Azure AI Document Intelligence
An Azure service that turns documents into usable data. Previously known as Azure Form Recognizer.
1,443 questions
A huge latency added to Document OCR request
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(38.4, 40%, 13%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAH%3C/text%3E%3C/svg%3E)
Ahmad Hammoudeh
0
Reputation points
I'm experiencing a latency issue with the Document Intelligence service while applying OCR on a large number of images of books. Each image takes about 5 seconds to respond, which I believe is not solely due to the OCR step. I have checked the Azure resource request latency metric, and the latency for my request is around 100ms. After spending time researching the issue, I have ruled out the latency of sending the image itself cause it's light (image info is mentioned below), so I am considering factors such as the asynchronous nature of the request, how the Azure server handles the request, and the tier of the service.
Regarding the issue, I have several questions:
- What could be the real reason behind this latency, is there something that I have missed?
- Regarding my configurations and image info mentioned below, is this latency normal?
- I am contemplating using Azure functions with Azure storage to process a large number of documents I have. Will I expect the same response time for each image?
Configurations:
- Free Tier F0, Location East US, Version 4.0
- SDK version = 1.0.0b2, python
- Model type: Read
- Environment: Google Colab, the server located in the US
- I followed the quick start tutorial to send the request
Image info:
- Link to the image
- File Type: TIFF
- File Size: 173 KB
- Dimensions: 2339 x 1654
Thanks for your time.
{count} votes
-
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(6.4, 61%, 10%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ES%3C/text%3E%3C/svg%3E)
santoshkc 5,080 Reputation points Microsoft Vendor2024-05-13T07:52:09.79+00:00 Hi @Ahmad Hammoudeh,
Thank you for reaching out to Microsoft Q&A forum!
The time to analyze a document depends on its size and content. Document Intelligence is a multitenant service, so latency for similar documents is comparable but not always identical. Occasional variability in latency and performance is inherent in any microservice-based, stateless, asynchronous service that processes images and large documents at scale.
Although we're continuously scaling up the hardware and capacity and scaling capabilities, you might still have latency issues at runtime.
To reduce the latency of document recognition using Azure Document Intelligence, you can try the following:
- Use a smaller model: Smaller models have fewer parameters, which means they can be loaded faster and run faster.
- Use a pre-built model: Try using a pre-built model. Pre-built models are optimized for performance and can be faster than custom models.
- Use a GPU: Try using GPUs, which can accelerate the processing of models and reduce the latency of document recognition.
- Use batch processing: If you are processing multiple documents, try using batch processing. Batch processing can reduce the overall processing time by processing multiple documents at once.
I hope you understand! Thank you.
-
Ahmad Hammoudeh 0 Reputation points
2024-05-14T07:40:48.3866667+00:00 Thanks for your response @santoshkc .
Regarding the points you mentioned to reduce latency:
- For points 1 and 2: I mentioned that I use the pre-built model 'read'
- For point 3: As I'm using the pre-built model by requesting the API, I think the processing is already done on the GPU.
- For point 4: I didn't find an option to send a batch of images to the API unless I group them in a PDF
So just to confirm what I understood, as Document Intelligence is a multitenant service, a 5s latency is normal for a request of a light image, even if I use Azure functions with Azure storage?
-
santoshkc 5,080 Reputation points Microsoft Vendor
2024-05-14T09:11:27.37+00:00 Hi @Ahmad Hammoudeh,
Thank you for your follow-up query.
It's difficult to say whether a 5s latency is normal for a request of a light image, even if you use Azure functions with Azure storage. However, it's important to note that occasional variability in latency and performance. You can try using Azure functions with Azure storage to process a large number of documents you have.
Thank you.
-
santoshkc 5,080 Reputation points Microsoft Vendor
2024-05-15T11:38:19.9633333+00:00 Hi @Ahmad Hammoudeh,
We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others. Thank you.
-
Ahmad Hammoudeh 0 Reputation points
2024-05-21T08:52:41.7766667+00:00 Hi @santoshkc
I tried the pay-as-you-go tier and got the same response time. So I went with the following solutions:
- Parallelize sending requests (15 requests per second)
- group images into a PDF instead of sending each image alone, this is similar to a batch request
I tried the Azure functions with Azure storage. As I expected, I didn't achieve less response time.
Currently, I'm working on testing the container solution, I will report the response time when I finish.
Thanks for your support
Sign in to comment