Skip to content
kenjtecJournal
Search the Journal

Docling makes PDFs usable for AI: why layout is part of the knowledge

When importing a PDF, a lost column can change the meaning. A reference document you create yourself shows what a structured export has to preserve.

Hanne cuts up a two-column sheet and sorts the pieces into shallow bowls. AI illustration.

On the left is the material for the basic course, on the right the additional material for the advanced course. After text extraction, the words alternate line by line between the two columns. All the terms are there, but the instructions are no longer correct. This is exactly why layout is part of the knowledge in document processing.

Docling prepares documents for further processing. The official advanced options also describe local use with prepared model artefacts. The FAQ adds prerequisites for offline operation. A complete export has to do more than deliver as many words as possible.

A reference document with three known relationships

First build a small document of your own. It contains two columns with different types of course, a table with a shared heading and a restriction below the table. Because you design the original yourself, you can describe its relationships before processing.

The following HTML template, written by us, can be opened in the browser and output as a PDF using the print function. It is a teaching document, not a claimed parser output. You should look at the actual print rendering before the test.

<!doctype html>
<meta charset="utf-8">
<style>
.spalten { display: grid; grid-template-columns: 1fr 1fr; gap: 24px; }
table { border-collapse: collapse; margin-top: 24px; }
th, td { border: 1px solid; padding: 8px; }
</style>
<section class="spalten">
  <div><h2>Basic course</h2><p>Four folders and two pens.</p></div>
  <div><h2>Advanced course</h2><p>Additionally, two exercise books.</p></div>
</section>
<table>
  <tr><th colspan="2">Material per group</th></tr>
  <tr><th>Item</th><th>Quantity</th></tr>
  <tr><td>Folders</td><td>4</td></tr>
</table>
<p>Open sealed packages only after approval.</p>

Note the target structure: the basic course materials stay together. “Additionally” refers to the advanced course. The number four belongs to the Folders table row. The approval condition remains accessible as a restriction. This list describes meaning, not just the order of individual characters.

Keeping preparation and processing separate

An existing Docling installation may download models on first use. For a planned isolated test, the required artefacts must be in place beforehand. The documentation names a targeted model download for this, followed by pointing to the prepared path.

docling-tools models download
docling --artifacts-path="./vorbereitete-modelle" referenz.pdf

The first command is a download option that is explicitly meant to be run in advance. In the second example, you must use the actual existing artefact path; the directory name shown is an example name of our own. These commands were not run for this article. A model download and the later document processing are different operations.

Also record which pipeline and which options you actually use. An application started locally can involve additional services if it is configured to do so. Offline operation follows from the prepared set-up, not from the program name alone.

Judging the export against the original

Open a supported structured export from the version you have chosen and compare it with the target structure. First look for the two course sections. Then check the table heading and the assignment of the number. Finally, examine whether the approval condition is preserved.

A long output can contain the same content several times or lose relationships. A shorter output can be more helpful for the subsequent search if headings, paragraphs and table references fit together. Sheer volume of text is therefore not a sufficient quality measure.

For a second round, change only one layout feature, such as the spacing between the columns. Save the new PDF file separately. This lets you relate an observed change to a known input condition instead of comparing a different parser run with a document that has changed unnoticed.

Why the later AI depends on this

A search or answer can only work with the information that is actually provided to it. If two separate lists become a single one, even an assistant that phrases things well can adopt the wrong assignment of material. A source link alone does not make this lost structure visible again.

The small test therefore helps even before a large archive import. On a document you understand, you check which relationships your pipeline preserves and where you need to rework. Docling is one processing step within a chain. Technical quality shows in whether the meaning remains traceable from the original through to later use.

Sources and status

Sources as of 25 October 2026. Docling advanced options — offline usage · Docling FAQ — completely offline.

Hamburg is our starting point.

The useful part travels with you.

Practical AI from Hamburg, explained for everyone.

Saved articles