Is it possible to train and create a custom model in Document Intelligence without using Document Intelligence studio

Hi Team,

Is it possible to train and create a custom model in Document Intelligence without using Document Intelligence studio. Please let us know if there are any REST API's which can be utilized for creation of custom models, so that if we specify the co-ordintaes and field name of a document and call a REST API, a model should be created in Document Intelligence.

In short our query is if we provide the PDF and the label.json file do we have any REST API in document intelligence which would create the model automatically?

Thanks

santoshkc 8,690 Reputation points Microsoft Vendor

2024-09-25T13:50:26.4333333+00:00

Hi @Sriramsubramaniyan Nadarajan,

Thank you for reaching out to Microsoft Q&A forum!

To train and create a custom model in Azure Document Intelligence without using Document Intelligence Studio, you can refer to: Build Document Model.

Thank you.
Sriramsubramaniyan Nadarajan 76 Reputation points

2024-09-25T15:08:11.5666667+00:00

Hi @santoshkc ,

Thanks for your response, however we need the model with custom labels, while checking the build document model api, we can see specification to provide training files blob location, but there is no parameter to define the custom fields, can you please let me know how it works
Sriramsubramaniyan Nadarajan 76 Reputation points

2024-09-25T15:16:51.6233333+00:00

Also , please provide if we have any option where an end user who does not have access to Azure document intelligence can create models, we are exploring options where users can create labels.json in sample labelling tool, and we can create the file.ocr.json using document analyze API, using these a custom model needs to be created
santoshkc 8,690 Reputation points Microsoft Vendor

2024-09-26T12:45:53.51+00:00

Hi @Sriramsubramaniyan Nadarajan,

Thank you for you follow-up queries.

The Build Document Model API allows you to specify the blob location for your training files, but it does not include a direct parameter for defining custom fields. Instead, you need to ensure that your training documents are labeled appropriately in your labels.json file. The model will use this labeled data to learn the custom fields during the training process. Make sure your documents and labels are structured according to the API requirements for optimal results.

Currently, there isn't a direct option for end users without access to Azure Document Intelligence to create models. However, you can facilitate this by allowing users to create a labels.json file using a sample labeling tool. Once they have this file, you can utilize the Document Analyze API to create the file.ocr.json. You can then use this output along with the labels.json to build the custom model programmatically, ensuring users can still contribute to the model creation process without direct Azure access.

I hope you understand. Thank you.
Sriramsubramaniyan Nadarajan 76 Reputation points

2024-09-26T13:10:02.9533333+00:00

Hi @santoshkc , Thanks for the detailed explanation, we were also exploring the same options as you said, however in some faq's we can see that it was mentioned that using sample labelling tool and creating labels.json ourselves can cause some inconsistency, and its better to use document intelligence studio. However we are not in a position to grant access to all end users to azure studio portal, so sample labelling tool would be the best option., Can you please let me know if there are any known issues that could happen, if we take the same approach which you mentioned?
santoshkc 8,690 Reputation points Microsoft Vendor

2024-09-27T15:06:54.44+00:00

Hi @Sriramsubramaniyan Nadarajan,

Using a sample labeling tool to create the labels.json file can work well, but ensure users follow standardized guidelines to maintain consistency and accuracy. Validate the labels before training the model, and confirm that the labels.json structure aligns with API requirements. Conducting small-scale tests can help identify potential issues, and providing clear documentation and support will enhance the process.

Thank you.
Sriramsubramaniyan Nadarajan 76 Reputation points

2024-09-30T05:17:54.0933333+00:00
Hi @santoshkc , thanks for the detailed response. We will try the same and check for any issues, one last question, I am using labelling tool and I had created the labels.json file, and also used document analyze api and created the file.ocr.json, what would be the next step for me to create the model,

Should i upload both labels.json, file.ocr.json, and the file to blob and use same container url and call the custom model build api. for ex:

POST {endpoint}/documentintelligence/documentModels:build?api-version=2024-07-31-preview

Please advise on custom model creation process alone. Thanks for your help

{ "modelId": "m1", "description": "Custom model description", "buildMode": "template", "azureBlobSource": { "containerUrl": "" }, "tags": { "createdBy": "user" } }
santoshkc 8,690 Reputation points Microsoft Vendor

2024-09-30T11:20:31.8533333+00:00

Hi @Sriramsubramaniyan Nadarajan,

Thank you for your follow-up query.

To create the custom model, you need to upload both the labels.json, file.ocr.json, and the document file itself to your Azure Blob Storage. Once uploaded, you can use the container URL and call the Custom Model Build API as you mentioned.

Make sure the containerUrl points to the blob container that contains all the necessary files (labels.json, file.ocr.json, and the document).

Thank you.

Share via

Is it possible to train and create a custom model in Document Intelligence without using Document Intelligence studio

Your answer