Document Intelligence composed custom models

What is a composed model?

With composed models, you can group multiple custom models into a composed model called with a single model ID. For example, your composed model might include custom models trained to analyze your supply, equipment, and furniture purchase orders. Instead of manually trying to select the appropriate model, you can use a composed model to determine the appropriate custom model for each analysis and extraction.

Some scenarios require classifying the document first and then analyzing the document with the model best suited to extract the fields from the model. Such scenarios can include ones where a user uploads a document but the document type isn't explicitly known. Another scenario can be when multiple documents are scanned together into a single file and the file is submitted for processing. Your application then needs to identify the component documents and select the best model for each document.

In previous versions, the model compose operation performed an implicit classification to decide which custom model best represents the submitted document. The 2024-07-31-preview implementation of the model compose operation replaces the implicit classification from the earlier versions with an explicit classification step and adds conditional routing.

Benefits of the new model compose operation

The new model compose operation requires you to train an explicit classifier and provides several benefits.

• Continual incremental improvement. You can consistently improve the quality of the classifier by adding more samples and incrementally improving classification. This fine tuning ensures your documents are always routed to the right model for extraction.

• Complete control over routing. By adding confidence-based routing, you provide a confidence threshold for the document type and the classification response.

• Ignore document specific document types during the operation. Earlier implementations of the model compose operation selected the best analysis model for extraction based on the confidence score even if the highest confidence scores were relatively low. By providing a confidence threshold or explicitly not mapping a known document type from classification to an extraction model, you can ignore specific document types.

• Analyze multiple instances of the same document type. When paired with the splitMode option of the classifier, the model compose operation can detect multiple instances of the same document in a file and split the file to process each document independently. Using splitMode enables the processing of multiple instances of a document in a single request.

• Support for add on features. Add on features like query fields or barcodes can also be specified as a part of the analysis model parameters.

• Assigned custom model maximum expanded to 500. The new implementation of the model compose operation allows you to assign up to 500 trained custom models to a single composed model.

Development options

Next steps

Learn to create and compose custom models: