A PDF is a view; data needs a structure
By Oliver Schmalzl · Written: · Archive discovery date:
23 April 2026 · Archive date
A PDF can look completely clear on screen. Two columns, a heading and a small table are easy to spot at once. For automatic processing, however, these are separate tasks: recognising characters, working out their order and keeping the relationships between the contents.
A sample sheet you make yourself shows this especially clearly. On the left there is a description, on the right a list of materials, and below that a table with three invented paper sizes.
Correctly recognised words can be put together wrongly
If the words are read line by line straight across both columns, the result is a text nobody meant. A table also loses meaning when numbers no longer belong to their column headings.
With a scanned page, character recognition is added on top. A badly recognised "1" can change a measurement. That is why it makes sense to record errors in characters, order and structure separately.
OpenDataLoader PDF describes how it outputs structured document elements. The find is interesting because it can take position and order into account as well as text. No independent quality comparison with other tools is drawn from it here.
A small test can be checked
For the sample sheet, the desired reading order is fixed in advance. After processing, headings, paragraphs and table values are compared one by one. It should also stay visible which page a statement comes from.
A later AI summary can only work with the structure that has been kept. If the amount of material has already been assigned to the wrong row, a nice answer sentence does not make the error any better. The most important result of the exercise is therefore a traceable link between the original view and the extracted data.
Original finds
These links were the starting point for this post. The assessment in the text is mine; the linked posts are not my own tests.
- German-OCR-3 and Nano for local recognition of German documents; check the quality and privacy claims.
- OpenDataLoader PDF with structured elements and position data; check the benchmark claims.
- GOCR as a compact local German OCR for text and positions; check the benchmark figures.
- PixelRAG processes visual page renderings instead of purely extracted text; check the original paper and benchmarks.
About the date: This post was written on 11 October 2026. The archive date matches the earliest related find in the collection, not a publication at that time. Technical assessments and sources added later refer to the time of writing.