docling.rs

docling.rs converts documents — PDF, Office, HTML, scans, audio — into Markdown, docling JSON and the other shapes an LLM or a RAG pipeline can read. It is a Rust port of IBM's docling: the same output, byte-for-byte, from one static binary. No Python, no PyTorch.

Pulling text out of a PDF is easy. Keeping its structure — headings in order, tables as tables, scanned pages read — takes layout analysis, table recognition and OCR. docling has the models for that, and a Python stack to match. docling.rs runs the same models from a single binary: PDF 4.3× faster than Python docling at under half the RAM, Office and HTML 20–60× faster, and tables that survive — cell recall 0.94 where MarkItDown gets 0.31 (numbers, method).

Coming from MarkItDown? pip install docling-rs, swap one import — from docling_rs.markitdown import MarkItDown — and keep its API, every format it converts included, with real layout analysis, tables and OCR behind it: guide, measured side by side.

Coming from Marker? pip install docling-rs, swap the imports — from docling_rs.marker.converters.pdf import PdfConverter — and keep its API: create_model_dict, text_from_rendered, the JSON block tree, chunks, marker_single's CLI and output layout, with docling's layout analysis, TableFormer and OCR behind it and no PyTorch or inference server: guide, measured side by side.

Coming from MinerU? pip install docling-rs, swap the imports — from docling_rs.mineru.parser import parse — and keep MinerU 4's SDK: ParseResult, markdown(), the Middle JSON, Content Lists and Structured Content, save()'s layout and mineru-kit parse, with docling's layout analysis, TableFormer and OCR behind it and no DocVortex or llama.cpp: guide, measured side by side.

Install

# CLI + models (Linux x64/arm64/s390x + macOS arm64 prebuilt; builds from source elsewhere)
curl -fsSL https://raw.githubusercontent.com/docling-project/docling.rs/master/scripts/install/install.sh | bash
docling-rs example.pdf          # Markdown on stdout

# or from crates.io; --cli puts the models where the installed binary looks
cargo install docling-cli
curl -fsSL https://raw.githubusercontent.com/docling-project/docling.rs/master/scripts/install/download_dependencies.sh | sh -s -- --cli

pip install docling-rs          # Python — drop-in for docling's, MarkItDown's, Marker's and MinerU's APIs
npm install docling.rs          # Node.js / Bun
cargo add docling               # Rust library

# containers, models baked in (-cuda variants for NVIDIA)
docker run --rm -v "$PWD:/data" ghcr.io/docling-project/docling-rs paper.pdf
docker run -p 127.0.0.1:5001:5001 ghcr.io/docling-project/docling-rs-serve

Office, HTML, Markdown and e-mail need no models. PDF, images and audio need the ONNX models (layout, OCR, TableFormer, Whisper; under 1 GB): install.sh, --cli and the containers bring them along; with pip, run python -c "import docling_rs; docling_rs.download_models()" once; with npm or a checkout, run download_dependencies.sh where you convert (details). Windows: docs/WINDOWS.md.

▶ Try it in your browser first — the converter as wasm, OCR included; nothing leaves your device.

Related MCP server: md-server

Use

docling-rs report.pdf > report.md               # Markdown (streamed page by page)
docling-rs report.pdf --to json                 # docling JSON; also html, text, latex, dclx, chunks, images
docling-rs --input ./docs --output ./out        # a directory tree; --jobs N in parallel
docling-rs https://example.com/paper.pdf        # URLs, data: URIs and stdin (`-`) too
docling-rs https://www.youtube.com/watch?v=ID   # a YouTube video: title, description, caption transcript (--to vtt keeps the timing)
docling-rs --help                               # every flag: OCR, page ranges, images, enrichment, redaction …

from docling_rs import DocumentConverter       # was: from docling.document_converter import …
doc = DocumentConverter().convert("report.pdf").document
print(doc.export_to_markdown())

import { convertFileAsync } from 'docling.rs'
const { content } = await convertFileAsync('report.pdf', { to: 'markdown' })

use docling::{DocumentConverter, SourceDocument};
let result = DocumentConverter::new().convert(SourceDocument::from_file("report.pdf")?)?;
println!("{}", result.document.export_to_markdown());

curl -F file=@report.pdf 'localhost:5001/v1/convert?to=md'   # the server; OpenAPI at /openapi.yaml

Same options, same outputs on every surface — every option in docs/OPTIONS.md.

Highlights

• From Python docling: from docling_rs import DocumentConverter; result.document is a genuine DoclingDocument (migration notes).

• Every docling format, and a few it lacks — the table under Status, each validated against live docling (Conformance).

• Every output docling has: Markdown (docling's, or a stricter dialect), docling JSON, HTML, plain text, DocLang, LaTeX, Pandoc AST (→ DOCX, EPUB, Typst… through pandoc), WebVTT subtitles, RAG chunks and page images (output formats).

• From any language: Rust, Python, Node.js / Bun, the browser (wasm, fully client-side), C, C++, C#, Go, Java and Swift through one C header, any HTTP client, the shell (Integration).

• A drop-in server for docling-serve: Open WebUI uses it as its Docling content-extraction engine — point Docling Server URL at http://<host>:5001 — and n8n, Dify and LangChain clients work unchanged (docling-rs serve, Open WebUI setup).

• AI agents (MCP): docling-rs mcp gives Claude Desktop, Claude Code, Cursor or VS Code MarkItDown's convert_to_markdown tool, plus convert for JSON, chunks and the rest — uvx docling-rs-mcp, or claude mcp add docling -- docling-rs mcp (MCP server).

• LangChain: pip install "docling-rs[langchain]" — langchain-docling's DoclingLoader on the Rust engine, one import to swap (details).

• AutoGen: pip install "docling-rs[autogen]" — DoclingIndexer fills any AutoGen memory with chunks that carry their pages, DoclingConvertTool lets an agent read documents (details).

• Microsoft Agent Framework: pip install "docling-rs[agent-framework]" — AutoGen's successor: DoclingIndexer fills any Agent Framework vector store, DoclingContextProvider puts the relevant chunks with their pages into each turn (details).

• Your own formats: register a backend for a format docling.rs does not read — or to read one of its formats your way — from Rust, Python (@docling_rs.register_backend(extensions=["rtf2"]), and MarkItDown's register_converter and plugins) or Node.js, ahead of the built-in formats (pluggable backends).

• RAG: docling-rag — chunk, embed and search with a REST API and a web UI (RAG subsystem).

• Runs where you are: Linux x64 / arm64 / s390x, macOS (Apple Silicon) and Windows prebuilt (CLI, wheels, npm), Intel macs from source, containers for amd64 and arm64; NVIDIA CUDA and TensorRT, DirectML (a prebuilt Windows CLI), CoreML and WebGPU in the browser as opt-in accelerators (GPU, IBM Z).

Developed with Claude Code and TENET (minimalistic AI-driven development framework).

Status

Every format in docling's pipeline converts — plus a few docling lacks. A file with no or an unknown extension, stdin, an upload or a data: URL is typed by its content (InputFormat::detect, docling_rs.detect_format, detectFormat: the packages, the OLE files, PDF, images, audio and video, iWork, XML, JSON, notebooks, messages, …), and a file that fails under its extension is checked by content and converted as what it really is, with a warning.

Every extension, with what each backend does: docs/MIGRATION.md.

Any ffmpeg ≥ 4.x on PATH works; docling.rs shells out to the binary and parses its output, so no dev headers/libraries are needed.

• Debian/Ubuntu: sudo apt-get install ffmpeg

• Fedora: sudo dnf install ffmpeg-free (or ffmpeg from RPM Fusion)

• Alpine: apk add ffmpeg

• macOS: brew install ffmpeg

• Windows: winget install ffmpeg (or choco install ffmpeg, or scoop install ffmpeg). Installing from a downloaded zip (gyan.dev / BtbN) also works — either add the extracted bin\ folder to PATH, or skip PATH entirely and point DOCLING_FFMPEG at the exe: set DOCLING_FFMPEG=C:\tools\ffmpeg\bin\ffmpeg.exe

Check with ffmpeg -version. DOCLING_FFMPEG overrides the binary used on any OS; the docling-rs-serve Docker image ships ffmpeg preinstalled.

A trimmed, decode-only build (--disable-everything --enable-…) works too (#853); docling.rs needs only: the file protocol; the demuxers of your containers (mov for mp4/mov/m4v, matroska for mkv/webm, avi, mpeg); the video decoders of your tracks (h264, hevc, vp8, vp9, av1, mpeg4, mpeg2video, …) with their parsers; the filters select, metadata, fps, scale, showinfo; the encoders png and wrapped_avframe; the muxers image2 and null. No audio component is used — the transcript is decoded in Rust — and the scene scan maps the video stream alone. Missing a component, ffmpeg fails the scene scan (Encoder not found, Invalid argument) and the video converts as transcript only; DOCLING_RS_DEBUG=1 shows the message.

Conformance with Microsoft MarkItDown

MarkItDown converts the same kinds of documents to Markdown for LLMs. Measured side by side — MarkItDown 0.1.9 (markitdown[all]) and docling.rs master after 2.51.3 (release build), 4 vCPU, no GPU, 2026-10-25 — on this repository's test corpus and on MarkItDown's own OCR test set (scripts/conformance/markitdown_compare.py):

Accuracy is scored against Python docling's groundtruth, which favours docling.rs by construction; the timings do not depend on it. Text drawn into images (MarkItDown's own OCR test set) reads at 1.00 with docling.rs' local OCR, against 0.00–0.33 for MarkItDown without an LLM. Method, caveats and the full OCR comparison: docs/MARKITDOWN.md.

Conformance with Marker

Marker (datalab's marker-pdf) converts PDFs and, through a PDF render, Office files to Markdown for LLMs with a layout model and OCR. Measured side by side — Marker 2.0.0 (marker-pdf[full], its CPU default fast mode, llama.cpp for its OCR) and docling.rs master after 2.51.3 (release build), 4 vCPU, no GPU, 2026-10-25 ³ — on this repository's test corpus (scripts/conformance/marker_compare.py):

A failed file scores 0. ¹ Only one XLSX groundtruth carries a heading. ² The median HTML file needs OCR in Marker: a cold call starts its llama.cpp server and loads the 1.2 GB OCR model first. ³ The HTML row is the earlier run's (docling.rs 2.46.2, another host): its re-run here was stopped after 3 h 45 min — its warm pass done, the first cold marker_single call had spent 97 minutes on one 855-byte file, Marker's OCR model looping on one render (5,000+ tokens for a single block).

Accuracy is scored against Python docling's groundtruth, which favours docling.rs by construction; the timings and memory do not depend on it. Marker reads Office files, HTML and EPUB by rendering them to PDF and running its vision models on the render; on CPU every page its text check flags goes through a 1.2 GB vision-language OCR model in a llama.cpp server that pip install marker-pdf does not bring (without it 4 PDFs, 9 DOCX, 10 HTML and the EPUB fail outright). Peak RSS counts Marker's client and its inference servers. Marker's balanced mode (its published benchmarks) needs a GPU. Method, caveats and the API map: docs/MARKER.md.

Conformance with MinerU

MinerU (OpenDataLab) converts PDFs, images and Office files to Markdown and JSON for LLMs in four tiers — flash (text layer, no models), basic (ONNX layout / table / OCR models), standard and advanced (a vision-language model through llama.cpp). Measured side by side — MinerU 4.0.12 and docling.rs master after 2.51.3 (release build), 4 vCPU, no GPU, 2026-10-25 — on this repository's test corpus (scripts/conformance/mineru_compare.py):

A failed file scores 0. ¹ Only one XLSX groundtruth carries a heading.

Accuracy is scored against Python docling's groundtruth, which favours docling.rs by construction; the timings and memory do not depend on it. MinerU parses Office files, HTML and EPUB at flash only; basic is its model tier on CPU; its standard tier (a 1.2 B vision-language model through llama.cpp) took 138 s for one page on a 4 vCPU CPU in the earlier run and is not in the table. Method, caveats and the API map: docs/MINERU.md.

Conformance with Python docling

Every declarative format is compared against the live library — Markdown byte for byte, JSON structurally (every key and value but origin/version and re-encoded image bytes). Latest sweep, docling 2.142.1 with the docling-core 2.103.1 it locks, over its mirrored test corpus (scripts/conformance/full_conformance.py; ✅ all = every file matches):

¹ Deliberate: in 1706.03762 a \multirow / \multicolumn cell's row_span / col_span agree with its offsets, where upstream leaves them at 1 (docling#4817's LaTeX rewrite is ported in full, PDF figures rendered at 144 dpi like upstream's).

² xlsx_02_sample_sales_data: 20 date cells print by docling PR #4628's number-format rules (#634), which the released docling does not apply yet.

³ Deliberate: a table inside a list item stays a table after the item (upstream prints its grid right under the item line); two Wikipedia links keep docling-core 2.104's %28/%29 escaping.

⁴ Whitespace only: a one-run footnote prints on its bullet, where docling 2.142 leaves the bullet empty and puts the text on the next line.

⁵ Deliberate: a code span keeps its literal text (&), which upstream decodes when a raw HTML block sends the file through its HTML backend.

⁶ Ahead of the release: pptx_chart's 3-D column chart is read as a bar chart with its data, as docling main reads it since docling#4811 (after 2.142.1).

The residuals, with the exact files: docs/MIGRATION.md. The PDF/image pipeline is pinned by a deterministic snapshot baseline and scored against docling's groundtruth in docs/PDF_CONFORMANCE.md.

Performance

One fixture per format through published Python docling and the Rust release binary (scripts/test/performance.sh); docling ÷ docling.rs, so bigger means Rust wins by more. Measured 2026-10-25: docling 2.142.1 (the default PyPI install, torch included) against docling.rs master after 2.51.3, 4 vCPU, no GPU.

† A 4 KB CSV converts in microseconds, so Rust's own process startup dominates its number; end to end it is 780× faster.

Declarative formats hold 17–28 MB against docling's 750–780 MB (it imports torch even for them) and finish 200–890× sooner end to end; the PDF peaks at 0.46 GB against 2.2 GB and converts 19× faster end to end. Where the time goes, and what was measured not to help: docs/PDF_CONFORMANCE.md.

Integration — using docling.rs from your language

One engine, several front doors, the same options and outputs everywhere.

Rust when you're already in Cargo; the CLI for scripts and batch jobs; serve when several services share one warm pipeline; Node / Python in-process; wasm when the document must not leave the user's machine; FFI for anything that can call C.

Rust API

use docling::{DocumentConverter, SourceDocument};

let converter = DocumentConverter::new();
let result = converter.convert(SourceDocument::from_file("input.md")?)?;

println!("{}", result.document.export_to_markdown()); // Markdown
println!("{}", result.document.export_to_json());     // docling DoclingDocument JSON

Every conversion knob is one serializable struct, docling::ConvertOptions, filled the same way by every surface (docs/OPTIONS.md). The library guide covers the output formats, error handling, chunking, image extraction, streaming, OCR, the PDF switches, the remote VLM pipeline and registering your own backends (#680).

HTTP conversion API — docling-rs serve

docling-serve exposes the converter over HTTP with one warm PDF/image pipeline shared across requests (repeat PDFs skip the model load), streamed Markdown, batches, ZIP inputs, async jobs, an interactive test form at GET / and an OpenAPI 3.1 description at GET /openapi.yaml. It also serves Python docling-serve's routes, so Open WebUI, n8n, Dify and LangChain clients work unchanged.

docker run -p 127.0.0.1:5001:5001 ghcr.io/docling-project/docling-rs-serve
curl -F file=@paper.pdf localhost:5001/v1/convert                # Markdown
curl -F file=@report.docx 'localhost:5001/v1/convert?to=json'    # docling JSON
curl -F file=@a.pdf -F file=@b.docx localhost:5001/v1/convert    # a batch

Request options, server flags, async jobs, the docling-serve routes and the Open WebUI setup: crates/docling-serve/README.md; containers and sizing: docs/DEPLOYMENT.md.

MCP server — docling-rs mcp

docling-rs mcp is a Model Context Protocol server in the default docling-rs binary and image. Agents get MarkItDown's convert_to_markdown(uri) tool — same name and contract, so a markitdown-mcp setup switches by changing the command — plus convert(uri, to, options) and list_input_formats().

{
  "mcpServers": {
    "docling": { "command": "docling-rs", "args": ["mcp", "--root", "/Users/me/Documents"] }
  }
}

claude mcp add docling -- docling-rs mcp --root ~/Documents     # Claude Code
uvx docling-rs-mcp --root ~/Documents                           # the same server from PyPI

Local reads stay under --root; URLs need --allow-url-fetch. Listed in the official MCP registry and on Smithery. The PyPI launcher, Docker, Streamable HTTP (--http), the security model and memory behaviour: crates/docling-cli/README.md.

RAG subsystem

crates/docling-rag turns documents into chunks, embeds them (Ollama, Gemini or a local ONNX model), stores them (SQLite + sqlite-vec or PostgreSQL + pgvector) and searches them — hybrid, multi-query and HyDE retrieval, an evaluation harness, a REST API and a built-in search UI.

In the browser — docling-wasm

The declarative converters and the PDF text layer compile to WebAssembly: npm i docling.rs-wasm, then convert(bytes, filename, to) runs fully client-side — no server, a ~3.4 MB gzipped module, optional in-browser OCR. Details and the demo page: crates/docling-wasm.

Node.js, Python and C

• Node.js / Bun — docling.rs on npm, a prebuilt N-API addon: sync and async conversion, streaming, chunking, ZIP batches, a warm Pipeline.

• Python — docling-rs on PyPI (docling-rs-cuda for NVIDIA): docling's DocumentConverter with a genuine docling-core DoclingDocument, MarkItDown's API, and the LangChain, AutoGen and Microsoft Agent Framework integrations.

• C ABI — docling-ffi: one library and one docling.h for C, C++, C#/.NET, Go, Java, Swift and Zig, prebuilt in every published release.

Deploy in a container

Compose files (standalone, Caddy TLS, CUDA), environment variables, sizing and observability: docs/DEPLOYMENT.md.

Getting the ML models

The PDF/image pipeline needs ONNX models — RT-DETR layout, PP-OCR, and TableFormer — and audio needs Whisper. download_dependencies.sh fetches them from this repository's releases into ./.models, where every surface looks (--cli: beside a cargo-installed binary instead):

scripts/install/download_dependencies.sh                       # from a checkout
curl -fsSL https://raw.githubusercontent.com/docling-project/docling.rs/master/scripts/install/download_dependencies.sh | sh         # anywhere

Every asset and option (ASR presets, enrichment models, NER for PII redaction, INT8, fonts), GPU execution providers and the environment overrides: docs/MODELS.md. IBM Z and other targets without a prebuilt ONNX Runtime: docs/IBMZ.md.

Contributing

Bug reports and pull requests are welcome — see CONTRIBUTING.md for the build and test commands, the conformance workflow, and the conventions a change is expected to follow.

License

MIT, matching upstream docling.