Hi there! We're working with a client who has a bunch of Excel and PDF files that come in all sorts of formats, and they're likely to keep changing. Some files have data laid out in standard tables, while others may use pivot tables or more complex layouts. We're tasked with extracting useful information from these files and loading it into normalized tables. With so many different templates, and the possibility of new ones appearing, what's the best pipeline and tech stack to use? Should we focus on Document Intelligence or Content Understanding? Are these technologies reliable enough for accurately deciphering file formats and extracting data?
4 Answers
In my experience using Document Intelligence for over five years, it's been pretty reliable. Content Understanding seems newer and I’m not entirely sure about its advantages yet. With more than 50 document types and custom templates, we’ve found Document Intelligence works well for our extraction needs.
Why not try both techs and see which one matches your requirements better? Sometimes firsthand experience is the best way to decide.
You might need to put in the effort to develop a good classification model first, then follow up with an extraction model. In our case, we have a classification model that manages over 80 classes with about 96% accuracy. After classifying, you can use key-value pairs along with custom logic to get the desired results. Depending on your structure, you can also employ transformer models to align everything properly for normalization. We’ve been doing this at a top-tier rate for enterprise clients.
I recommend leveraging AI to convert your data into JSON format. With easy access to open-source solutions nowadays, there's really no reason to get locked into a pricey cloud provider for this task. It's simpler than ever, especially with tools that can break documentation down into markdown format while using a cost-effective LLM for structured outputs.
Totally agree! It’s wild how much flexibility we have with open-source LLMs these days. Makes it a lot easier to avoid vendor lock-in.
One thing to note is that Content Understanding can extract data from audio and video as well. It really shines with custom tasks where you can design schemas for it to identify information.