All posts
Product2026-11-184 min readby Jasper de Winter, Founder, PDFluent

PDFluent SDK 1.5.0 adds documents, OCR and columnar output

sdkreleasepythonextractionocrbenchmarkparquet

PDFluent SDK 1.5.0 adds document creation, page images, OCR copies, barcode reading and columnar output to the Python package. If you build PDF pipelines, you can now create documents and page images, run OCR, and write extraction results as Parquet, Apache Arrow or JSONL.

Documents, page images and codes

The Python package now builds PDFs as well as reading them. images_to_pdf makes a PDF from images, create_blank_pdf makes one from blank pages, and Document.rasterize renders pages as images and returns a PDF with no text layer. Document.ocr writes an OCR copy, Document.read_barcodes reads barcodes and QR codes, and Document.read_mrz reads machine-readable zones on identity documents. The release notes list what the other bindings have.

Extraction results as data

pdfluent.columnar writes extraction results as JSONL, Apache Arrow and Parquet. You pass your own embeddings; the SDK does not compute them. pdfluent.storage reads and writes documents by URI. The default wheel covers local file, in-memory and HTTP. Cloud services need a wheel built with the storage-services feature. The package needs no network connection to be imported or to extract Markdown from a local file.

PDF to Markdown with a table

pip install pdfluent==1.5.0 installs the Python package. The example below runs on one synthetic page with a heading, a paragraph and a four-row table of revenue by region.

bash
pip install pdfluent==1.5.0
pdfluent models install layout tables
python
from pdfluent import Document

with Document("report.pdf") as doc:
    print(doc.extract_markdown())
markdown
# Quarterly report

Revenue grew in every region this quarter. The table below breaks it down by region.

## Revenue by region

| Region | Q1 | Q2 | Q3 |
| --- | --- | --- | --- |
| Europe | 412 | 455 | 498 |
| North America | 380 | 401 | 437 |
| Asia | 205 | 246 | 291 |
| Total | 997 | 1102 | 1226 |

Figures are in thousands of euros.

In the output, check that the table kept its rows and that the numbers are unchanged.

The scores

I ran the benchmark with its own harness and evaluator. The footnote under the numbers tells you how to read them.

On opendataloader-bench, SDK 1.5.0 in its default auto tier scores 0.9288 overall*, 0.9463 on reading order, 0.9140 on table structure and 0.8789 on heading hierarchy, on all 200 documents, with the benchmark's own harness and evaluator. SDK 1.3.0 scored 0.9154, 0.9359, 0.8695 and 0.8678 with the same harness and settings.

We measure the fast tier separately with the same harness.

On opendataloader-bench, SDK 1.5.0 in its fast tier scores 0.8870 overall*, 0.9147 on reading order, 0.7213 on table structure and 0.8601 on heading hierarchy, on all 200 documents, with the benchmark's own harness and evaluator.

*The scores are measured on the benchmark's own 200 documents. One run, no confidence interval. Both PDFluent rows keep page headers and footers (keep_furniture=True), because the benchmark's ground truth counts them as content. The table structure score depends on the platform; PDFluent 1.5.0 was measured on macOS arm64, both rows. Measured 17 November 2026 on a Mac, with the layout, table and formula models installed.

We also measured on our own held-out half of four public datasets.

On our own held-out half of four public datasets (19 extraction measures over FinTabNet, PubTables-1M, DocLayNet and ParseBench), SDK 1.5.0 averages 0.7304 with quality="auto", against 0.7244 for SDK 1.3.0 measured the same way; with quality="fast" it averages 0.6722 against 0.6533. Measured 17 November 2026, one run.

The table model was trained on the source datasets of PubTables-1M and FinTabNet, and most of the held-out pages of those two are in their training split, so the measures on them favour the table model.

The two comparison pages have side-by-side figures against Docling and Unstructured, with their footnotes. Docling, Unstructured. The full tables are on the benchmarks page.

Fixes and upgrade notes

The fixes and the upgrade notes are in the release notes and the changelog.

Licences

The SDK is AGPL-3.0 or the PDFluent Commercial Licence, your choice. The pdfluent-mcp MCP server is MIT licensed from version 0.3.1. It depends on the PDFluent SDK, which is AGPL-3.0 or commercial; see pdfluent.com/license. See licensing and SDK pricing.

Install

Install the binding you need with the commands below.

Calls enter the core and return.
bash
pip install pdfluent==1.5.0
cargo add pdfluent@1.5.0
npm install @pdfluent/node@1.5.0
npm install @pdfluent/sdk-wasm@1.5.0
dotnet add package PDFluent --version 1.5.0
com.pdfluent:pdfluent:1.5.0

Report problems or ask questions in GitHub Discussions.

Frequently asked questions

What does PDFluent SDK 1.5.0 score on opendataloader-bench?
On opendataloader-bench, SDK 1.5.0 in its default auto tier scores 0.9288 overall*, 0.9463 on reading order, 0.9140 on table structure and 0.8789 on heading hierarchy, on all 200 documents, with the benchmark's own harness and evaluator. SDK 1.3.0 scored 0.9154, 0.9359, 0.8695 and 0.8678 with the same harness and settings. *The scores are measured on the benchmark's own 200 documents. One run, no confidence interval. Both PDFluent rows keep page headers and footers (keep_furniture=True), because the benchmark's ground truth counts them as content. The table structure score depends on the platform; PDFluent 1.5.0 was measured on macOS arm64, both rows. Measured 17 November 2026 on a Mac, with the layout, table and formula models installed.
Is PDFluent SDK 1.5.0 free to use?
It is free under AGPL-3.0 if that licence works for your project. If you cannot use AGPL-3.0, you need the PDFluent Commercial Licence.
Which licence does the PDFluent MCP server use?
The pdfluent-mcp MCP server is MIT licensed from version 0.3.1, as of 18 November 2026. It depends on the PDFluent SDK, which is AGPL-3.0 or commercial; see pdfluent.com/license.
Does PDFluent SDK 1.5.0 need a network connection?
The package needs no network connection to be imported or to extract Markdown from a local file.
Which new functions read barcodes and identity documents?
Document.read_barcodes reads barcodes and QR codes, and Document.read_mrz reads machine-readable zones on identity documents, both in the Python package.
What should I check when upgrading from 1.4?
The fixes and the upgrade notes are in the release notes and the changelog.