Document parsing pipeline for purchase order acknowledgments

Overview

An e-commerce distribution team receives purchase order acknowledgments from many suppliers, each in its own format. We built the parsing layer that turns those documents into structured data with text and vision language models, one schema and parser per supplier and customer account, and we keep fixing errors as suppliers change.

The challenge

Every supplier words and lays out its acknowledgments differently, and small mistakes such as a wrong unit of measure or promised date reach the order data.

What we built

1. A JSON schema and a parser for each supplier and customer account, built on a document parsing service
2. Testing of text-based and vision-based extraction models for each supplier
3. Fixes for parsing errors such as unit-of-measure mismatches and promised dates
4. Tuning of parsers for accuracy and efficiency
5. Purchase order PDF designs for several customers
6. Support and handover notes for resolving parsing errors