Field note: the three edge cases that break document pipelines

Three things break document pipelines in production. None of them survive a generic OCR pass. We've seen all three more than once.

If your pipeline is being benchmarked against a clean test set and accuracy looks great, congratulations: you've validated a demo. The cases that follow are what production looks like.

1. Stamps and overlays on top of text

A 'received' stamp covering a date. A redaction box across half a paragraph. A signature looped through three lines of an address. Generic OCR sees one layer; the page has two.

What we do: a multi-pass approach. First pass identifies overlay regions (stamps, redactions, signatures) and masks them. Second pass extracts text from the unmasked regions. Third pass attempts to recover the partially-occluded content from contextual fields elsewhere in the document. Cross-reference against the entity-resolution graph. If the recovered value disagrees with what's elsewhere, route to a human. If it doesn't appear elsewhere, the field is marked 'occluded' rather than 'missing' — those are different states downstream.

2. Mixed-language paragraphs

A US filing with a paragraph quoted in German. A Spanish-language contract with English boilerplate at the bottom. A scanned letter where one sentence switches mid-line because the writer borrowed a phrase.

Single-language LLMs misclassify or skip these. Single-language OCR is worse — it'll silently drop characters that don't fit the alphabet it expects. The fix isn't a fancier model. It's segmentation: detect language at the span level (not the page level), route each span to the right extractor, then reassemble.

3. Visual tables that aren't tables

A row of indented bullets that someone formatted to look like a table. A two-column layout where the left column is figure captions and the right column is unrelated body text. A 'table' with no borders, just whitespace alignment.

Layout-aware models help, but they only get you so far. The reliable fix is to convert anything table-shaped into a graph of cells with explicit relationships — header, sibling, sub-row — and then ask the model to reason over the graph, not the visual structure. Tables that aren't tables semantically (the bulleted-list-pretending-to-be-a-table case) get marked as such and excluded from table extraction entirely. They go through the regular paragraph pipeline.

Edge cases like these are the difference between a working demo and a production system. They're also what eats the budget if you don't plan for them — every one of these patterns took us a couple of weeks to solve the first time and an afternoon to recognize the second.