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 Foundry portal is only available in North Central US region:
    • East US
    • West US2
    • West Europe
    • North Central US

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

This content applies to: checkmark v3.1 (GA) | Latest version: purple-checkmark v4.0 (preview) | Previous versions: blue-checkmark v3.0 blue-checkmark v2.1

This content applies to: checkmark v3.0 (GA) | Latest versions: purple-checkmark v4.0 (preview) purple-checkmark v3.1 | Previous version: blue-checkmark v2.1

This content applies to: checkmark v2.1 | Latest version: blue-checkmark v4.0 (preview)

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.

Use the model compose operation

  • Start by creating a list of all the model IDs you want to compose into a single model.

  • Compose the models into a single model ID using the Studio, REST API, or client libraries.

  • Use the composed model ID to analyze documents.

Billing

Composed models are billed the same as individual custom models. The pricing is based on the number of pages analyzed. Billing is based on the extraction price for the pages routed to an extraction model. For more information, see the Document Intelligence pricing page.

  • There's no change in pricing for analyzing a document by using an individual custom model or a composed custom model.

Composed models features

  • Custom template and custom neural models can be composed together into a single composed model across multiple API versions.

  • The response includes a docType property to indicate which of the composed models was used to analyze the document.

  • For custom template models, the composed model can be created using variations of a custom template or different form types. This operation is useful when incoming forms belong to one of several templates.

  • For custom neural models the best practice is to add all the different variations of a single document type into a single training dataset and train on custom neural model. The model compose operation is best suited for scenarios when you have documents of different types being submitted for analysis.

Compose model limits

  • With the model compose operation, you can assign up to 500 models to a single model ID. If the number of models that I want to compose exceeds the upper limit of a composed model, you can use one of these alternatives:

  • Analyzing a document by using composed models is identical to analyzing a document by using a single model. The Analyze Document result returns a docType property that indicates which of the component models you selected for analyzing the document.

  • The model compose operation is currently available only for custom models trained with labels.

Composed model compatibility

Custom model type Models trained with v2.1 and v2.0 Custom template and neural models v3.1 and v3.0 Custom template and neural models v4.0 preview Custom Generative models v4.0 preview
Models trained with version 2.1 and v2.0 Not Supported Not Supported Not Supported Not Supported
Custom template and neural models v3.0 and v3.1 Not Supported Supported Supported Not Supported
Custom template and neural models v4.0 preview Not Supported Supported Supported Not Supported
Custom generative models v4.0 preview Not Supported Not Supported Not Supported Not Supported
  • To compose a model trained with a prior version of the API (v2.1 or earlier), train a model with the v3.0 API using the same labeled dataset. That addition ensures that the v2.1 model can be composed with other models.

  • With models composed using v2.1 of the API continues to be supported, requiring no updates.

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

Document Intelligence v3.1:2023-07-31 (GA) 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

Document Intelligence v3.0:2022-08-31 (GA) 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

Document Intelligence v2.1 supports the following resources:

Feature Resources
Custom model Document Intelligence labeling tool
REST API
Client library SDK
Document Intelligence Docker container
Composed model Document Intelligence labeling tool
REST API
C# SDK
Java SDK
• JavaScript SDK
Python SDK

Next steps

Learn to create and compose custom models: