Zovda
Run a PDF through a text extractor and you get a stream of words. Run a question about the second column of a table through the same extractor and you get nothing useful. The words survived; the structure did not.
Document corpora fail on layout far more often than on language.
Reading order is a field, not a guess
Two columns, an inset figure and a sidebar will all flatten into one paragraph if nothing decides the reading order. A model trained on that learns to associate unrelated sentences.
Record the reading order as structure: paragraph index, column id, table id. Then a page can be reconstructed without the original file.
Tables need their grid
A table extracted as running text loses which value belongs to which row. Keep the grid, the header row, and the unit. A footnote that changes the meaning of a number belongs with the table, not at the end of the document.
Those three facts turn a table from decoration into trainable content.
Keep the page, not only the text
Layout-aware tasks need the page image as well as the extracted text. If the corpus only ships text, every downstream model has to guess where things were.
We keep the page image, the text layer, and the structural index together, so a buyer can train extraction and reasoning on the same material.
The checklist we run
- Is reading order explicit for multi-column pages?
- Do tables keep their grid, header row and units?
- Are footnotes attached to the element they modify?
- Can a page be reconstructed without the source file?
- Is the page image available alongside the text layer?
If a supplier cannot answer these, the corpus is a word salad with page numbers. Zovda builds document families with the Datuva catalogue so the structure travels with the text.
Where to start
Start from the questions your users ask about documents, then check whether the structure needed to answer them survived extraction. The corpus catalogue shows where document material sits across stages, and the image collection covers page imagery.
Tell us the document types and volume. We reply within one business day.