How to validate PDF extraction before using the spreadsheet

Spend your trial checking the result you would actually use

Count the source checks, corrections and export steps in your comparison with manual entry. ParseMill lets you inspect invoice fields beside the PDF and correct them before downloading Excel, so you can evaluate a usable result rather than a completion message.

Inspect the interactive invoice sample, then try your own invoices. Eligible new workspaces get 10 pages for 7 days after email verification, with no card or automatic subscription.

Use three separate acceptance checks

1. Structure: are all expected pages, rows and columns present, without duplicate headers or added rows?
2. Values: does each field match the source, including units, precision, identifiers and qualifiers?
3. Evidence: does the source reference point to the actual supporting text, and can the reviewer open it?

For a new document class, start with a manually enumerated expected output. The five-row material example gives you 25 field values to compare across two pages. Getting those right would validate that fixture, not all supplier documents.

An extracted cell is still unreviewed

The material sample’s cells begin as unreviewed. Its density value 1,234 is preserved but flagged as unsupported for numeric interpretation. An exception is a reason to investigate, not proof the raw text is wrong. Likewise, an unflagged cell is not proof the extraction is right.

For your own invoice PDFs, use the invoice workspace to inspect extracted fields beside the source and save corrections before export. Keep your review notes with the handoff. The material sample above is a separate public demonstration, not a promise that the invoice workflow processes every technical report.

A worked review decision

Question: can the density value 1,234 g/cm3 be used in a calculation?

Evidence: page 2 contains exactly that text, and the report explicitly says the comma is ambiguous.

Decision: retain the raw string and leave numeric interpretation unresolved. Ask for the source’s number convention or another authoritative clarification. Do not silently choose 1.234 or 1234. The synthetic example intentionally provides no answer to that question.

Measure correction effort honestly

For your own test set, record review time, fields corrected, unresolved fields and why each correction was needed. Separate time spent understanding the source from time spent fixing extraction. We have not measured human correction time for the public sample, so no time-saving or accuracy percentage is claimed.

Read the reproducible evaluation methodology for the distinction between a fixture check and a broader benchmark.