Zovda
A scanned page run through text extraction produces something that looks like the page and answers none of the questions the page was designed to answer.
The text survives. The structure that told you which number belonged to which column does not.
Text is the easy layer
Character accuracy is the part everyone measures and the part that matters least at the end. A page of dense text with one swapped digit reads fine and teaches a wrong number.
What a downstream model needs is where each block sits, what it is, and what it belongs to.
The fields that keep a page intact
- Block type: heading, paragraph, table, caption, footnote, figure
- Reading order, explicit rather than inferred at load time
- Bounding box, in page coordinates with the page size
- For tables: row and column spans, header cells
- For figures: the caption it belongs to
- The page number and the document it came from
Without reading order, a two-column layout turns into interleaved nonsense at extraction time.
Tables are the hard case
A merged cell is a relationship, not a formatting detail. Once flattened, a total row and its subtotal become indistinguishable, and a model trained on that learns to add things that should not be added.
Keep the spans. Keep the empty cells as empty. A table with holes is more faithful than a table that has been quietly filled.
Check by reconstruction
The cheapest quality test is to rebuild the page from the record and compare it to the original. If the rebuild loses a caption or moves a column, the record is wrong, regardless of character accuracy.
Zovda builds these records through Datuva, so block type and reading order stay with the text. The image collection lists the full field set.
Tell us the document types you need to read. We reply within one business day.