Separate a prose specification from a real data table

Where the extra work starts

Recognize that difference before choosing a table-extraction workflow so you do not spend the trial repairing a result the input could never support.

For supplier reports and technical tables, start with ParseMill’s synthetic material-register example. Compare the source references and output with the register you need, then discuss a document pilot with an agreed scope and fee. The self-service trial covers invoice PDFs.

The distinction is practical: a flat-table workflow needs consistent relationships among headings, rows and values. It does not become a general document interpretation system because several numbers happen to align.

A practical way to get it right

Inspect the target region and identify repeated column roles. Ask whether several rows share the same fields and whether a value's meaning can be determined from its row and heading. Then look for exceptions such as prose sentences, diagrams, section labels and values explained only in surrounding notes.

Write a candidate row definition. For a real test table, it might be one test result per sample. For a product specification paragraph, there may be no meaningful repeated row without a separate manual interpretation step. Record that difference instead of forcing both into the same schema.

If the source contains a genuine table inside a larger prose page, inventory that region separately and preserve its surrounding qualifications. Test only the supported region with a method that actually provides the required boundary control; do not assume the entire page is one table.

W3C: PDF table markup describes tables through rows, headers and data cells. Use those relationships to distinguish a data table from visually nearby prose.

What this looks like on a real task

Synthetic example. The people, references and figures below are invented to demonstrate the method.

The first region has a repeated schema. The others may contain useful facts, but their relationships require a different approach. Copy the region, structure and method columns to document the boundary. Do not describe an excluded prose section as missing extraction when it was never within the declared table task.

Catch the cases that create rework

A decorative grid can make prose look tabular, while a genuine table may have no visible borders. Judge the repeated relationships, not the presence of lines. A short two-row table can still be legitimate; the issue is whether its field meanings are consistent and reviewable.

If a number depends on a sentence outside the region, preserve that qualification and ask the reviewer whether a standalone cell would be misleading.

Microsoft: Power Query PDF connector imports detected tables and exposes them for selection. Inspect the detected region and its meaning instead of assuming every page element belongs to a table.

Decide whether the change is worth keeping

The fit decision should identify the exact table region, row meaning and excluded content. A colleague should be able to repeat the classification from the source. Use a bounded trial only where the row model is clear, and retain a manual route for prose or diagram content rather than promising arbitrary PDF interpretation.