Service quotas and limits

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

This article contains both a quick reference and detailed description of Azure AI Document Intelligence service Quotas and Limits for all pricing tiers. It also contains some best practices to avoid request throttling.

Model usage

Document types supported	Read	Layout	Prebuilt models	Custom models	Add-on capabilities
PDF	✔️	✔️	✔️	✔️	✔️
Images: JPEG/JPG, PNG, BMP, TIFF, HEIF	✔️	✔️	✔️	✔️	✔️
Microsoft Office: DOCX, PPTX, XLS	✔️	✔️	✖️	✖️	✖️

✔️ = supported ✖️ = Not supported

For Document Intelligence v4.0 2024-11-30 (GA) supports page and line features with the following restrictions:

• Angle, width/height, and unit aren't supported.
• For each object detected, bounding polygon or bounding regions aren't supported.
• Page range (pages) isn't supported as a parameter.
• The lines object isn't supported.

• Document Intelligence SDKs
• Document Intelligence REST API
• Document Intelligence Studio v3.0

Quota	Free (F0)1	Standard (S0)
Analyze transactions Per Second limit	1	15 (default value)
Adjustable	No	Yes 2
Get operations Per Second limit	1	50 (default value)
Adjustable	No	Yes 2
Model management operations Per Second limit	1	5 (default value)
Adjustable	No	Yes 2
List operations Per Second limit	1	10 (default value)
Adjustable	No	Yes 2
Max document size	4 MB	500 MB
Adjustable	No	No
Max number of pages (Analysis)	2	2000
Adjustable	No	No
Max size of labels file	10 MB	10 MB
Adjustable	No	No
Max size of OCR json response	500 MB	500 MB
Adjustable	No	No
Max number of Template models	500	5000
Adjustable	No	No
Max number of Neural models	100	500
Adjustable	No	No

Custom model usage

• Custom template model
• Custom neural model
• Composed classification models
• Composed custom models

Quota	Free (F0) 1	Standard (S0)
Compose Model limit	5	500 (default value)
Adjustable	No	No
Training dataset size * Neural and Generative	1 GB 3	1 GB (default value)
Adjustable	No	No
Training dataset size * Template	50 MB 4	50 MB (default value)
Adjustable	No	No
Max number of pages (Training) * Template	500	500 (default value)
Adjustable	No	No
Max number of pages (Training) * Neural and Generative	50,000	50,000 (default value)
Adjustable	No	No
Custom neural model train	10 hours per month 5	no limit (pay by the hour), start with 10 free hours each month
Adjustable	No	Yes 3
Max number of pages (Training) * Classifier	10,000	10,000 (default value)
Adjustable	No	No
Max number of document types (classes) * Classifier	500	500 (default value)
Adjustable	No	No
Training dataset size * Classifier	1GB	2GB (default value)
Adjustable	No	No
Min number of samples per class * Classifier	5	5 (default value)
Adjustable	No	No

¹ For Free (F0) pricing tier see also monthly allowances at the pricing page.
² See best practices, and adjustment instructions.
³ Neural models training count is reset every calendar month. Open a support request to increase the monthly training limit. Starting with the v4.0 API, training requests over 20 requests in a calendar month are billed on the training tier. See pricing for details.

⁴ This limit applies to all documents found in your training dataset folder prior to any labeling-related updates.

⁵ This limit applies for v 4.0 (2024-11-30 GA) custom neural models only. Starting from v 4.0, we support training larger documents for longer durations (up to 10 hours for free, and incurring charges after). For more information, please refer to custom neural model page.

Detailed description, Quota adjustment, and best practices

The default limits can be extended by requesting an increase via a support ticket. Before requesting a quota increase (where applicable), ensure that it's necessary. Document Intelligence service uses autoscaling to bring the required computational resources on-demand, keep the customer costs low, and deprovision unused resources by not maintaining an excessive amount of hardware capacity.

If your application returns Response Code 429 (Too many requests) you are over the threshold for one or more of the transactions per second limits (TPS):

• Analyze transactions Per Second limit The TPS for submitting analyze requests (POST)
• Get operations Per Second limit The TPS for polling for results on analyze operations (GET)
• Model management operations Per Second limit Operations related to model management like build/train and copy.
• List operations Per Second limit Operations related to listing models, operations.

General best practices to mitigate throttling during autoscaling

To minimize issues related to throttling (Response Code 429), we recommend using the following techniques:

• Implement retry logic in your application
• Avoid sharp changes in the workload. Increase the workload gradually
  Example. Your application is using Document Intelligence and your current workload is 10 TPS (transactions per second). The next second you increase the load to 40 TPS. The result is a 429 response code for some requests as you are over the 15 TPS limit for submitting analyze operations. You could either back off the processing to stay under the 15 TPS or request an increase on the TPS to support your higher volumes.

The next sections describe specific cases of adjusting quotas. Jump to Document Intelligence: increasing concurrent request limit

Increasing transactions per second request limit

By default the number of transactions per second is limited to 15 transactions per second for a Document Intelligence resource. For the Standard pricing tier, this amount can be increased. Before submitting the request, ensure you're familiar with the material in this section and aware of these best practices.

The fist step would be to enable auto scaling. Follow this document to enable auto scaling on your resource * enable auto scaling. With auto scaling enabled your resource can continue to accept requests over the TPS limits configured if there's capacity on the service. It can still result in request throttled.

Increasing the Concurrent Request limit does not directly affect your costs. Document Intelligence service uses "Pay only for what you use" model. The limit defines how high the Service can scale before it starts throttle your requests.

The existing value of different request limit categories is available via Azure portal, under the monitoring tab on the resource overview blade.

Create and submit support request for TPS increase

Initiate the increase of transactions per second(TPS) limit for your resource by submitting the Support Request:

• Sign in to the Azure portal
• Select the Document Intelligence Resource for which you would like to increase the TPS limit
• Select -New support request- (-Support + troubleshooting- group). A new window appears with autopopulated information about your Azure Subscription and Azure Resource
• Enter -Summary- (like "Increase Document Intelligence TPS limit")
• Select "Quota or usage validation" for problem type field.
• Select -Next: Solutions-
• Proceed further with the request creation
• Enter the following information in the -Description- field, under the Details tab:
  • a note, that the request is about Document Intelligence quota.
  • Provide a TPS expectation you would like to scale to meet. While TPS increases are free, you should only request a TPS that is reasonable for your workload.
  • Azure resource information
  • Complete entering the required information and select -Create- button in -Review + create- tab
  • Note the support request number in Azure portal notifications. Look for Support to contact you shortly for further processing.

Example of a workload pattern best practice

This example presents the approach we recommend following to mitigate possible request throttling due to Autoscaling being in progress. It isn't an exact recipe, but merely a template we invite to follow and adjust as necessary.

Let us suppose that a Document Intelligence resource has the default limit set. Start the workload to submit your analyze requests. If you find that you're seeing frequent throttling with response code 429 when checking for completion, start by implementing an exponential backoff on the GET analyze response request. By using a progressively longer wait time between retries for consecutive error responses, for example a 2-5-13-34 pattern of delays between requests. In general, we recommended not calling the get analyze response more than once every 2 seconds for a corresponding POST request. The analyze response also contains a retry-after header that indicates how long you should wait in seconds before checking for completion of that request.

If you find that you're being throttled on the number of POST requests for documents being submitted, consider adding a delay between the requests. If your workload requires a higher degree of concurrent processing, you then need to create a support request to increase your service limits on transactions per second.

Generally, we recommended testing the workload and the workload patterns before going to production.

Next steps

Learn about error codes and troubleshooting