Docling inference server

This project provides a FastAPI wrapper around the docling document parser to make it easier to use in distributed production environments.

Running

The easiest way to run this project is using docker. There are two image families, one for cuda machines and one for cpu:

• Cuda: ghcr.io/aidotse/docling-inference:rev
• CPU: ghcr.io/aidotse/docling-inference:cpu-rev

# Create volumes to not have to download models every time
docker volume create hf_cache
docker volume create ocr_cache

# Run the container
docker run -d \
  --gpus all \
  -p 8080:8080 \
  -e NUM_WORKERS=8 \
  -v hf_cache:/root/.cache/huggingface \
  -v ocr_cache:/root/.EasyOCR \
  ghcr.io/aidotse/docling-inference:latest

Docker compose

services:
  docling-inference:
    image: ghcr.io/aidotse/docling-inference:latest
    ports:
      - 8080:8080
    environment:
      - NUM_WORKERS=8
    volumes:
      - hf_cache:/root/.cache/huggingface
      - ocr_cache:/root/.EasyOCR
    restart: always
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

volumes:
  hf_cache:
  ocr_cache:

Local python

Dependencies are handled with uv in this project. Follow their installation instructions if you do not have it.

# Create a virtual environment
uv venv

# Install the dependencies
uv sync --extra cpu
# OR if you have cuda devices
uv sync --extra cu121

# Activate the shell
source .venv/bin/activate

# Start the server
python src/main.py

Using the API

When the server is started you can find the interactive API documentation at the /docs endpoint. If you're running locally with the example command, this will be http://localhost:8080/docs.

There are two main methods to parse documents that take the data on two different formats. You can use the /parse/url to parse a document from a download link. To call it with curl from the command line:

curl -X 'POST' \
  'http://localhost:8080/parse/url' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "include_json": false,
  "output_format": "markdown",
  "url": "https://arxiv.org/pdf/2408.09869"
}'

You can also parse files directly with the /parse/file endpoint:

curl -X 'POST' \
  'http://localhost:8080/parse/file' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'file=@file-path.pdf;type=application/pdf' \
  -F 'data={"include_json":false,"output_format":"markdown"}'

Tip: You can use a service like https://curlconverter.com/ to convert curl commands to your favourite http client, e.g. requests.

For a full list of available options, please refer to the interactive documentation.

Building

Build the project docker image with one of the following commands

• Cuda: docker build -t ghcr.io/aidotse/docling-inference:dev .
• CPU: docker build -f Dockerfile.cpu -t ghcr.io/aidotse/docling-inference:dev .

Configuration

Configuration is handled through environment variables. Here is a list of the available configuration variables. They are defined in src/config.py

• NUM_WORKERS: The number of processes to run.
• LOG_LEVEL: The lowest level of logs to display. One of DEBUG, INFO, WARNING, CRITICAL, ERROR.
• DEV_MODE: Sets automatic reload of the service. Useful during development
• PORT: The port to run the server on.
• AUTH_TOKEN: Token to use for authentication. Token is expected in the Authorization: Bearer: <token> format in the request header. The service is unprotected if this option is omitted.
• OCR_LANGUAGES: List of language codes to use for optical character optimization. Default is "es,en,fr,de,sv". See https://www.jaided.ai/easyocr/ for the list of all available languages.
• DO_CODE_ENRICHMENT: Use a code enrichment model in the pipeline. Processes images of code to code.
• DO_FORMULA_ENRICHMENT: Use a formula enrichment model in the pipeline. Converts formulas to LaTeX.
• DO_PICTURE_CLASSIFICATION: Use a picture classification model in the pipelinese. Classifies the type of image into a category.
• DO_PICTURE_DESCRIPTION: Use a picture description model in the pipeline. Uses a small multimodal model to describe images.