OCR action incorrectly identifying text from PDF.

(0) Share

Report

Posted on by CU05081759-2

Hey,

I have been trying to create a flow in power automate to create extract some fields from my invoices dynamically, into excel. I am using the 'Recognize text in an image or document' action and input my invoice pdf via the 'Get file content' action. I have realized that some fields are not being extracted correctly such as in my invoice the number '156' is appearing as '951' in the output of the OCR. Another example includes the phrase 'GEL FRESH 700ML' being recognized as 'GEL FRESH JOUNL' as well the the number '38.50' being outputted as '18.50'. These are just some of the many incorrectly recognized texts I have provided. Anyone facing the same issue? Please provide a solution.

Thanks.

Categories:

AI Builder

I have the same question (0)

All responses (4)

Answers (0)

Sort by

Suggested answer

venturemavenwill 1,198 Super User 2026 Season 1 on at

Like
a
(0)

Report
Copy link

Link copied!

Have you tried to use AI Prompt in the flow instead?

This allows you to use the GPT4 model, which I find more reliable than standard OCR

1 people found this reply helpful.

Was this reply helpful? Yes No
Michael E. Gernaey 53,963 Moderator on at

Like
a
(0)

Report
Copy link

Link copied!

Hi @CU05081759-2

Are you training a model and using AI Builder to leverage your Model? There is an Invoice one OOB that you can try but I would train a model with some of your files so that you can see if you get better results.

You can use https://learn.microsoft.com/en-us/ai-builder/prebuilt-invoice-processing

Was this reply helpful? Yes No
CU05081759-2 2 on at

Like
a
(0)

Report
Copy link

Link copied!

Hi @Michael E. Gernaey,

I have used Invoice processing model previously, but the moment the position of any of my fields changes or in some cases where there are more of each, the identification fails.

Was this reply helpful? Yes No
Suggested answer

takolota1 4,980 Moderator on at

Like
a
(0)

Report
Copy link

Link copied!

You could try providing a GPT prompt both the raw file & the OCR text & see if that improves accuracy. But frankly I always tell my prompts to default to the OCR text values because I find them to be more accurate than whatever file reader they bolted on to GPT.

Maybe in some cases your prompt is getting confused on which number goes to which label & returning the wrong number, in which case I find it helpful to provide the OCR text as a text replica that replicates a lot of the vertical & horizontal spacing of text in the file.

https://community.powerplatform.com/galleries/gallery-posts/?postid=31e67eea-3f73-47b4-95b7-fe4a7b646389

Was this reply helpful? Yes No