Automated invoice processing with AI, Snowflake and CocoIndex - with incremental processing
I recently worked with a clothing manufacturer who wanted to simplify their invoice process. Every day, they receive around 20–22 supplier invoices in PDF format. All these invoices are stored in Azure Blob Storage. The finance team used to open each PDF manually and copy the details into their system. This took a lot of time and effort. On top of that, they already had a backlog of 8,000 old invoices waiting to be processed.
At first, I built a flow using n8n. This solution read the invoices from Azure Blob Storage, used Mistral AI to pull out the fields from each PDF, and then loaded the results into Snowflake. The setup worked fine for a while. But as the number of invoices grew, the workflow started to break. Debugging errors inside a no-code tool like n8n became harder and harder. That’s when I decided to switch to a coding solution.
I came across CocoIndex, an open-source ETL framework designed to transform data for AI, with support for real-time incremental processing. It allowed me to build a pipeline that was both reliable and scalable for this use case.