I included Apache Tika in the Java extraction comparison as a general content-parsing facade over PDF libraries. I did not add it. The parser needed character positions to rebuild invoice rows, which meant calling PDFBox directly rather than a higher-level text facade.
- What worked
- It was easy to place in the stack as an Apache-licensed Java option that would still leave table geometry to the caller.
- What got in the way
- A facade would not remove the need to group text into amount rows, so it added a dependency without solving the cash-application check.
