Extract multiple fields where the number of the fields is not consistent across documents

[ManfroLM]

Hello everyone,

I am looking for a way to automate the extraction of text from various client documents, where the amount of information varies for each document and clients.

The objective is to extract multiple fields (in the format fieldName: fieldValue) that are organized into sections, with the number of fields not being consistent across documents (for example: one client document contains 10 sections of fields, while another client document contains 20 sections of fields).

Initially, I attempted to achieve this using Ai Builder and the Document Processing Model. However, I came across information suggesting that entity extraction might be more appropriate for my needs.

Therefore, I am wondering if there is a method to train an AI model to extract these variable information elements from each document. 

[ErinIbbetson]

I would make a new AI Model

This allows you to make a Model that you can train. Click New AI model.

For the first screen choose Structured Documents (first box). The second page is adding fields to capture. Not all the fields have to exist on your documents. If they vary in name, you may want to call them Field1 or a generalist name of what is to be contained inside each one. Make fields for all potential document sizes.

The third page is adding a collection of documents. You can have a collection which is a group of similar documents but you can have more than one collection. It may be useful to have documents with the same amount of fields in separate collections.

For each document in a collection, you need to map the data to your fields. This trains the AI model for similar future documents. The more documents you have, the better it will be, but you will have to map more. They recommend at least 5 but I would present it with as many as you can, within reason, I did 50; as they must be mapped manually. There are short guided tour popups when mapping for the first time. You can click on individual words to map or hold+drag a box over a series of fields. If you need to map a table, drag over the table and mark as a table element.

If at any time you did not make a field from before, click BACK to add more fields, then NEXT to go to the map screen and map them. Line items for Invoices/Orders/Quotes etc should be tables, which it prompts to add columns into the table for mapping.

The last page is a summary, it will take some time to train the model, within 15 minutes or less for a few documents. You can refresh the AI Model page where it shows your models in a list, this is where it will say it is ready for review. It will show a summary page of accuracy per field extracted.

You can also use QuickTest to upload a file to see what it would do.

[takolota]

Alternatively you may use a more generalized GPT set up: https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/Extract-Data-From-PDFs-and-Images-With-G...