Form Recognizer from PA

[IshDoss]

Hi Everyone,

I have a requirement to extract data from 500 different invoice documents. I wanted to make use of the Invoice prebuilt model in the form recognizer. I'm trying to call the form recognizer from power automate. Here are my queries.

1. Is AI builder in PA cheaper than form recognizer for 500+ invoice pdf docs extraction? Will form recognizer be the better approach?

2. Can the prebuilt model be used without a Storage Account and blob container in the Azure portal?

3. How to pass the input file from PA when using a form recognizer connector?

@Brad_Groux  Kindly suggest

[Sundeep_Malik]

Hey @IshDoss 

I have used form recognizer in Power Automate. So, I might be able to answer some of your questions.

I have the answer for the 3rd question. 

-> For me I had a scheduled trigger that would run everyday. Then I used to run list blobs, this would pick all the files from the blob storage. After that I used apply to each loop and sent the list I got from list blobs.

Inside that loop, I used an action called Create SAS URI by path (V2).

This would create a link for that file. This link we can use in an http request body to send to form recognizer. 

By this way we can input files from PA To form recognizer. 

Inside HTTP Body:

I am not sure about the second question.

For the 1st question, I am tagging AI Builder admins, they might be able to help.

@JoeF-MSFT @plarrue 

I hope this helps 🙂

[IshDoss]

Thanks a lot for sharing, Sundeep! Is this for a custom model? My input PDFs are invoice documents, so the prebuilt model works fine. All the blogs have mentioned the need for a storage account and blob container for a custom model. Not sure if it is required for Prebuilt as well. Thanks for your input!

[Sundeep_Malik]

@IshDoss 

Yes mine is a custom model

[IshDoss]

Thank you! Also, I'm trying to parse the lengthy JSON output (attached )that I received from the GET endpoint. Do you have any sample on how to get each field values string in an Excel sheet? I tried this way 

body('Parse_JSON')?['analyzeResult']?['documents'][0]?['TotalDue']?['valueString']. But its returning null. I'm not sure how to pass the items and header values..

[Sundeep_Malik]

@IshDoss 

Nice work in deciphering the json data. My data was also really huge and I took a lot of time to decipher it. Hehe.

You missed a field called: fields.

Rest everything looks good in your expression. 

So, an easy way to decipher the big json structure, use Notepad++. That will give you the exact brackets. So you will be easily able to know what is inside what. 

For me I have done it in the following way(This technique I used 6 months back, there can be many alternatives to it):

Saved the body in a object variable:

Took another variable and stored analyze result part in it:

Then used another object variable to store documents:

I know above actions can be used together in just a single variable, but this I did 6 months ago, now I will do it differently.

Next I used select action for fields variable:

After that used a parse json action and passed above select action in it. 

Then you can use an apply to each action, pass the body of parse json in it and extract all the values.

I hope this helps 🙂

[IshDoss]

Wow. Thank you so much for sharing in detail! I've fixed it with your support!

[IshDoss]

For the prebuilt model like Invoice, I tried like this and its working!

[IshDoss]

Sundeep,

Are all your input files in the azure data lake? Could you share the complete sequence of the flow, if possible?

[takolota]

So I built this template that processes any document just using the less expensive image recognition AI Builder service & GPT.

It seems much less expensive with a much lower barrier to entry only requiring a premium Power Automate license with 5000 AI Builder credits per month & some on-going LLM / GPT costs…

https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/Extract-Data-From-PDFs-and-Images-With-G...

If I’m looking at this right, this method should cost 

Image recognition +.0017

GPT prompt & output +.007

Flow actions +.00375

So about $ .013 per page.

Less expensive than any of the other microsoft document/receipt/invoice processing options. And several of those take about 30 seconds to run. 
And the 5000 AI Builder credits that come with the now $15 premium Power Automate license is enough for 1500 Image recognition pages. So you could probably process 500-750 documents per month with just 1 premium Power Automate license.

So it seems this method is faster, less expensive, easier to adjust, & requires less set-up than other standard AI builder set-ups.

Once the GPT actions are in general availability, I don’t know why I’d use any of the other services. 🤷‍