Custom extraction model can’t improve table detection when Layout model fails

Hi,

I’m working with Azure Document Intelligence (Custom Extraction Model) and running into a limitation around table detection.

• The Layout model sometimes fails to detect tables in my documents (e.g., transaction listings).
• I tried to use Custom Extraction to train on these cases, but the labeling interface doesn’t let me add a table region unless I predefine a Table field. Even then, I can only label structured columns inside that table I can’t train the system to find new tables when the Layout engine doesn’t detect one at all.

Is there any way to train a custom model to actually improve table layout recognition itself (i.e., teaching it to find tables where the default Layout model fails)?

Layout model?

If this is a current limitation, is table detection improvable through custom neural models or future roadmap?

thanks