Hey n8n Community,
Over the last year I have built document extraction into a bunch of client workflows. The models keep getting better, but the same handful of mistakes cause most of the pain, and almost none of them are about the model itself. Here are the five I run into most.
1. Building only the happy path.
Everyone tests on a clean PDF and ships, then production hands you a phone photo, a stamped scan, a two-language invoice. Test on your ugliest real samples first, because that is what the workflow actually gets fed.
2. Building for one layout.
It works on your one vendor’s invoice, then a new supplier’s template quietly breaks it. Test across as many real layouts as you can, and key off what a field means rather than where it sits on the page. If you do not have different layouts in your test data, I just ask my favourite LLM to generate a few alternative layouts with the same data points on them. It is a quick way to stress test against formats you have not seen yet.
3. Losing table structure.
Pull a table as one flat blob and rows drift, so quantities line up with the wrong description. This one I struggled with for a long time until I found the right way to do it: extract line items as structured rows with named columns, not free text. I made a short video walking through table extraction if it helps: https://youtu.be/qj4-zz9a13o
4. Ignoring what a 1% error rate means at volume.
99% sounds perfect until 1,000 docs a day is 10 silent errors daily, nothing crashes, the numbers are just wrong. Decide which fields must be right and put a check or review gate on those specifically. I usually wire an error route off those gates that pings me on Slack the moment something looks off, especially during onboarding, so I can react before the client even has to call.
5. Reaching for an agent when a rule would do.
People wrap an LLM around plain logic like date math or ID matching, and it breaks quietly instead of loudly. If you can sketch it on paper with the edge cases, make it deterministic and save the model for the fuzzy part. It is not just about robustness either, an agent can burn a lot of tokens on a task a simple rule would handle for free.
If you want to see what I have been building, I keep 25+ free workflow templates in one repo, most of them around document processing, and a star helps me out a lot if any of them save you time:
Happy to hear what everyone else struggles with most. Is it the messy inputs, the layouts that keep shifting, or something I have not run into yet?
Best,
Felix