Skip to content

WINDOWS.md

Latest commit

 

History

History
74 lines (56 loc) · 2.55 KB

WINDOWS.md

File metadata and controls

74 lines (56 loc) · 2.55 KB
Raw

docling.rs on Windows (native)

The pipeline runs natively on Windows — no WSL required, and usually faster than under it: ONNX Runtime's prebuilt Windows binaries (pulled automatically by the ort crate at build time) use the same AVX2/VNNI kernels, without the VM/filesystem overhead.

Prerequisites

  • Rust (MSVC toolchain) — rustup default stable-x86_64-pc-windows-msvc plus the Visual Studio Build Tools with the "Desktop development with C++" workload (the linker).
  • Windows 10 1803+ — curl.exe and tar.exe ship with the OS; the download script needs nothing else.

Build & fetch the ML dependencies

git clone https://github.com/docling-project/docling.rs
cd docling.rs

scripts\install\download_dependencies.bat   :: models\ + .pdfium\lib\pdfium.dll
cargo build --release -p docling-cli

The script mirrors download_dependencies.sh: layout / OCR / TableFormer / picture-classifier ONNX models (INT8 variants included — preferred automatically; set DOCLING_RS_FP32=1 to opt out), Whisper-tiny for audio (--no-asr to skip), the chunker tokenizer, and pdfium.dll (fetched from pdfium-binaries, which the Linux-only asset in our models release doesn't cover).

Run

Everything resolves relative to the current directory (or next to the executable), so from the repo root no environment variables are needed:

target\release\docling-rs paper.pdf
target\release\docling-rs --to json report.docx

From another directory, point the two roots at the checkout — cmd:

set PDFIUM_DYNAMIC_LIB_PATH=C:\src\docling.rs\.pdfium\lib
set DOCLING_LAYOUT_ONNX=C:\src\docling.rs\models\layout_heron_int8.onnx

or PowerShell:

$env:PDFIUM_DYNAMIC_LIB_PATH = "C:\src\docling.rs\.pdfium\lib"

Every DOCLING_* / DOCLING_RS_* variable from the main README works the same way. The HTTP server is the same crate as on Linux:

cargo run --release -p docling-serve            :: 127.0.0.1:5001, docs at /

Notes & limits

  • Model export / quantization / conformance scripts are bash — run those under WSL or Git Bash; the prebuilt models from the release make this unnecessary for normal use.
  • Audio decoding is pure Rust (symphonia), no ffmpeg needed on Windows either.
  • The optional --features web-browser pre-render needs a Chromium install and is untested on Windows.
  • Benchmark tip: target\release\docling-rs --bench-warm 3 paper.pdf prints warm seconds/conversion — handy for comparing against a WSL build of the same checkout.