Package Architecture

Package Architecture#

Docling ships two PyPI distributions from a single monorepo: docling-slim (the real implementation) and docling (a backwards-compatible meta-package). The split was introduced in PR #3285 (April 2026) to allow leaner installs without breaking existing pip install docling workflows.

The Two-Package Model#

Package	Role	Ships Python modules?
`docling-slim`	Implementation — all source code, all extras	Yes
`docling`	Meta-package — depends on `docling-slim[standard]`	No

docling-slim is defined at the repository root pyproject.toml. Its minimal base install pulls in ~50 MB across 8 packages: pydantic, docling-core, pydantic-settings, filetype, requests, certifi, pluggy, and tqdm .

docling is a pure dependency wheel defined at packages/docling/pyproject.toml. It ships no Python modules to prevent collisions on install , and simply pins docling-slim[standard] as its sole dependency .

The repository is a UV workspace with docling as the only explicit workspace member:

[tool.uv.workspace]
members = ["packages/docling"]

Note: the root pyproject.toml is the docling-slim package definition — there is no separate packages/docling-slim/pyproject.toml.

Backwards Compatibility#

The docling meta-package re-exports slim extras under their legacy names so existing installs keep working :

Legacy extra (`docling[…]`)	Maps to
`easyocr`	`docling-slim[feat-ocr-easyocr]`
`tesserocr`	`docling-slim[feat-ocr-tesserocr]`
`ocrmac`	`docling-slim[feat-ocr-mac]`
`vlm`	`docling-slim[models-vlm-inline]`
`rapidocr`	`docling-slim[feat-ocr-rapidocr-onnx]`
`asr`	`docling-slim[format-audio]`
`htmlrender`	`docling-slim[format-html-render]`
`remote-serving`	`docling-slim[models-remote]`
`onnxruntime`	`docling-slim[models-onnxruntime]`
`xbrl`	`docling-slim[format-xml-xbrl]`

CLI entry points (docling, docling-tools) are re-declared in both wheels so that uv tool install docling correctly links them into ~/.local/bin; at runtime, both launchers resolve their modules from the docling-slim dependency .

Format Extras#

All format extras are defined in docling-slim's pyproject.toml:

Extra	Key dependencies	Notes
`format-pdf`	pypdfium2, docling-parse	Bundle of `format-pdf-pypdfium2` + `format-pdf-docling`
`format-office`	python-docx, python-pptx, openpyxl	Bundle of docx + pptx + xlsx
`format-web`	beautifulsoup4, marko	Bundle of html + markdown
`format-opendocument`	odfdo	ODT, ODS, ODP
`format-latex`	pylatexenc
`format-email`	mail-parser (+ html)
`format-xml-jats`	lxml, beautifulsoup4	JATS/NLM scientific XML
`format-xml-xbrl`	arelle-release	Financial XBRL
`format-xml-uspto`	defusedxml, beautifulsoup4	USPTO patent XML
`format-html-render`	playwright	Headless browser rendering
`format-audio`	mlx-whisper (Apple Silicon), openai-whisper + whisper-s2t (others)	Platform-conditional

OCR Engine Extras#

OCR engines are fully optional and installed independently :

Extra	Engine	Platform / Python constraint
`feat-ocr-rapidocr`	RapidOCR (no inference backend)	Any
`feat-ocr-rapidocr-onnx`	RapidOCR + onnxruntime	Python < 3.14
`feat-ocr-easyocr`	EasyOCR + scikit-image	Any
`feat-ocr-tesserocr`	tesserocr + pandas	Any
`feat-ocr-mac`	ocrmac (Apple Vision)	macOS only
`feat-ocr-nemotron`	Nemotron OCR	Linux x86_64, Python == 3.12, CUDA 13.x

At runtime, OcrAutoModel selects an engine via a cascading try/except ImportError loop: macOS → Nemotron (Linux) → RapidOCR-ONNX → EasyOCR → RapidOCR-torch . If no engine is available, conversion proceeds without OCR rather than crashing.

Platform and Python Version Constraints#

Several extras carry python_version, sys_platform, or platform_machine markers that constrain installation :

feat-ocr-nemotron: python_version == "3.12" and sys_platform == "linux" and platform_machine == "x86_64"
feat-ocr-mac: sys_platform == "darwin"
models-onnxruntime: onnxruntime<1.24 on macOS; onnxruntime-gpu<1.24 on Linux/Windows; both gates require python_version < "3.14" — effectively not installed on Python 3.14+
feat-ocr-rapidocr-onnx: onnxruntime dependency guarded python_version < "3.14"
models-vlm-inline: mlx-vlm only on darwin arm64; transformers version ceiling differs on macOS vs other platforms
format-audio: mlx-whisper on darwin arm64; whisper-s2t-reborn on all others

Convenience Bundles#

Two aggregate extras reduce install complexity :

standard: format-pdf + models-local + feat-ocr-rapidocr + format-office + format-web + format-latex + format-email + feat-chunking + extract-core + service-client + cli
all: standard + VLM + audio + HTML render + all XML formats + remote models + ONNX + EasyOCR + tesserocr + macOS OCR

Installing pip install docling resolves to docling-slim[standard] — the same as pip install docling-slim[standard].

Key Files#

pyproject.toml — docling-slim definition: all extras, format support, OCR engines, platform markers
packages/docling/pyproject.toml — docling meta-package: backwards-compat re-exports and CLI entry points