DoclingDocument Data Model#

DoclingDocument is the unified output format of Docling's document parsing pipeline. It is the canonical, schema-versioned JSON representation of a parsed document — carrying all texts, tables, images, structure, and layout provenance — used across Python, Java, and the HTTP API surface.

Python canonical source: docling_core/types/doc/document.py (Pydantic models)
Java source: docling-core/src/main/java/ai/docling/core/DoclingDocument.java 1 (Jackson-annotated, Lombok-backed)

Top-Level Structure#

The root class carries these fields 2:

The body and furniture nodes are GroupItems serving as tree roots; all other items are stored in flat lists and referenced by JSON Pointer ($ref, e.g. #/texts/0) via RefItem.

Item Types#

Text (BaseTextItem / texts list)#

A sealed interface with concrete implementations 3:

All text items share: selfRef, parent, children, contentLayer, label, prov (provenance list), orig (original text), text (normalized), formatting, hyperlink, source, comments.

Tables (TableItem / tables list)#

TableItem holds a data: TableData field. TableData contains:

• tableCells: flat List<TableCell> with per-cell text, span indices, columnHeader/rowHeader/rowSection flags, bbox, and an optional ref for rich cells
• grid: the reconstructed 2-D num_rows × num_cols matrix
• orientation: optional Orientation enum (ROT_0 / ROT_90 / etc.)

Pictures (PictureItem / pictures list)#

PictureItem has prov, image (ImageRef with mimetype, dpi, size, uri), captions, references, footnotes, and a rich PictureMeta holding classification (predicted class names + confidence), description, molecule (SMILES), tabular_chart, and code.

Key-Value / Forms (KeyValueItem, FormItem)#

KeyValueItem and FormItem both carry a graph: GraphData field. GraphData stores cells (GraphCell with label ∈ {KEY, VALUE, CHECKBOX}, text, prov) and links (GraphLink with directional label between cell IDs).

Provenance & Layout#

Every item carries a prov: List<ProvenanceItem>. ProvenanceItem records:

• pageNo: 1-indexed page number
• bbox: BoundingBox with l, t, r, b coordinates and optional coordOrigin ("TOPLEFT" or "BOTTOMLEFT")
• charspan: 2-element list [start, end] (0-indexed, end-exclusive)

PageItem in the pages map provides the page's physical size (width, height) and optional rendered image reference.

Content Classification#

Each node carries two classification fields 4:

• contentLayer (ContentLayer enum): BODY, FURNITURE, BACKGROUND, INVISIBLE, NOTES — used to filter items during export; default is BODY.
• label (DocItemLabel enum): structural type — TITLE, SECTION_HEADER, TABLE, PICTURE, PARAGRAPH, LIST_ITEM, CODE, FORMULA, CAPTION, FOOTNOTE, PAGE_HEADER, PAGE_FOOTER, CHECKBOX_SELECTED, CHECKBOX_UNSELECTED, etc.

Metadata (BaseMeta / FloatingMeta / PictureMeta)#

All nodes accept an optional meta field. BaseMeta provides: summary, language (BCP 47), entities (NER mentions with charspan), keywords, topics. FloatingMeta adds description; PictureMeta further adds classification, molecule, tabular_chart, code. Keywords and topics are deduplicated and order-preserving 5.

Java Library Details#

The Java model lives in docling-project/docling-java under docling-core/. It mirrors the Python wire format exactly using Jackson @JsonProperty for snake_case mapping and @JsonSubTypes for polymorphic text item deserialization. All builders are Lombok-generated and @Jacksonized. Null collection fields deserialize as empty lists via @JsonSetter(nulls = Nulls.AS_EMPTY) 6.

Test coverage in DoclingDocumentTests.java 7 provides concrete JSON round-trip examples for all major sub-types.

Key Source Files#