Hi there, I have a modern Document Understanding model that does classification and extractions for invoices and payment advices. How can I setup my code in Studio such that when this process is in Production:
Classification confidence < threshold: Business User receives AC form to select correct category for document. This selection and document then gets uploaded to the DU model so that it can learn actively.
Extraction confidence < threshold: Same as above
What code, activities, and packages must I use so that these Business User corrected documents get sent to the model?
Lastly, is my understanding correct that these documents will be sent to a “Retraining Candidates” section in the DU model, and then I can:
Manually review them
Start the re-training for the DU model
Publish & Deploy new DU model
I will use an Asset to represent the DU’s model version and update it accordingly?
I see there is a post from 2024 addressing some of this, but a lot has changed with the IXP implementation.
Do check the latest Document Understanding Process template in the Studio, it has build-in setup for training the classifiers and extractors.
First you need to check where do you want to store your retraining documents, in some shared folder, or directly in dataset?
Then configure the training activities that is by filling the project, dataset and authorization details.
Enable retraining from the config file.
Schedule recurring pipeline or run manually for classic DU projects. For modern projects you just annotate the document (with the help of pre-labelling) and retraining will be done automatically.
But, rathar than relying on user validated documents (from action center), annotate by developer to ensure consistency.
The method is similar as earlier. The major change with IXP is Unstructured and Complex documents processing with Gererative AI capabilities, for retraining this type of model, you need to store documents seperately then use those to refine the prompts.