Handling multiple Invoice Formats with Document Understanding and integrate with OIC

Question

Hi Oracle Community,

Requirement:

We have a requirement to read PDF files/invoices through Oracle Integration Cloud (OIC), extract the information from the PDF, and then use that information/details to create transactions in Oracle via OIC. The PDF may reside in sources such as SFTP or any other supported locations.

Approach:

We understand, as part of OIC Gen 3, OIC provides a Document Understanding Adapter to connect with the OCI Document Understanding service(OCI's AI  service that extracts text, tables, and key-value pairs from digital and scanned documents) and have decided to proceed with this approach. However, one of the main challenges which we foresee is - if we receive multiple invoices from multiple vendors with different invoice layouts and formats, we would need to create and train separate custom models for each invoice type/template to achieve reliable extraction.

Could you please confirm whether our understanding is correct or not and if this is the recommended approach ? or is there a way to train a single model that can intelligently recognize and extract data from multiple invoice formats with out maintaining separate models for each template.

We would like to understand if anyone has implemented a similar solution , with multiple PDF formats/structures/use cases. Any guidance would be greatly appreciated!

Also if anyone has implemented a different approach(other than using Document Understanding) that provides similar or better results for data extraction, we would love to hear about that.

Thank you in advance for your support!

Brendan T · Answer

OCI AI Document has a model and demo that illustrates something like this

Gianni Ceresa · Answer

In the OCI AI Document Understanding service, the customization of models is optional.

By default, it will use the available model and work on whatever you feed it with. The model is "generic", it isn't like a OCR that would require a standardized structure to give a meaning to recognised information. Being an AI model, it will do its best to extract all the required info from any format of document you will provide.

If you decide to take the optional route to use a custom trained model, if you decide to use a model trained on a single type of document it will obviously not work well (or at all) on different types of documents. It's your duty when training the custom model to do it using all the expected type of documents.

But still, you maybe should first try to use the built-in models and see what the result looks like. It's maybe good enough to not require to train a custom model.

Gianni Ceresa · Answer

Welcome to the Oracle Analytics and AI community Atul.

The question is more about the OCI AI Document Understanding service than OIC, therefore this (the "Oracle AI Services Forum") is the right place (the fact the OCI AI service will be called from OIC isn't the main topic).