#atom

A structured format for representing extracted document elements with positional information

Core Idea: DocTags is a markup format used by document understanding models like SmolDocling to represent both the content and the structural information of processed documents, including element types and spatial positioning.

Key Elements

• Format Structure:

  • HTML-like hierarchical structure
  • Element type identification
  • Positional coordinates for elements
  • Nested representation of document hierarchy
  • Content extraction within tags

• Element Types:

  • Text blocks
  • Headings
  • Lists and list items
  • Tables
  • Images
  • Code blocks
  • Formulas
  • Charts
  • Other specialized document elements

• Positional Information:

  • Bounding box coordinates
  • Page numbers (for multi-page documents)
  • Relative positioning data
  • Layout flow indicators

• Example Structure:

<doc> <text loc="[x1, y1, x2, y2]">Extracted text content</text> <list> <item loc="[x1, y1, x2, y2]">List item 1</item> <item loc="[x1, y1, x2, y2]">List item 2</item> </list> <code loc="[x1, y1, x2, y2]"> def example(): return "code block" </code> </doc> 

Applications

• Intermediate representation in document processing pipelines
• Input format for downstream document analysis tasks
• Preserving document structure for accurate recreation
• Enabling spatial reasoning about document elements
• Supporting document element retrieval by type or location

Advantages

• Combines content and structure in a single representation
• Preserves spatial relationships between elements
• Enables selective processing of specific element types
• Supports reconstruction of visual layout
• Facilitates document navigation and search

Further Processing

• Conversion to other formats (Markdown, HTML, etc.)
• Extraction of specific element types for specialized processing
• Integration with LLMs for content analysis
• Database storage of structured document information
• Visualization of document structure

Connections

• Related Concepts: Document Understanding Models, Structured Document Representation
• Implementation Context: SmolDocling (uses this format), Document Conversion Pipelines
• Similar Formats: ALTO XML, hOCR (other document markup formats)
• Applications: Document Layout Analysis, Content Extraction Systems

References

1. SmolDocling paper and documentation
2. DocTags format specification
3. Document understanding literature

#DocumentMarkup #StructuredData #DocumentAI #InformationExtraction #DocumentRepresentation

Connections:

Sources:

• From: Sam Witteveen - SmolDocling ¿la solución SmolOCR