SoulkeyResearch

Practical guide / sample v1.0

Repeated PDF headers can become extra CSV records

By SoulkeyResearch · 30 September 2026

A table can continue onto a second page while repeating its column labels for the reader. An extractor may correctly return the table on each page, including both headers. The extra work is deciding which rows are data.

Our two-page sample contains 12 purchasing records and two identical header rows. Concatenating pdfplumber's page-level tables produces 14 rows. Removing only the first header leaves 13 rows, including a header in the middle of the data.

The correction in this sample is deliberately simple: remove a row only when all six cells exactly match the expected column labels. That produces 12 records and 72 of 72 exact data cells. The package keeps both the naive and cleaned CSVs so the difference is inspectable.

This rule is narrow. A real file might use abbreviated labels, wrapped headings or a different table on page two. A broad rule such as “delete every row containing Date” could discard a legitimate record.

For a repeatable workflow:

  1. Record the expected column labels.
  2. Keep the raw extraction before cleaning.
  3. Apply a documented, specific header rule.
  4. Compare row counts and record IDs across pages.
  5. Flag unexpected layouts instead of silently dropping them.

A clean-looking CSV is not proof that every record survived. Header handling is part of the extraction workflow and should be reported alongside the tool name.

These results cover one synthetic repeated-header layout. No commercial converter has been tested.

Inspect the before and after

StageRowsWhat remains
Concatenate both extracted tables1412 records + 2 headers
Drop only the first header1312 records + 1 header
Remove exact six-cell header matches1212 records

Download the naive output and the filtered output. The rule removes matching headers only; it does not repair data values.

Scope reminder: these are three clean synthetic fixtures. The OCR test used known cell positions. No paid converter was tested, and no result here is a general accuracy or vendor-ranking claim.