Parsed Page Access and Segmentation#

Page.parsed_page is the central entry point for backend-produced text cells and their geometry. After a backend parses a page, it stores a SegmentedPdfPage (or the base SegmentedPage) in Page.parsed_page 1. The pipeline then accesses the cells via the page.cells property, which is a thin alias for parsed_page.textline_cells 2.

Data Model#

All types are defined in docling_core/types/doc/page.py.

TextCell and PdfTextCell#

TextCell is the base unit. Key fields:

PdfTextCell extends TextCell with rendering_mode (PDF32000 text rendering mode), widget, font_key, and font_name.

BoundingRectangle#

BoundingRectangle stores four explicit corner points (r_x0/r_y0 … r_x3/r_y3) rather than a simple LTRB box, enabling representation of rotated text. The default coord_origin is BOTTOMLEFT (matching PDF native coordinates). Use to_top_left_origin(page_height) / to_bottom_left_origin(page_height) to convert. to_bounding_box() returns an axis-aligned BoundingBox for overlap arithmetic.

SegmentedPage / SegmentedPdfPage#

SegmentedPage holds three granularity levels:

The has_chars / has_words / has_lines boolean flags distinguish "not computed" from "computed but empty" 3. Additional fields: bitmap_resources, widgets, hyperlinks, and an optional rendered image.

SegmentedPdfPage narrows dimension to PdfPageGeometry (exposes all five PDF boundary boxes: art, bleed, crop, media, trim) and refines the cell lists to list[Union[PdfTextCell, TextCell]]. It also holds shapes (vector graphics as PdfShape instances).

TextCellUnit is the enum (CHAR, WORD, LINE) used by iterators and query methods.

Coordinate System#

docling-parse emits coordinates in bottom-left origin (native PDF). Immediately after get_page() returns, _ensure_parsed() flips every cell list to top-left origin so the rest of the pipeline operates in a single consistent frame 4. All downstream code—layout models, table assembly—therefore receives top-left coordinates.

BoundingBox in docling_core/types/doc/base.py stores CoordOrigin alongside every box and enforces origin consistency in overlap/IoU calculations.

How Backends Populate parsed_page#

docling-parse (default PDF backend)#

DoclingParsePageBackend._ensure_parsed() is lazily called before any cell access 4. It creates a ContentConfig via _make_docling_parse_page_content_config() 5 which controls what is materialized:

• textline_cells — always materialized (the pipeline's primary input)
• word_cells — materialized when create_words=True
• char_cells — computed but not materialized unless keep_chars=True; char-level data drives word grouping but is discarded by default
• bitmap_resources — rectangle only (bitmap bytes never loaded)

get_text_cells() returns textline_cells for the layout pipeline 6. get_text_in_rect() filters textline cells by ≥50% intersection-over-self with a query bbox—used by table-cell text matching 7.

Other backends#

All backends expose the same PdfPageBackend interface (get_text_cells(), get_text_in_rect(), get_segmented_page()), making cell access polymorphic regardless of backend.

Key Query Methods on SegmentedPdfPage#

All defined in docling_core/types/doc/page.py:

• iterate_cells(unit_type) — yields cells for a given TextCellUnit
• get_cells_in_bbox(cell_unit, bbox, ios=0.8) — returns cells with intersection-over-self > threshold; handles cross-origin comparisons automatically
• crop_text(cell_unit, bbox) — extracts concatenated text from cells fully contained in a bbox
• export_to_textlines(cell_unit, ...) — debug-friendly text export with location, font, and direction metadata
• render_as_image(cell_unit, ...) — PIL image rendering with overlaid cell rectangles, shapes, bitmaps, and hyperlinks

Primary Source Files#