Document Content Classification#

Document content classification in Docling operates on two orthogonal axes:

1. ContentLayer — assigns each element a functional layer (body, furniture, background, etc.) for selective filtering during export and iteration.
2. DocItemLabel — assigns each element a structural type (text, table, picture, page header, etc.) used by downstream models and renderers.

These two attributes live on every NodeItem in the document tree. The classification is assigned during PDF pipeline assembly (primarily by the ReadingOrderModel) and by format-specific backends (Word, HTML, PPTX). All export and iteration methods accept included_content_layers to filter which content surfaces in the output.

ContentLayer Enum#

Defined in docling_core/types/doc/document.py:

DEFAULT_CONTENT_LAYERS = {ContentLayer.BODY} 1 is the fallback used when callers pass included_content_layers=None. This means CLI and API exports silently drop furniture, watermarks, and notes unless the caller explicitly opts in.

Every NodeItem defaults to ContentLayer.BODY at construction time.

DocItemLabel Enum#

Defined in docling_core/types/doc/labels.py. Selected values relevant to layer classification:

There is no dedicated WATERMARK DocItemLabel; watermark content is classified at the ContentLayer level (BACKGROUND) rather than as a distinct structural label.

For fine-grained picture classification, see PictureClassificationLabel, which includes chart types, photographs, barcodes, and chemistry structures — used by the DocumentPictureClassifier model.

Layer Assignment#

PDF Pipeline (Reading Order Model)#

The ReadingOrderModel is the primary site for layer assignment during PDF processing. It runs as part of StandardPdfPipeline._assemble_document(). Two methods apply layer logic:

• _add_child_elements: When building child items in composite elements, items labeled PAGE_HEADER or PAGE_FOOTER receive ContentLayer.FURNITURE; all others receive ContentLayer.BODY.
• _handle_text_element: Same logic applied when handling top-level text elements from the layout model output.

Format-Specific Backends#

• Word/DOCX: Headers and footers discovered during _add_header_footer() are tagged ContentLayer.FURNITURE directly by MsWordDocumentBackend.
• HTML: When infer_furniture=True, content before the first heading is tagged ContentLayer.FURNITURE; subsequent content is BODY.
• PPTX: Supports included_content_layers at export time but does not auto-assign non-BODY layers.

The BACKGROUND and INVISIBLE layers are defined in the enum but are not yet assigned by any built-in model or backend — watermark detection is not a current pipeline stage. They are reserved for future use or custom assignment.

Filtering by Content Layer#

All traversal and export methods accept included_content_layers: Optional[set[ContentLayer]]:

• DoclingDocument.iterate_items(): Yields only nodes whose content_layer is in the provided set. Falls back to DEFAULT_CONTENT_LAYERS (BODY only) when None.
• export_to_markdown(), export_to_html(), export_to_text(): All thread the same parameter through to iterate_items().

To include headers/footers in Markdown output:

from docling_core.types.doc import ContentLayer

doc.export_to_markdown(
    included_content_layers={ContentLayer.BODY, ContentLayer.FURNITURE}
)

The VLM pipeline passes included_content_layers=set(ContentLayer) 2 — all layers — when assembling the intermediate per-page document, ensuring no content is dropped before final post-processing.

Default behavior summary:

Key Source References#