Docling
chirashi.docling
A Docker Compose file for Docling alone, generated from the catalog and checked by the safety inspector. No account needed.
Document conversion toolkit parsing PDF, Office, HTML, image and audio inputs into a structured representation, recovering layout, reading order and table structure for retrieval pipelines.
Docling processes documents through modular pipelines that analyse page geometry, detect regions, reconstruct multi-column reading order and infer table structure, producing a typed document object exportable to Markdown, HTML or JSON. Layout analysis and a dedicated table model do most of the work visually, so OCR is applied only where a page genuinely lacks a text layer, which is both faster and less error-prone than treating every page as an image. Pluggable OCR engines cover the scanned cases, and integrations exist for the common retrieval frameworks. Conversion runs locally, so documents are not transmitted anywhere during processing.
You know it worked when
- A text-layer PDF converts to Markdown with its headings preserved.
- A table in that document appears as a structured table rather than run-together text.
- Multi-column pages are returned in correct reading order.
- A scanned page produces text once an OCR engine is configured.
- The JSON export contains element positions and provenance.
Known sharp edges
- Layout and table models download on first run, so the first conversion is far slower than later ones and requires outbound access unless the models are baked into the image.
- Conversion is CPU-intensive and scales with page count, so a large batch occupies the container for a long time with little visible progress.
- OCR engines are optional extras, and a scanned document silently yields almost no text when none is installed.
- Model weights are licensed separately from the MIT codebase, which matters if converted output feeds a commercial product.