Azure AI Document Intelligence – custom model training data size limit and custom neural model availability in South Africa North

Question

Azure AI Document Intelligence – custom model training data size limit and custom neural model availability in South Africa North

Ehsan Datasmith 0

Hi,

We're building a production OCR solution on Azure AI Document Intelligence using a custom extraction model, and we're blocked by the training data size limit.

Our setup:

Pricing tier: Standard (S0)
Region: South Africa
API version: 2024-11-30 / v4.0
Model type: Template
Training through: Document Intelligence Studio
Dataset: 300 plus labeled documents, ~250 MB total

The issue:

Training fails with a size limit error at around 150 MB. The exact error message is:

"InvalidTrainingContentLength: Training document collection is limited to 157286400 total bytes."

The service limits documentation lists 50 MB for template models, so we'd like to understand where the 150 MB limit comes from and which limit applies to our configuration.

Our questions:

Can the training dataset size limit be increased for our subscription (through a support request or quota increase)? If so, what's the process?
Is training a custom neural model (documented at ~1 GB of training data) supported for a Document Intelligence resource in South Africa North? Our subscription and data residency requirements are South Africa based. If neural training isn't available in this region, what is the recommended approach, for example training in another region and copying the model to South Africa North?

Why we need this:

We’ve developed an AI-based automated invoicing system that is currently running in production, processing invoices from many vendors and suppliers across multiple branches, each with its own distinct set of suppliers. Each supplier may use one or more different invoice formats, and as we’ve added more suppliers and layouts to our training data, we’ve hit this size limit, which now blocks us from extending the model to cover new formats.

Thanks in advance for any guidance.

0 comments

1 answer

Your answer

Answer 1

Hi @Ehsan Datasmith

Thanks for laying it out so clearly. Short version on the two questions:

Size limit: the service limits page marks training dataset size as not adjustable for both template and neural models, so a support request won't raise it. The docs say 50 MB for template and the service is enforcing ~150 MB for you. I can't find anything official explaining that gap, but either way it's a hard ceiling per model.
Neural in South Africa North: it isn't in the list of regions that support custom neural training. Microsoft's documented route is to train in a supported region and copy the model to any other region. That means your training data would sit in that region during training, so check it against your residency requirements before going that way.

For your scenario (many suppliers, many layouts), you can stay in South Africa North and get around the limit by splitting the training set and composing the models:

Group your labeled documents by supplier or layout family, keeping each group well under the size limit.
Train a separate template model for each group in Document Intelligence Studio.
Train a custom classification model, with one folder per group (folder name = document type). In v4.0 a composed model needs this explicit classifier to route each document. Check that custom classification is available on your resource in Studio before you start.
Compose the classifier and the extraction models. v4.0 allows up to 500 extraction models in one composed model, and you call it like a single model. Note that classification is billed on every page, on top of extraction.
When a new supplier comes in, train a small model just for them, add a folder to the classifier, and recompose, instead of retraining everything.

This also scales better over time. One giant template model gets harder to keep accurate as layouts pile up.

Roughly how many distinct invoice layouts are you dealing with today?

If this resolved your question, please consider accepting it as the answer. If composing runs into routing or accuracy issues, let me know and I'll be happy to keep helping.

Reference: