Docling as a Service: FastAPI, Celery, and a Queue for PDF-to-MD

What it does

docling-api is a production-oriented shim around IBM’s Docling library. It exposes Docling’s PDF, Word, PowerPoint, and image parsing through a FastAPI interface, adding a Redis-backed Celery task queue and a Flower dashboard so conversions don’t block the main thread. You can throw single files or batches at it and get Markdown back, with optional GPU acceleration for faster processing.

The interesting bit

Rather than just exposing a library, it bundles the whole operational stack—distributed workers, job tracking, and Docker Compose profiles for CPU and GPU—so you can scale horizontally without writing your own plumbing. The README even includes a side-by-side quality comparison against Marker, PyPDF, and PyMuPDF4LLM, suggesting the author cares about output fidelity as much as throughput.

Key highlights

• Supports PDF, DOCX, PPTX, HTML, images, CSV, AsciiDoc, and Markdown inputs.
• Offers synchronous, asynchronous, and batch conversion endpoints with job status tracking.
• Uses Celery and Redis for distributed task processing; includes a Flower monitoring dashboard.
• Provides Docker Compose setups for both CPU-only and GPU-accelerated deployments.
• Exposes configuration for image resolution scaling and table-to-image extraction.

Verdict

Reach for this if you need a self-hosted, queue-backed document conversion API rather than a command-line tool. If you only convert documents occasionally, run Docling directly and skip the infrastructure.