Document Intelligence composed custom models

Important

• Document Intelligence public preview releases provide early access to features that are in active development. Features, approaches, and processes may change, prior to General Availability (GA), based on user feedback.
• The public preview version of Document Intelligence client libraries default to REST API version 2024-07-31-preview.
• Public preview version 2024-07-31-preview is currently only available in the following Azure regions. Note that the custom generative (document field extraction) model in AI Studio is only available in North Central US region:
  • East US
  • West US2
  • West Europe
  • North Central US

This content applies to: v4.0 (preview) | Previous versions: v3.1 (GA) v3.0 (GA) v2.1 (GA)

Important

The model compose operation behavior is changing from api-version=2024-07-31-preview. The model compose operation v4.0 and later adds an explicitly trained classifier instead of an implicit classifier for analysis. For the previous composed model version, see Composed custom models v3.1. If you are currently using composed models consider upgrading to the latest implementation.

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.

How to use model compose

• Start by collecting samples of all your needed documents including samples with information that should be extracted or ignored.

• Train a classifier by organizing the documents in folders where the folder names are the document type you intend to use in your composed model definition.

• Finally, train an extraction model for each of the document types you intend to use.

• Once your classification and extraction models are trained, use the Document Intelligence Studio, client libraries, or the REST API to compose the classification and extraction models into a composed model.

Use the splitMode parameter to control the file splitting behavior:

• None. The entire file is treated as a single document.
• perPage. Each page in the file is treated as a separate document.
• auto. The file is automatically split into documents.

Billing and pricing

Composed models are billed the same as individual custom models. The pricing is based on the number of pages analyzed by the downstream analysis model. Billing is based on the extraction price for the pages routed to an extraction model. With the addition of the explicit classification charges are incurred for the classification of all pages in the input file. For more information, see the Document Intelligence pricing page.

Development options

Document Intelligence v4.0:2024-07-31-preview supports the following tools, applications, and libraries:

Feature	Resources
Custom model	• Document Intelligence Studio • REST API • C# SDK • Java SDK • JavaScript SDK • Python SDK
Composed model	• Document Intelligence Studio • REST API • C# SDK • Java SDK • JavaScript SDK • Python SDK

Next steps

Learn to create and compose custom models:

Build a custom model Compose custom models