Why would Document Intelligence start taking a long time to analyze text?

Question

Hi, I'm using the Document Intelligence Read model to extract text from PDF and other documents as part of a workflow. This has been working well for many months on a variety of differenct documents, typically returning results within a second or two of submission.

Over the last few days, we've been experiencing unusually large delays in receiving processed results back. Sometimes 20-30 minutes for a single file (and sometimes longer), which is messing up the workflow.

This is an S0 tier using the v4.0 API (2024-11-30) on Canada Central. Data needs to stay on Canadian servers and there is no other region available with this service at this time.

Any ideas as to why Document Intelligence might be so delayed in processing?

Answer 1

Hi Bryan Bates

Thank you for reaching out on Microsoft Q&A.

The most common causes of long Azure Document Intelligence job latency (20–30 minutes) are regional capacity/backlogs and hitting S0 quotas. Bursty workloads or aggressive polling can make it worse. Large files or cross-region storage can also increase processing time.

What to check:

Service Health: Check Azure Service Health for your region.

Quotas: Default S0 limits: 15 TPS POST, 50 TPS GET; exceeding these causes throttling.

Polling: Avoid GET calls more than once every ~2 seconds.

Files & storage: Very large PDFs or cross-region storage increase end-to-end latency.

References:

Azure Document Intelligence - Troubleshooting Issues

Azure Service Health Overview

If this information is useful, please accept the answer and upvote so it can assist other community members.

Answer 2

Hi Bryan Bates,

Welcome to Microsoft Q&A and Thank you for reaching out.

I understand your experiencing long delays sometimes 20 to 30 minutes or more when using the Document Intelligence Read model, even though it used to return results within a couple of seconds, there are several common reasons this might be happening.

The size and complexity of the documents themselves play a big role. Larger files with more pages naturally take longer to process. Documents that include scanned images, handwriting, tables, or high-resolution pictures require more time and computing power to analyze. Even small changes like higher image quality or additional embedded content can significantly increase processing time.

Document Intelligence is a multitenant, asynchronous microservice. This means it shares resources across many users and processes requests asynchronously. Because of this architecture, latency can vary even for similar documents and there is no guaranteed SLA for latency and processing times can fluctuate depending on backend capacity, operational load, or maintenance activities.

API versions (like 2024-11-30) and the use of custom extraction models have been reported to cause increased processing times or resource exhaustion in some regions. Custom models usually require more processing than prebuilt ones and may be more sensitive to backend resource availability or recent software updates.

Your current service tier (S0) provides good baseline capacity but has limitations compared to higher tiers or dedicated resources. If there is increased demand in the Canada Central region or if you hit quota limits, your requests may be queued or throttled, leading to longer wait times. Additionally, regional outages or partial service degradations can cause delays.

Sometimes network issues or the size of the documents being uploaded and downloaded can add to delays. Poorly formatted or corrupted documents may also require extra time for the system to process, causing longer latency than usual.

To troubleshoot, start by comparing documents you process now versus earlier ones to check for size or complexity differences. Monitor Azure service health for Canada Central to see if there are ongoing incidents. Review your resource usage metrics for throttling or error spikes and ensure you aren’t hitting quota limits. You might also want to try processing documents with prebuilt models or different API versions to compare performance.

Consider breaking large documents into smaller batches or scheduling processing during off-peak hours to reduce load. Implement retry and timeout logic in your workflow to gracefully handle occasional long-running requests without stalling the entire process.

For your References:

Troubleshooting Issues

Azure Service Health Overview

I Hope this helps. Do let me know if you have any further queries.

Thank you!