AI Builder - Error while using "Recognize text in an image or a PDF document”

[MartineA]

Hello,

In Power Automate, after many successful executions of the AI Builder action “Recognize text in an image or a PDF document” (more than 45,000 executions), the action is now suddenly failing and throwing an error "CognitiveService: Error while invoking method for action Predict”.

Full error:

{"operationStatus":"Error","error":{"type":"Error","code":"InternalServerError","message":"CognitiveService: Error while invoking method for action Predict"}}

Does anybody know about that error and how it can be solved?

Thank you for your help.

[antoinec]

Hi Martine, Sorry you're getting this error. Are you getting it for a file only or is the model systematically failing for you now? Please feel free to send me a private message with the id of your model so we can look into product telemetry to see what's happening.
Here's how to get to the id from the Power Automate Flow authoring experience:

[antoinec]

Oops sorry, please ignore the second half of my message, that would only be useful if you were using a custom model, which it seems you're not.

[MartineA]

Hi,

I am getting the error for any file, so it seems the action is systematically failing now.
I will send you the code in a private message, noting that the action used is "Recognize text in an image or a PDF document" (the code does not contain an ID but other parameters) and not "Extract information from forms".

Thank you very much!

[antoinec]

Thanks. Can you also share the flow URL you're seeing this behavior for? It would look something like https://us.flow.microsoft.com/manage/environments/xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx/flows/xxxxxxx....

[MartineA]

I have just sent you the link in the private message.

Thank you.

[antoinec]

Thanks! It seems that some the documents you're trying to process are too big (one example from telemetry was at 45 MB where our documentation states 20 MB as the maximum size). We will be updating the limit to 50 MB like we recently did for other model types in the next few weeks. In the interim, I would recommend filtering out documents which are too large from your automation. Note that the vast majority of the documents you're trying to process are smaller and do seem to still be going through correctly. Do let me know if you have smaller documents reproducing the problem, maybe you're seeing more than one issue.

[MartineA]

I just looked a little more, all the remaining files seem to be larger than 20 MB or have more than 5,000 characters (isn't there a limitation on the number of characters too ?), since they all have the same error.

And for smaller files, the action worked normally in a local test that I did few minutes ago.

I am just wondering how we could get around this, by maybe using a method to retrieve the content of the first page of the document only, since we really need to process all these files too and I only need to extract the information of the first page of the document.

Thank you.