DCC Docling Serve

A docling-serve image with the DCC docling plugins and a few patches for speed and operations.

GitHub Repository: DCC-BS/dcc-docling-serve

Features

• Bundled Plugins: docling-pp-doc-layout, docling-glm-ocr and docling-pp-ocrv6, selectable per request. Plugin versions and model weights are pinned and baked into the image, so pods do not download anything at startup.
• Web UI: Upstream's UI at /ui, unchanged. The plugins show up among its OCR and layout choices.
• Word Bounding Boxes: include_word_boxes=true adds a word_boxes field to the answer: per page, every word and text line with its box and whether it came from OCR.
• RapidOCR on the GPU: CUDA images ship onnxruntime-gpu, so the default OCR engine runs on the GPU instead of the CPU.
• Faster Born-Digital PDFs: Vector shapes (table rules, underlines, background fills) no longer send regions with PDF text to OCR. On our test set (976 pages) conversion time drops from 522 s to 165 s with the same output. Disable with DCC_OCR_IGNORE_SHAPES=0.
• Quiet Health Probes: GET /health requests are not logged. Disable with DCC_MUTE_HEALTH_LOGS=0.
• Multi-Environment Support: Images for CPU (amd64, arm64), CUDA 12.8 and CUDA 13.0.
• Full Stack: Docker Compose setup with the Docling API and a vLLM server for GLM-OCR.

Installation

1. Prerequisites: Docker with NVIDIA GPU support and a HuggingFace token.
2. Configuration: Copy .env.example to .env and set your HF_TOKEN.
3. Start: Run make docker-up to launch the services.

Usage

• API: Send a POST request to http://localhost:5001/v1/convert/source. Select the plugins with "ocr_engine": "glm-ocr-remote" or "pp-ocrv6", and "layout_custom_config": { "kind": "ppdoclayout-v3" }.
• Word boxes: Add "include_word_boxes": true (needs a dlparse PDF backend, the default). With PP-OCRv6 this also reads each page whole and returns a box per OCR'd word.
• UI: Open http://localhost:5001/ui.

Environment Variables

Variable	Description	Default
DCC_WORD_BOXES	0 leaves the API as upstream has it (no include_word_boxes)	1
DCC_OCR_IGNORE_SHAPES	0 restores docling's OCR rule for vector shapes	1
DCC_MUTE_HEALTH_LOGS	0 logs health probes again	1
DCC_MUTE_HEALTH_PATHS	Comma-separated request paths muted in the access log	/health

The plugins read their own variables, listed on their pages.

Releases

Images are published to ghcr.io/dcc-bs/dcc-docling-serve[-cpu|-cu128|-cu130] by a manually dispatched workflow that takes a docling-serve release tag. Our image gets the same tag: dcc-docling-serve:v1.36.0 is docling-serve v1.36.0 with our plugins and patches.