rest of rows. However: in the table element, the expected rows are all present.

We've tried to upgrade the SDK from FormRecognizer 4.1 to Azure.AI.DocumentIntelligence 1.0.0-beta and to go to the latest API version. And we've tried to create new resources in a new region, but the issue persists.

We also, in addition, have some internal tools where we measure success rate over a large subset of invoices.

And after July17th we observed this behavior starting, leading to dramatic decreases in success rate of invoice row totals vs. invoice headers (headers that we get from a partner).

Expected behavior

Invoice rows to be returned correctly in Items (as before)

Actual behavior

Does not return invoice lines correctly in Items

Reproduction Steps

I'm not sure it's easily re-producible but we have tons of examples if you wish to get in contact with us.

Environment

.NET6.0

[jsquire]

@kinelski: Would you please offer thoughts on whether this behavior is due to client logic or the service?

[jsquire]

Thank you for your feedback. Tagging and routing to the team member best able to assist.

[maxneovici]

Thanks for picking it up. This is still ongoing, any updates? If we can help out by providing examples or something else, please let us know. :)