Docling-Serve

Docling-Serve

Docker app from xxBeanSproutxx's Repository

What is Docling? Docling is an open-source toolkit (from IBM Research) that converts documents (PDF, DOCX, images, HTML, etc.) into structured Markdown or JSON.

Requirements


**CPU-only deployments**
- Select the `cpu` branch (`quay.io/docling-project/docling-serve-cpu`) for CPU-only operation.
- `DOCLING_DEVICE=cpu` is a runtime hint; it does NOT replace choosing the CPU image branch.

**GPU deployments (NVIDIA, optional)**
- Install the NVIDIA Driver plugin and reboot.
- Add `--gpus all` in Extra Parameters (keep `--user 0` too).
- Branch `cu126` exists upstream but is no longer updated; prefer `cu128`.
- Optional: set `DOCLING_DEVICE=cuda` (or `cuda:0`) to force GPU execution.
Overview
What is Docling?

Docling is an open-source toolkit (from IBM Research) that converts documents (PDF, DOCX, images, HTML, etc.) into structured Markdown or JSON. It's great for RAG and local document processing.

Highlights

  • Multi-format parsing with layout understanding and table extraction.
  • Simple API + optional Web UI.
  • Runs locally on your Unraid box; keep your data private.

Default Endpoints

  • API: http://[IP]:[PORT:5001]
  • Docs: http://[IP]:[PORT:5001]/docs
  • Web UI: http://[IP]:[PORT:5001]/ui (set DOCLING_SERVE_ENABLE_UI=1)

Permissions

  • The template runs the container with --user 0 so first-run model downloads can write to appdata on any Unraid ownership. If you removed it and see PermissionError/Errno 13, either add --user 0 back in Extra Parameters or run chown -R 1001:1001 /mnt/user/appdata/docling from the Unraid host.

First-Run Model Download

  • On a fresh install the models directory will be empty. Docling must download RapidOCR and other artifacts on first boot.
  • First run downloads models from upstream model hosts (e.g., HuggingFace); needs outbound HTTPS.
  • After the first successful start and model cache is populated, you may set DOCLING_SERVE_ENABLE_REMOTE_SERVICES back to false if you prefer a fully local-only deployment.
  • Keep DOCLING_SERVE_LOAD_MODELS_AT_BOOT=true so any download failures show up immediately in startup logs rather than at first OCR request.

Image Size

  • The image is large (up to ~8 GB) — make sure the docker image share has free space before installing.

Persistent Paths

  • Models/artifacts are persisted in appdata so restarts do not re-download everything.
  • If logs show artifacts_path is set to an invalid directory, verify the models path exists and matches DOCLING_SERVE_ARTIFACTS_PATH.

Docling

Docling Serve

Running Docling as an API service.

📚 Docling Serve documentation

[!NOTE] Migration to the v1 API. Docling Serve now has a stable v1 API. Read more on the migration to v1.

Getting started

Install the docling-serve package and run the server.

# Using the python package
pip install docling-serve
docling-serve run --enable-ui

# Using container images, e.g. with Podman
podman run -p 5001:5001 -e DOCLING_SERVE_ENABLE_UI=1 quay.io/docling-project/docling-serve

The server is available at

API documentation

Try it out with a simple conversion:

curl -X 'POST' \
  'http://localhost:5001/v1/convert/source' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "sources": [{"kind": "http", "url": "https://arxiv.org/pdf/2501.17887"}]
  }'

Container Images

The following container images are available for running Docling Serve with different hardware and PyTorch configurations:

📦 Distributed Images

Image Description Architectures Size
ghcr.io/docling-project/docling-serve
quay.io/docling-project/docling-serve
Base image with all packages installed from the official PyPI index. linux/amd64, linux/arm64 4.4 GB (arm64)
8.7 GB (amd64)
ghcr.io/docling-project/docling-serve-cpu
quay.io/docling-project/docling-serve-cpu
CPU-only variant, using torch from the PyTorch CPU index. linux/amd64, linux/arm64 4.4 GB
ghcr.io/docling-project/docling-serve-cu128
quay.io/docling-project/docling-serve-cu128
CUDA 12.8 build with torch from the cu128 index. linux/amd64 11.4 GB
ghcr.io/docling-project/docling-serve-cu130
quay.io/docling-project/docling-serve-cu130
CUDA 13.0 build with torch from the cu130 index. linux/amd64, linux/arm64 TBD

[!IMPORTANT] CUDA Image Tagging Policy

CUDA-specific images (-cu128, -cu130) follow PyTorch's CUDA version support lifecycle and are tagged differently from base images:

  • Base images (docling-serve, docling-serve-cpu): Tagged with latest and main for convenience
  • CUDA images (docling-serve-cu*): Only tagged with explicit versions (e.g., 1.12.0) and main

Why? CUDA versions are deprecated over time as PyTorch adds support for newer CUDA releases. To avoid accidentally pulling deprecated CUDA versions, CUDA images intentionally exclude the latest tag. Always use explicit version tags like:

# ✅ Recommended: Explicit version
docker pull quay.io/docling-project/docling-serve-cu130:v1.18.0

# ❌ Not available for CUDA images
docker pull quay.io/docling-project/docling-serve-cu130:latest

🚫 Not Distributed

An image for AMD ROCm 6.3 (docling-serve-rocm) is supported but not published due to its large size.

To build it locally:

git clone --branch main https://github.com/docling-project/docling-serve.git
cd docling-serve/
make docling-serve-rocm-image

For deployment using Docker Compose, see docs/deployment.md.

Coming soon: docling-serve-slim images will reduce the size by skipping the model weights download.

Demonstration UI

An easy to use UI is available at the /ui endpoint.

Input controllers in the UI

Output visualization in the UI

Get help and support

Please feel free to connect with us using the discussion section.

Contributing

Please read Contributing to Docling Serve for details.

References

If you use Docling in your projects, please consider citing the following:

@techreport{Docling,
  author = {Docling Contributors},
  month = {1},
  title = {Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion},
  url = {https://arxiv.org/abs/2501.17887},
  eprint = {2501.17887},
  doi = {10.48550/arXiv.2501.17887},
  version = {2.0.0},
  year = {2025}
}

License

The Docling Serve codebase is under MIT license.

IBM ❤️ Open Source AI

Docling has been brought to you by IBM.

Categories

Related apps

Explore more like this

Explore all

Links

Details

Repository
quay.io/docling-project/docling-serve
Last Updated2026-11-22
First Seen2025-10-06

Runtime arguments

Web UI
http://[IP]:[PORT:5001]/ui
Network
bridge
Shell
sh
Privileged
false
Extra Params
--user 0

Template configuration

WebUI PortPorttcp

Docling Serve API/UI port

Target
5001
Default
5001
Docling Models PathPathrw

Persistent Docling model artifacts (required; must exist and be writable; must match DOCLING_SERVE_ARTIFACTS_PATH)

Target
/opt/app-root/src/.cache/docling/models
Default
/mnt/user/appdata/docling/models
HuggingFace CachePathrw

Persistent HuggingFace cache

Target
/opt/app-root/src/.cache/huggingface
Default
/mnt/user/appdata/docling/huggingface_cache
EasyOCR CachePathrw

Caches OCR models

Target
/opt/app-root/.EasyOCR
Default
/mnt/user/appdata/docling/easyocr_cache
Enable UIVariable

Enable the /ui playground (1/0)

Target
DOCLING_SERVE_ENABLE_UI
Default
1
Load Models At BootVariable

Preload/download models at startup (recommended for easier first-run diagnostics).

Target
DOCLING_SERVE_LOAD_MODELS_AT_BOOT
Default
true
Enable Remote ServicesVariable

Allow remote model/service calls. Enabled by default so first-run model downloads work automatically. Set to false after initial setup if you prefer fully offline operation.

Target
DOCLING_SERVE_ENABLE_REMOTE_SERVICES
Default
true
Artifacts Path (env)Variable

Directory used by Docling to load/store model artifacts

Target
DOCLING_SERVE_ARTIFACTS_PATH
Default
/opt/app-root/src/.cache/docling/models
Force DeviceVariable

Runtime device: auto, cpu, cuda, cuda:0, mps. Use 'auto' to let Docling decide. For CPU-only, select the cpu image branch above.

Target
DOCLING_DEVICE
Default
auto
NVIDIA_VISIBLE_DEVICESVariable
Default
all
NVIDIA_DRIVER_CAPABILITIESVariable
Default
compute,utility