Docling Rust SDK (docling.rs / docling-rs)#

Overview#

docling.rs is the official full Rust port of the Python docling library, hosted under docling-project on GitHub 1. It converts documents (PDF, DOCX, HTML, XLSX, PPTX, EPUB, and many more) into DoclingDocument — the same unified format used by the Python pipeline — for downstream AI workflows. The project is experimentally available as a Python package (docling-rs) on PyPI 2, and also ships Node.js/Bun bindings (docling.rs on npm) and a WebAssembly build for in-browser conversion.

The repository originated from a community-contributed Rust rewrite and was officially adopted by the docling-project GitHub organization 3.

Format Support#

The format migration is described as complete in the README 1. Supported input/output formats include:

• Declarative: Markdown, CSV, HTML, AsciiDoc, DOCX, PPTX, XLSX, legacy DOC/XLS/PPT, EPUB, ODF, WebVTT, Email, MHTML, JATS, USPTO, XBRL, LaTeX, JSON, DocTags (.doctags/.dt), DocLang (.dclg/.dclx)
• ML pipeline: PDF, images, METS, audio (via Whisper ASR), video
• MHTML is a docling.rs-only extension — Python docling has no MHTML backend 1. Requested in issue #659, docling-rs was cited as the solution 4.

Performance vs. Python#

Benchmarks from scripts/test/performance.sh (Python docling ÷ docling.rs — higher = Rust wins by more) 1:

Community benchmarks on a DOCX file showed 736× faster end-to-end (whole process) and 54× less peak memory; for a multi-page PDF, 4.3× faster end-to-end and 4.8× less memory 3.

For chunking: a large 14 MB HTML file that took ~2 hours with Python HybridChunker (O(N²) tokenization bottleneck) was chunked in ~2 min 49 s with docling-rs 5.

INT8 quantized models (fetched by default) give an additional ~2.4× layout-inference speedup on CPUs with AVX-512 VNNI, ~1.4–1.8× end-to-end 1.

GPU (docling-rs-cuda): 1.5–2.1× end-to-end on multi-page digital PDFs; 8.7× on a 1913-page manual 2.

Python Bindings (docling-rs on PyPI)#

Install (CPU):

pip install docling-rs

GPU (Linux x86-64, CUDA 12 + cuDNN 9):

pip install docling-rs-cuda

• Latest release: 0.52.4 (August 1, 2026) 2
• Wheels: abi3 (Python ≥ 3.9), Linux x86-64/arm64 (manylinux_2_28), Windows x86-64. macOS users install from the sdist (source build).
• Only the document processor is Rust. result.document is a genuine docling_core.DoclingDocument — all of docling's Python serializers (export_to_markdown(), export_to_dict(), export_to_doctags()) and chunkers work unchanged 2.
• Models (~700 MB) are fetched on first use via docling_rs.download_models() to ~/.cache/docling.rs 2.

Migration from Python docling#

The migration guide in crates/docling-py/README.md describes a three-step process 2:

1. Swap the package: pip install docling-rs (coexists with docling)
2. Rewrite imports: docling. → docling_rs. (mechanical rename)
3. Fetch models: python -c "import docling_rs; docling_rs.download_models()"

Known divergences: full-VLM pipeline (SmolDocling), per-format backend selection, and inline formatting rendered as text rather than structured formatting fields. HybridChunker(tokenizer=...) takes a path to a tokenizer.json (no transformers install required).

Crate Layout#

1

Key References#

• Repository: github.com/docling-project/docling.rs 1
• PyPI: pypi.org/project/docling-rs 2
• PyPI (GPU): pypi.org/project/docling-rs-cuda
• npm: npmjs.com/package/docling.rs
• Browser demo: docling-project.github.io/docling.rs 1
• Python migration guide: crates/docling-py/README.md
• PDF conformance / perf details: docs/PDF_CONFORMANCE.md
• Architecture & Python→Rust mapping: docs/MIGRATION.md