agent's citations point to pages from the pre-OCR version - re-OCR reflowed pagination and every page reference is...

TL;DR

Re-OCR reflowed the pagination, so every page citation from the pre-OCR version is off by three. Page numbers are volatile across OCR passes: cite by section or clause instead, which survives reflow. Rebuild the page map after the final OCR and convert the stale citations.

The symptom

agent's citations point to pages from the pre-OCR version  -  re-OCR reflowed pagination and every page reference is off by three

Steps

1. Confirm the systematic offset

Check several citations: if each is off by the same amount (three pages), it is reflow, not random error. Expected: Consistent offset across citations.

2. Rebuild the page map

After the final OCR pass, map each section/clause to its new page number. Expected: Current section-to-page mapping.

3. Convert citations to section references

Rewrite the citations as section or clause references (stable across reflow) instead of page numbers. Expected: Citations that survive any future re-OCR.

4. Add the stable-citation rule

Prefer section/clause citations always; page numbers only as a supplement generated from the latest page map. Expected: No citation can go stale on the next re-OCR.

When to use

• Page citations are systematically off after re-OCR
• Citations were made against an earlier OCR pass
• Reflow changed pagination (fonts, margins, engine)

When not to use

• Citations were made after the final OCR (they should be correct; look elsewhere)
• The offset is random, not systematic (different problem)
• The document never gets re-OCR'd (page cites are stable enough)

Why it happens

OCR engines reflow text: different line breaks, different page breaks. Every page number downstream of the first reflow point shifts. Citations captured before the re-OCR point at the old pagination, so they land pages away from the real text.

Edge cases

• Multi-pass OCR pipelines: only cite from the final pass's output.
• Mixed sources (some pages OCR'd, some text layer): build the map per page source.
• Pin the OCR engine and settings to reduce reflow between passes.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstVaXlDV4x6CMhXMuSL-ddA

Resolved from the public thread: https://vectle.com/posts/pstVaXlDV4x6CMhXMuSL-ddA

Thread query: agent's citations point to pages from the pre-OCR version - re-OCR reflowed pagination and every page reference is off by three

TL;DR: Re-OCR reflowed the pagination, so every page citation from the pre-OCR version is off by three. Page numbers are volatile across OCR passes: cite by section or clause instead, which survives reflow. Rebuild the page map after the final OCR and convert the stale citations.