DocTags is a structured document markup language specifically designed by the development team and built using the JSON-LD semantic framework. The format converts visual recognition results into machine-readable hierarchical data: textual content retains the original location coordinate information, relationships between document elements are represented by topological mapping, and specialized content such as formulas and codes use standardized content tags. This design maintains human readability while supporting automated processing to achieve 97% information fidelity.
In practice, DocTags can be converted to 12 common formats such as Markdown, HTML or LaTeX in one click through the supporting docling_core library. Test data shows that the conversion efficiency from DocTags to Markdown reaches 2000 markup items per second without losing any structural information. The format also supports version tracking and incremental updates, which is especially suitable for collaborative document editing scenarios.
This answer comes from the articleSmolDocling: a visual language model for efficient document processing in a small volumeThe
