Retraining Out of the box Invoice model in AI Fabric

Hi Team,
I need to retrain out of the box Invoice Model available in AI fabric for our internal Invoices, like adding some new fields or getting right data for the already available fields in the model. Do we have any document or training for retraining the models?


Hello hopefully I can help answer a few of your queries.

Short of creating your own Python solution and uploading to Ai Fabric, I do not believe there is any way to add fields to the out of the box solutions. The code isn’t there for fields not listed so it would have to be created by a developer.

As far as training existing fields, I would suggest creating a Document Understanding workflow and use both the classifier trainer and extractor trainer and run through many documents in either an attended or semi attended way, where a user corrects the fields if they are extracted incorrectly. You should see documents begin to classify themselves with higher confidence. Also after uploading the files from the extractor trainer to data manager, and then to Ai Fabric, the extractor should improve themselves as well. If you keep both of these in your future workflows then you can maintain a system that automatically changes and finds better solutions over time.

@JosephNehl Your answer seems to be wrong.

@prince.adlakha According to this

If you are adding new fields to the model, you need 30-50 docs per new field to get good results.

or this

Using state-of-the-art transfer learning capabilities, this model can be retrained on additional labelled documents and tailored to specific use-case […]. The dataset used may have the same fields, a subset of the fields, or have additional fields.

you can add new fields to the model.


So first you’d have to use the data manager to label your data (and create the new field), then you have to retrain the model in AIFabric.

Here’s another short article on that topic:

And here’s an extremely detailled video by UiPath on how to do all of this step by step:

1 Like

@T0Bi This great! Thank you for providing this information. My understanding was out of date as I had asked a similar question back in June and was not able to find any information about a way to make this work. Thank you for correcting me and giving the new up-to-date information. I apologize for providing the old solution. Great catch!

1 Like