The takeaway

A document pipeline is a small ETL system bolted to your agent. Treat it with the same rigor: validate inputs, store originals immutably, isolate the messy conversion step and make it fail soft, extract with citation-grade location markers, version everything, and use an explicit status state machine so the whole thing is recoverable and ready to go asynchronous when scale demands. The agent gets the credit, but the pipeline is what makes its document work trustworthy.