# PDFluent SDK 1.4.0: Better PDF Extraction

Source: https://pdfluent.com/blog/sdk-1-4-0/

All posts

Product 2026-11-13 3 min read by [Jasper de Winter](https://www.linkedin.com/in/jasperdewinter/), Founder, PDFluent

# PDFluent SDK 1.4.0 raises the extraction scores and fixes empty OCR pages

sdk release extraction ocr pdf-a benchmark

PDFluent SDK 1.4.0 is live on crates.io, PyPI, npm, NuGet and the PDFluent Maven repository. It scores higher than 1.3.0 on reading order, table structure and heading hierarchy on opendataloader-bench*, and it fixes OCR that returned an empty page without an error. It also reads more than PDF.

Upgrade if you run OCR in production, or if your documents carry formulas. Re-run the model install once, as described at the end.

## The scores

On opendataloader-bench, SDK 1.4.0 in its default auto tier scores 0.9282 overall*, 0.9463 on reading order, 0.9140 on table structure and 0.8767 on heading hierarchy, on all 200 documents, with the benchmark's own harness and evaluator.

The same harness and settings gave SDK 1.3.0 0.9154, 0.9359, 0.8695 and 0.8678.

| Row | Overall | Reading order | Table structure | Heading hierarchy |
| --- | --- | --- | --- | --- |
| PDFluent 1.4.0 auto | 0.9282 | 0.9463 | 0.9140 | 0.8767 |
| PDFluent 1.3.0 auto | 0.9154 | 0.9359 | 0.8695 | 0.8678 |

*Measured 11 Nov 2026 on the final macOS arm64 release wheel (Apple M1 Pro, macOS 26.5), quality="auto" with keep_furniture=True and the layout, table and formula models installed. Table structure counts 42 documents, heading hierarchy 107, reading order and overall 200.

- The scores are measured on the benchmark's own 200 documents.
- One run, no confidence interval.
- Both rows keep page headers and footers, because the benchmark's ground truth counts them as content.
- The table structure score depends on the platform, so the comparison is Mac against Mac.

The 1.3.0 results, with the comparison against the published leaderboard as it stood on 3 November 2026, stay on the [benchmarks page](https://pdfluent.com/benchmarks/) next to these.

## OCR and empty pages

SDK 1.4.0 fixes OCR returning an empty page without an error on installations whose Tesseract data folder holds only traineddata files and has no configs folder. SDK 1.3.0 and earlier are affected, so run OCR again on the originals if that was your setup.

There is no hotfix for 1.3.0 or earlier. The fix ships in 1.4.0 only.

## More inputs

The Python package reads CSV, TSV, JSON, XML, RTF, HTML, Markdown, plain text and e-mail (.eml) input into blocks, the same block structure a PDF page produces, and writes them as Markdown or JSON.

Every element in the canonical block output carries an element_id, and the same input gives the same ids in separate runs.

## Formulas

In SDK 1.4.0 the own formula model reads formula-heavy documents by default. A document is routed to it only when the layout model finds at least three display formulas, at least one per two pages, and the formulas cover at most a tenth of the page on average. formula_model=False switches it off, formula_latex=True reads every formula. A document the route leaves alone reads as in 1.3.0.

## PDF/A conversion

SDK 1.4.0 PDF/A conversion, measured once on a sealed holdout of 1,000 documents (400 Zenodo, 350 Wikimedia Commons, 250 arXiv, 2007 to 2026): 993 of 1,000 converted files conform (99.3%), checked with veraPDF 1.28.2. Real-word retention, 5th percentile over 841 documents: 98.4%.

Both internal gates, 99.5% conformance and 99.5% retention, were not met. The measurement used the converter as it stood on 7 November 2026; later changes were not measured again. Measured 7 Nov 2026.

## Known issues

Known issue: on the olmOCR-bench headers and footers category, SDK 1.4.0 quality="auto" scores 0.0013 lower than SDK 1.3.0 (one run, 6 November 2026). The other seven categories are equal or higher. keep_furniture=True keeps page headers and footers on every tier.

Pages where display equations cover more than a tenth of the page on average stay outside the formula route on purpose, so they read as in 1.3.0.

The PyMuPDF-compatible layer stays a Preview in SDK 1.4.0. [How it works](https://pdfluent.com/how-to/pymupdf-compat/).

## Models and licences

The model files come from a new models release, models-v2, each with a notice. The layout and formula models are PDFluent's own, under the same terms as the SDK. The layout models are fine-tuned from IBM's docling-layout-heron-101 (Apache-2.0). The table model is Microsoft's Table Transformer, unmodified, under MIT.

The PDFluent SDK is free to use under the AGPL-3.0, and a commercial licence is the alternative if you do not want to publish your own source. See [SDK pricing](https://pdfluent.com/sdk/pricing/).

## Install and upgrade

bash

```
pip install pdfluent==1.4.0
cargo add pdfluent@1.4.0
npm install @pdfluent/node@1.4.0
npm install @pdfluent/sdk-wasm@1.4.0
dotnet add package PDFluent --version 1.4.0
com.pdfluent:pdfluent:1.4.0
pdfluent models install layout tables formula
```

- Re-run `pdfluent models install layout tables formula` to fetch the models and their notices from models-v2.
- Pass `formula_model=False` to keep the 1.3.0 reading of formula-heavy documents.
- If OCR ran on an installation without a Tesseract configs folder, run it again on the originals.

- [GitHub release v1.4.0](https://github.com/pdfluent/pdfluent-sdk/releases/tag/v1.4.0)
- [Changelog](https://pdfluent.com/changelog/)
- [Benchmarks](https://pdfluent.com/benchmarks/#odl-140)
- [The SDK 1.3.0 release post](https://pdfluent.com/blog/sdk-1-3-0/)

Report problems or ask questions in [GitHub Discussions](https://github.com/pdfluent/pdfluent-sdk/discussions).

## Frequently asked questions

What does PDFluent SDK 1.4.0 score on opendataloader-bench?

On opendataloader-bench, SDK 1.4.0 in its default auto tier scores 0.9282 overall*, 0.9463 on reading order, 0.9140 on table structure and 0.8767 on heading hierarchy, on all 200 documents, with the benchmark's own harness and evaluator. The same harness and settings gave SDK 1.3.0 0.9154, 0.9359, 0.8695 and 0.8678. One run, measured 11 Nov 2026 on a Mac.

Which PDFluent SDK versions have the empty OCR page bug?

SDK 1.4.0 fixes OCR returning an empty page without an error on installations whose Tesseract data folder holds only traineddata files and has no configs folder. SDK 1.3.0 and earlier are affected, so run OCR again on the originals if that was your setup.

Which input formats does PDFluent SDK 1.4.0 read besides PDF?

The Python package reads CSV, TSV, JSON, XML, RTF, HTML, Markdown, plain text and e-mail (.eml) input into blocks, the same block structure a PDF page produces, and writes them as Markdown or JSON.

Does PDFluent SDK 1.4.0 convert every document to PDF/A?

No. SDK 1.4.0 PDF/A conversion, measured once on a sealed holdout of 1,000 documents (400 Zenodo, 350 Wikimedia Commons, 250 arXiv, 2007 to 2026): 993 of 1,000 converted files conform (99.3%), checked with veraPDF 1.28.2. Both internal gates, 99.5% conformance and 99.5% retention, were not met. The measurement used the converter as it stood on 7 November 2026; later changes were not measured again.

What do I run after upgrading to PDFluent SDK 1.4.0?

Run pdfluent models install layout tables formula to fetch the models and their notices from models-v2. Pass formula_model=False to keep the 1.3.0 reading of formula-heavy documents.

Is PDFluent SDK 1.4.0 free to use?

The PDFluent SDK is free to use under the AGPL-3.0, and a commercial licence is the alternative if you do not want to publish your own source.

PyMuPDF and MuPDF are trademarks of Artifex Software, Inc. PDFluent is an independent product of Innovation Trigger B.V. and is not affiliated with or endorsed by Artifex.

[All posts](https://pdfluent.com/blog/)[RSS feed](https://pdfluent.com/rss.xml)
