We would like to understand whether UiPath provides or plans to provide a continuous learning capability for Document Understanding where the solution can automatically improve from user corrections without requiring manual retraining and redeployment.
Business Requirement
Our customer processes invoices from more than 70 vendors. While the current document classification accuracy is above 80%, there are scenarios where invoices from a particular vendor fail extraction or classification and require manual validation.
The business expectation is:
-
When a user corrects a failed invoice in Validation Station/Action Center, the platform should remember that correction.
-
The next time an invoice with the same vendor layout is received, it should be processed successfully without requiring another human validation.
-
Ideally, this learning should happen automatically after a single correction or with minimal user intervention.
Current Understanding
Based on our understanding:
-
Public Document Understanding endpoints are inference-only and cannot be retrained.
-
Retraining currently requires collecting corrected documents, creating/updating a dataset, retraining the model, publishing a new model version, and redeploying it.
-
This introduces a delay before improvements are reflected in production.
Questions
-
Does UiPath provide any capability for automatic or incremental learning from Validation Station/Action Center corrections without requiring manual retraining and redeployment?
-
Is there any supported mechanism where a corrected invoice can immediately improve extraction or classification for subsequent invoices with the same layout?
-
For fixed-layout vendors, is there a recommended approach that avoids repeated manual validation while minimizing model maintenance?
-
Are there any roadmap items or upcoming features related to:
-
Continuous learning
-
Online/incremental model updates
-
Automatic template creation from validated documents
-
Automatic incorporation of human corrections into production models
-
-
What is UiPath’s recommended architecture for customers processing invoices from a large number of recurring vendors (70+ vendors) where the objective is to continuously reduce manual validation over time?
Expected Business Outcome
The desired behavior is that once a user validates or corrects an invoice from a recurring vendor, future invoices with the same structure should require little or no manual intervention, while maintaining high extraction and classification accuracy.
We would appreciate UiPath’s guidance on the best-supported approach to achieve this requirement using the current platform, or whether any upcoming capabilities are planned to address this scenario.