OCR latency: Azure AI Vision `/imageanalysis:analyze` vs Azure AI Document Intelligence Read

Thomas SANCHEZ 20 Reputation points
2026-06-18T08:23:39.0533333+00:00

Hello,

We perform OCR on document images.

Until now, we were using the Azure AI Vision Image Analysis endpoint, specifically /imageanalysis:analyze, for OCR. The processing time was suitable for our use case: around 300–500 ms per image.

Since this endpoint is planned for retirement in September 2028, we started integrating Azure AI Document Intelligence / Read, using the prebuilt-read model with API version 2024-11-30.

However, on the same documents, the response times are significantly different. With Document Intelligence Read, we are now seeing around 1,5 to 3 seconds per document.

I understand that Azure AI Vision Image Analysis and Azure AI Document Intelligence are not exactly the same services, and that Document Intelligence may perform more advanced processing. However, we would like to know whether there are any parameters, request options, best practices, or recommendations to reduce latency and get closer to the performance we had with /imageanalysis:analyze.

Our goal is to get the best possible OCR accuracy with the lowest possible latency. If available, we would like to know whether there are any settings or options to tune the trade-off between recognition quality and processing time.

For example:

  • disabling or limiting some analysis features;
  • optimizing the image format, resolution, or compression;
  • avoiding unnecessary processing;
  • using a more suitable request mode;
  • choosing another Azure OCR API better suited for low-latency document image OCR;
  • using specific parameters with prebuilt-read to improve response time.

Typical latency observed:

  • Azure AI Vision /imageanalysis:analyze: around 300–400 ms
  • Azure AI Document Intelligence prebuilt-read: around 1–2 seconds

Is there a recommended way to achieve lower latency with Document Intelligence Read, or is this difference expected because of the service architecture?

Thanks in advance for your advice.

Azure Document Intelligence in Foundry Tools

Answer accepted by question author

Karnam Venkata Rajeswari 4,515 Reputation points Microsoft External Staff Moderator
2026-06-18T09:56:43.4+00:00

Hello @Thomas SANCHEZ ,

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

Thank you for sharing the detailed comparison and latency information.

The increase from approximately 300–500 ms to 1.5–3 seconds is expected due to design differences between the two services.

Image Analysis OCR operates as a synchronous API, returning results immediately and optimized for low-latency image scenarios. In contrast, Document Intelligence Read (prebuilt-read) follows an asynchronous workflow:

  1. Submit the document
  2. Process in the service
  3. Retrieve results via polling

This approach, combined with higher-resolution OCR and document-level processing (lines, paragraphs, language detection), naturally introduces additional latency.

Sub-second latency similar to image-based OCR cannot typically be achieved with prebuilt-read, as it prioritizes document understanding over real-time speed.

Currently, there are no configurable parameters to trade OCR accuracy for latency.The pipeline is fully managed by the service.

While the core processing behavior is fixed, latency can still be reduced using the following high-impact improvements:

  1. Limiting processing scope
    1. Use the pages parameter to analyze only required pages
    2. Please avoid sending full multi-page documents when not needed
  2. Optimizing document inputs
    • Using clear images with moderate resolution (150–300 DPI)
    • Please avoid oversized files and excessive blank space
    • Prefer formats such as JPEG or PNG
  3. Reducing request overhead
    1. Send document bytes directly in the request body when possible
    2. When using URLs, ensure storage access latency is minimal
  4. Measuring latency correctly Observed latency often includes:
    • service processing time
    • network delay
    • polling interval delay
    Recommended validation is to
    1. Compare the same input across both services
    2. Measure processing time vs total end-to-end latency
  5. Optimizing polling behavior
    1. Use a balanced polling interval to reduce idle wait time
    2. Please avoid overly aggressive polling, which can increase overhead
  6. Aligning deployment regions Ensure all components are in the same region:
    • application
    • AI resource
    • storage
    This minimizes unnecessary network latency.
  7. Optimizing for throughput
    1. Use parallel processing for multiple documents
    2. Focus on overall throughput instead of single-request latency

Choosing the appropriate OCR service depends on the workload:

  1. Image Analysis OCR Best for:
    • real-time OCR
    • single-image scenarios
    • latency-sensitive applications
  2. Document Intelligence Read Best for:
    • PDFs and scanned documents
    • multi-page inputs
    • structured document extraction

For strict low-latency needs, the image-based OCR API remains the closest fit and is supported until September 2028.

In summary ,

The latency difference is expected due to architectural design and deeper document processing. While latency cannot be reduced to sub-second levels, it can be improved through:

  • limiting processed pages
  • optimizing document size
  • reducing request overhead
  • refining polling strategy
  • aligning deployment regions

For real-time OCR requirements, image-based OCR remains the better choice, while Document Intelligence Read is best suited for document-heavy scenarios requiring higher accuracy and structure.

The following references might be helpful , please check them out

Please let us know if the response was helpful

 

Thank you

 

Please "Accept" the answer with an "Upvote" if the response was helpful. This will be benefitting other community members who face the same issue.

 

 

Was this answer helpful?

2 people found this answer helpful.
0 comments

0 additional answers

Sort by: Most helpful
Most helpful Newest Oldest

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.