Conference 9–10 May 2023 📍 Berlin, Germany

PDF Days Europe 2023 — The PDF Association's Annual Technical Conference

PDF Days Europe 2023 took place in Berlin on 9–10 May 2023, bringing together PDF software developers, document engineers, and standards experts for two days of technical sessions on PDF 2.0, accessibility, digital signatures, and the future of the format.

Official event page →

Key takeaways

PDF/A-4 (ISO 19005-4:2020) is the first PDF/A standard built on PDF 2.0, with variant suffixes (4e for engineering, 4f for embedded files) replacing the former conformance level A/B/U structure.
PDF/UA-2 was in final balloting at the 2023 conference, with detailed review of MathML requirements, revised Matterhorn Protocol checkpoints, and validator previews from PAC and veraPDF.
AI document processing practitioners provided the first significant conference-level empirical evidence that PDF quality (text extraction order, hyphenation, scan-only pages) directly determines LLM pipeline performance.
AI-assisted PDF tagging had reached a maturity level where it was being evaluated as a production tool for accessibility remediation workflows.
The hybrid conference format was fully mature: in-person attendees in Berlin with simultaneous remote streaming and post-event recording access.

PDF Days Europe 2023 was held in Berlin in May 2023, organised by the PDF Association. Following the 2022 return to in-person, the 2023 conference operated fully in the established hybrid format with simultaneous in-person and remote attendance. The primary technical agenda centred on three developments: accelerating PDF 2.0 adoption, the near-final state of PDF/UA-2, and the introduction of PDF/A-4 — the first PDF/A standard built on PDF 2.0 rather than PDF 1.x. By 2023, AI-assisted document processing had moved from a speculative topic to an active deployment concern. Early adopters of large language models in document analysis pipelines were discovering that the quality of the input PDFs — tagging, structure, text extraction fidelity — directly determined the quality of AI outputs. This practical feedback loop gave the 2023 conference a new dimension: PDF quality was no longer just a standards compliance question but a direct determinant of AI pipeline performance.

PDF/A-4: the PDF 2.0 archiving standard

PDF/A-4 (ISO 19005-4:2020) was the major standards topic at PDF Days Europe 2023. Published in 2020 (though adoption accelerated in 2022–2023), PDF/A-4 is the first PDF/A standard built on PDF 2.0 (ISO 32000-2) rather than PDF 1.4–1.7. Key differences from earlier PDF/A standards: PDF 2.0 foundation — PDF/A-4 inherits all PDF 2.0 features not otherwise restricted by the PDF/A-4 specification. This includes new encryption options, new structure elements, and the improved tagged PDF model. No conformance level A/B/U split — Earlier PDF/A standards had conformance levels (a, b, u). PDF/A-4 uses a different approach: PDF/A-4 (no suffix), PDF/A-4e (for engineering, permits 3D/rich media), and PDF/A-4f (permits embedded files, similar to PDF/A-3). The former level structure is replaced by a capabilities-based approach. PDF/A-4e for engineering — PDF/A-4e explicitly supports 3D and rich media content for engineering drawings and product data, which previous PDF/A standards either excluded or handled ambiguously. PDF/A-4f for embedded files — PDF/A-4f permits embedded files with AFRelationship declarations — equivalent to PDF/A-3's embedded file capability but now in the PDF 2.0 framework. This maintains the ZUGFeRD/Factur-X invoice workflow capability. Sessions at PDF Days Europe 2023 discussed the migration path from PDF/A-1 and PDF/A-2 document archives to PDF/A-4, validator support status, and the longer timeline expected before enterprise document management systems would support PDF/A-4 natively.

PDF/UA-2: final stages before publication

At PDF Days Europe 2023, PDF/UA-2 (which would be published as ISO 14289-2:2024) was in its final balloting stages. Sessions provided detailed reviews of the near-final specification content: Revised Matterhorn Protocol — The Matterhorn Protocol test suite for PDF/UA conformance was being revised for PDF/UA-2. The original 136 checkpoints for PDF/UA-1 were being updated to reflect the new requirements, and new checkpoints were being added for the new structure elements (/Aside, /FENote) and MathML integration. MathML in practice — For scientific and academic publishers, MathML integration was one of the most significant PDF/UA-2 changes. Sessions from scientific publishers and accessibility specialists discussed the practical challenge of converting LaTeX mathematics to MathML for PDF embedding. The LaTeXML processor and the newer LaTeX-to-MathML pipeline via Pandoc were reviewed. The key requirement: each mathematical expression must be associated with a MathML representation that a screen reader supporting MathML (JAWS, NVDA + MathPlayer) can verbalise. Artifact handling improvements — PDF/UA-2 refined the specification of how decorative and presentational content is marked as an artifact (invisible to screen readers). Header/footer repetition, background images, decorative rules, and page numbers with correct label semantics were all addressed more precisely in PDF/UA-2 than in PDF/UA-1. PDF/UA-2 validator preview — PAC and veraPDF teams presented their PDF/UA-2 validation development status. Both were targeting availability coinciding with or shortly after the standard's publication.

AI and PDF quality: the feedback loop

PDF Days Europe 2023 was the first conference where AI-assisted document processing was a substantial conference topic, reflecting actual deployment experience rather than speculation. LLM quality is PDF quality — Practitioners deploying large language models on enterprise document corpora found that PDF text extraction quality directly determined LLM performance. Three PDF quality issues caused consistent LLM degradation: 1. Poor text extraction order — PDFs where the text content stream order did not match the visual reading order (common in multi-column layouts, scanned PDFs with OCR, and PDFs created from complex desktop publishing tools) caused LLMs to receive jumbled context. A two-column report where column 2 content was interleaved with column 1 in the text stream looked coherent on screen but produced nonsensical context for the model. 2. Hyphenation artifacts — PDFs that soft-hyphenate words at line breaks often encode the hyphenated form in the text stream. LLMs trained on unhyphenated text encounter vocabulary they did not see in training when "manage- ment" is not normalised to "management" during extraction. 3. Image-only pages — Scanned PDFs with no OCR text layer produce zero text for LLM context. AI pipelines deploying document understanding on mixed corpora needed to identify and OCR scan-only PDFs before indexing. AI-assisted tagging — On the creation side, AI-based automatic PDF tagging was mature enough to be reviewed as a production tool. Layout analysis models (trained on labeled PDF datasets) could reliably identify headings, paragraphs, and tables in typical business documents, significantly reducing the human effort required for accessibility remediation.

Frequently asked questions

What is PDF/A-4 and how does it differ from PDF/A-3?

PDF/A-4 (ISO 19005-4:2020) is the PDF/A standard based on PDF 2.0, while PDF/A-3 (ISO 19005-3:2012) is based on PDF 1.7. Key differences: (1) PDF 2.0 features — PDF/A-4 has access to PDF 2.0's new structure elements, improved tagged PDF model, and updated colour management. (2) No conformance level letters — PDF/A-3 has levels a (accessible), b (basic), u (unicode). PDF/A-4 uses variant designators: PDF/A-4e permits 3D/rich media content; PDF/A-4f permits embedded files with AFRelationship. (3) Tagged PDF requirements — PDF/A-4 does not require tagging (unlike PDF/A-4a which was not defined — the accessible requirement is implied by pairing with PDF/UA-2). (4) The embedded file capability in PDF/A-4f is functionally equivalent to PDF/A-3 — ZUGFeRD/Factur-X workflows using PDF/A-3 can migrate to PDF/A-4f.

How does AI use PDF content and what makes a PDF AI-readable?

AI systems (LLMs, document understanding models) access PDF content primarily through text extraction: they extract the text content stream from each PDF page and use it as input context. An AI-readable PDF has: (1) Selectable, searchable text — not a scan-only image. OCR text layers from tools like Tesseract or Acrobat's OCR improve extractability. (2) Correct text extraction order — the text stream order should match reading order, not rendering order. Properly tagged PDFs (with a logical structure tree) are inherently better for this. (3) Clean Unicode mapping — the ToUnicode CMap in each font must correctly map glyph codes to Unicode values; fonts without ToUnicode produce garbled text extraction. (4) No text rendered as curves — some PDF export pipelines outline text (convert text to vector curves) for font-independent rendering, which completely prevents text extraction. AI systems reading outlined text see nothing.

What is the Matterhorn Protocol?

The Matterhorn Protocol is a test suite for PDF/UA-1 (ISO 14289-1) conformance, published by the PDF Association. It defines 136 failure conditions organised into 31 categories, each of which represents a way in which a PDF can fail to meet PDF/UA-1 requirements. The failure conditions are categorised as either machine-checkable (can be verified automatically by a validator) or human-checkable (require human judgment). For example, failure condition 01-003 (Document does not have a valid language identifier) is machine-checkable; failure condition 09-004 (Content marked as Artifact is visually significant) requires human evaluation. PAC (PDF Accessibility Checker) implements the machine-checkable Matterhorn checkpoints. The Matterhorn Protocol for PDF/UA-2 was under development as of 2023–2024, with a revised version expected at or after the PDF/UA-2 publication.

Try it in your browser

The PDF tools and standards discussed at this event are available free — no account needed.

Edit PDF Annotate PDF Sign PDF Open editor