Files
docling-core/tests/test_docling_doc.py
Panos Vagenas 275503348b refactor(document): share page bbox traversal and add get_min_page_sizes (#1015)
Visit page bboxes through a single generator used by both the page-bounds
clamping and the page size computation, so the two cannot drift apart.
Rename get_page_extent to get_min_page_size, add get_min_page_sizes to
compute all pages in one pass, and document that negative coordinates are
still clamped.

Signed-off-by: Panos Vagenas <pva@zurich.ibm.com>
2026-11-12 12:56:56 +01:00

3727 lines
132 KiB
Python

import base64
import itertools
import os
import re
import warnings
from collections import deque
from collections.abc import Callable
from copy import deepcopy
from io import BytesIO
from pathlib import Path
from typing import Optional, Union
from unittest.mock import Mock
import pytest
import yaml
from PIL import Image as PILImage
from pydantic import AnyUrl, BaseModel, ValidationError
from docling_core.transforms.serializer.markdown import MarkdownDocSerializer
from docling_core.types.doc import (
BoundingBox,
CodeItem,
ContentLayer,
CoordOrigin,
DocItem,
DocItemLabel,
DoclingDocument,
DocumentOrigin,
FloatingItem,
Formatting,
FormItem,
FormulaItem,
GraphCell,
GraphCellLabel,
GraphData,
GraphLink,
GraphLinkLabel,
GroupLabel,
ImageRef,
ImageRefMode,
KeyValueItem,
ListGroup,
ListItem,
NodeItem,
Orientation,
PictureItem,
ProvenanceItem,
RefItem,
RichTableCell,
SectionHeaderItem,
Size,
TableCell,
TableData,
TableItem,
TextItem,
TitleItem,
)
from docling_core.types.doc.document import (
CURRENT_VERSION,
FieldHeadingItem,
FieldItem,
FieldRegionItem,
FieldValueItem,
PageItem,
)
from docling_core.types.doc.webvtt import WebVTTFile
from docling_core.utils.settings import settings
from .test_data_gen_flag import GEN_TEST_DATA
def test_doc_origin():
DocumentOrigin(
mimetype="application/vnd.openxmlformats-officedocument.wordprocessingml.document",
filename="myfile.pdf",
binary_hash="50115d582a0897fe1dd520a6876ec3f9321690ed0f6cfdc99a8d09019be073e8",
)
def test_overlaps_horizontally():
# Overlapping horizontally
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_horizontally(bbox2) is True
# No overlap horizontally (disjoint on the right)
bbox3 = BoundingBox(l=11, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_horizontally(bbox3) is False
# No overlap horizontally (disjoint on the left)
bbox4 = BoundingBox(l=-10, t=0, r=-1, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_horizontally(bbox4) is False
# Full containment
bbox5 = BoundingBox(l=2, t=2, r=8, b=8, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_horizontally(bbox5) is True
# Edge touching (no overlap)
bbox6 = BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_horizontally(bbox6) is False
def test_overlaps_vertically():
page_height = 300
# Same CoordOrigin (TOPLEFT)
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_vertically(bbox2) is True
bbox1_ = bbox1.to_bottom_left_origin(page_height=page_height)
bbox2_ = bbox2.to_bottom_left_origin(page_height=page_height)
assert bbox1_.overlaps_vertically(bbox2_) is True
bbox3 = BoundingBox(l=0, t=11, r=10, b=20, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.overlaps_vertically(bbox3) is False
bbox3_ = bbox3.to_bottom_left_origin(page_height=page_height)
assert bbox1_.overlaps_vertically(bbox3_) is False
# Same CoordOrigin (BOTTOMLEFT)
bbox4 = BoundingBox(l=0, b=20, r=10, t=30, coord_origin=CoordOrigin.BOTTOMLEFT)
bbox5 = BoundingBox(l=5, b=15, r=15, t=25, coord_origin=CoordOrigin.BOTTOMLEFT)
assert bbox4.overlaps_vertically(bbox5) is True
bbox4_ = bbox4.to_top_left_origin(page_height=page_height)
bbox5_ = bbox5.to_top_left_origin(page_height=page_height)
assert bbox4_.overlaps_vertically(bbox5_) is True
bbox6 = BoundingBox(l=0, b=31, r=10, t=40, coord_origin=CoordOrigin.BOTTOMLEFT)
assert bbox4.overlaps_vertically(bbox6) is False
bbox6_ = bbox6.to_top_left_origin(page_height=page_height)
assert bbox4_.overlaps_vertically(bbox6_) is False
# Different CoordOrigin
with pytest.raises(ValueError):
bbox1.overlaps_vertically(bbox4)
def test_intersection_area_with():
page_height = 300
# Overlapping bounding boxes (TOPLEFT)
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.intersection_area_with(bbox2) - 25.0) < 1.0e-3
bbox1_ = bbox1.to_bottom_left_origin(page_height=page_height)
bbox2_ = bbox2.to_bottom_left_origin(page_height=page_height)
assert abs(bbox1_.intersection_area_with(bbox2_) - 25.0) < 1.0e-3
# Non-overlapping bounding boxes (TOPLEFT)
bbox3 = BoundingBox(l=11, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.intersection_area_with(bbox3) - 0.0) < 1.0e-3
# Touching edges (no intersection, TOPLEFT)
bbox4 = BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.intersection_area_with(bbox4) - 0.0) < 1.0e-3
# Fully contained (TOPLEFT)
bbox5 = BoundingBox(l=2, t=2, r=8, b=8, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.intersection_area_with(bbox5) - 36.0) < 1.0e-3
# Overlapping bounding boxes (BOTTOMLEFT)
bbox6 = BoundingBox(l=0, t=10, r=10, b=0, coord_origin=CoordOrigin.BOTTOMLEFT)
bbox7 = BoundingBox(l=5, t=15, r=15, b=5, coord_origin=CoordOrigin.BOTTOMLEFT)
assert abs(bbox6.intersection_area_with(bbox7) - 25.0) < 1.0e-3
# Different CoordOrigins (raises ValueError)
with pytest.raises(ValueError):
bbox1.intersection_area_with(bbox6)
def test_x_overlap_with():
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=0, r=15, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox2) - 5.0) < 1.0e-3
# No overlap (disjoint right)
bbox3 = BoundingBox(l=11, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox3) - 0.0) < 1.0e-3
# No overlap (disjoint left)
bbox4 = BoundingBox(l=-10, t=0, r=-1, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox4) - 0.0) < 1.0e-3
# Touching edges
bbox5 = BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox5) - 0.0) < 1.0e-3
# Full containment
bbox6 = BoundingBox(l=2, t=0, r=8, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox6) - 6.0) < 1.0e-3
# Identical boxes
bbox7 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_overlap_with(bbox7) - 10.0) < 1.0e-3
# Different CoordOrigin
bbox_bl = BoundingBox(l=0, t=10, r=10, b=0, coord_origin=CoordOrigin.BOTTOMLEFT)
with pytest.raises(ValueError):
bbox1.x_overlap_with(bbox_bl)
def test_y_overlap_with():
# TOPLEFT origin
bbox1_tl = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2_tl = BoundingBox(l=0, t=5, r=10, b=15, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1_tl.y_overlap_with(bbox2_tl) - 5.0) < 1.0e-3
# No overlap (disjoint below)
bbox3_tl = BoundingBox(l=0, t=11, r=10, b=20, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1_tl.y_overlap_with(bbox3_tl) - 0.0) < 1.0e-3
# Touching edges
bbox4_tl = BoundingBox(l=0, t=10, r=10, b=20, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1_tl.y_overlap_with(bbox4_tl) - 0.0) < 1.0e-3
# Full containment
bbox5_tl = BoundingBox(l=0, t=2, r=10, b=8, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1_tl.y_overlap_with(bbox5_tl) - 6.0) < 1.0e-3
# BOTTOMLEFT origin
bbox1_bl = BoundingBox(l=0, b=0, r=10, t=10, coord_origin=CoordOrigin.BOTTOMLEFT)
bbox2_bl = BoundingBox(l=0, b=5, r=10, t=15, coord_origin=CoordOrigin.BOTTOMLEFT)
assert abs(bbox1_bl.y_overlap_with(bbox2_bl) - 5.0) < 1.0e-3
# No overlap (disjoint above)
bbox3_bl = BoundingBox(l=0, b=11, r=10, t=20, coord_origin=CoordOrigin.BOTTOMLEFT)
assert abs(bbox1_bl.y_overlap_with(bbox3_bl) - 0.0) < 1.0e-3
# Touching edges
bbox4_bl = BoundingBox(l=0, b=10, r=10, t=20, coord_origin=CoordOrigin.BOTTOMLEFT)
assert abs(bbox1_bl.y_overlap_with(bbox4_bl) - 0.0) < 1.0e-3
# Full containment
bbox5_bl = BoundingBox(l=0, b=2, r=10, t=8, coord_origin=CoordOrigin.BOTTOMLEFT)
assert abs(bbox1_bl.y_overlap_with(bbox5_bl) - 6.0) < 1.0e-3
# Different CoordOrigin
with pytest.raises(ValueError):
bbox1_tl.y_overlap_with(bbox1_bl)
def test_union_area_with():
# Overlapping (TOPLEFT)
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT) # Area 100
bbox2 = BoundingBox(l=5, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT) # Area 100
# Intersection area 25
# Union area = 100 + 100 - 25 = 175
assert abs(bbox1.union_area_with(bbox2) - 175.0) < 1.0e-3
# Non-overlapping (TOPLEFT)
bbox3 = BoundingBox(l=20, t=0, r=30, b=10, coord_origin=CoordOrigin.TOPLEFT) # Area 100
# Union area = 100 + 100 - 0 = 200
assert abs(bbox1.union_area_with(bbox3) - 200.0) < 1.0e-3
# Touching edges (TOPLEFT)
bbox4 = BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT) # Area 100
# Union area = 100 + 100 - 0 = 200
assert abs(bbox1.union_area_with(bbox4) - 200.0) < 1.0e-3
# Full containment (TOPLEFT)
bbox5 = BoundingBox(l=2, t=2, r=8, b=8, coord_origin=CoordOrigin.TOPLEFT) # Area 36
# Union area = 100 + 36 - 36 = 100
assert abs(bbox1.union_area_with(bbox5) - 100.0) < 1.0e-3
# Overlapping (BOTTOMLEFT)
bbox6 = BoundingBox(l=0, b=0, r=10, t=10, coord_origin=CoordOrigin.BOTTOMLEFT) # Area 100
bbox7 = BoundingBox(l=5, b=5, r=15, t=15, coord_origin=CoordOrigin.BOTTOMLEFT) # Area 100
# Intersection area 25
# Union area = 100 + 100 - 25 = 175
assert abs(bbox6.union_area_with(bbox7) - 175.0) < 1.0e-3
# Different CoordOrigin
with pytest.raises(ValueError):
bbox1.union_area_with(bbox6)
def test_x_union_with():
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=0, r=15, b=10, coord_origin=CoordOrigin.TOPLEFT)
# x_union = max(10, 15) - min(0, 5) = 15 - 0 = 15
assert abs(bbox1.x_union_with(bbox2) - 15.0) < 1.0e-3
# No overlap (disjoint)
bbox3 = BoundingBox(l=20, t=0, r=30, b=10, coord_origin=CoordOrigin.TOPLEFT)
# x_union = max(10, 30) - min(0, 20) = 30 - 0 = 30
assert abs(bbox1.x_union_with(bbox3) - 30.0) < 1.0e-3
# Touching edges
bbox4 = BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT)
# x_union = max(10, 20) - min(0, 10) = 20 - 0 = 20
assert abs(bbox1.x_union_with(bbox4) - 20.0) < 1.0e-3
# Full containment
bbox5 = BoundingBox(l=2, t=0, r=8, b=10, coord_origin=CoordOrigin.TOPLEFT)
# x_union = max(10, 8) - min(0, 2) = 10 - 0 = 10
assert abs(bbox1.x_union_with(bbox5) - 10.0) < 1.0e-3
# Identical boxes
bbox6 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
assert abs(bbox1.x_union_with(bbox6) - 10.0) < 1.0e-3
# Different CoordOrigin
bbox_bl = BoundingBox(l=0, t=10, r=10, b=0, coord_origin=CoordOrigin.BOTTOMLEFT)
with pytest.raises(ValueError):
bbox1.x_union_with(bbox_bl)
def test_y_union_with():
bbox1_tl = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2_tl = BoundingBox(l=0, t=5, r=10, b=15, coord_origin=CoordOrigin.TOPLEFT)
# y_union = max(10, 15) - min(0, 5) = 15 - 0 = 15
assert abs(bbox1_tl.y_union_with(bbox2_tl) - 15.0) < 1.0e-3
# No overlap (disjoint below)
bbox3_tl = BoundingBox(l=0, t=20, r=10, b=30, coord_origin=CoordOrigin.TOPLEFT)
# y_union = max(10, 30) - min(0, 20) = 30 - 0 = 30
assert abs(bbox1_tl.y_union_with(bbox3_tl) - 30.0) < 1.0e-3
# Touching edges
bbox4_tl = BoundingBox(l=0, t=10, r=10, b=20, coord_origin=CoordOrigin.TOPLEFT)
# y_union = max(10, 20) - min(0, 10) = 20 - 0 = 20
assert abs(bbox1_tl.y_union_with(bbox4_tl) - 20.0) < 1.0e-3
# Full containment
bbox5_tl = BoundingBox(l=0, t=2, r=10, b=8, coord_origin=CoordOrigin.TOPLEFT)
# y_union = max(10, 8) - min(0, 2) = 10 - 0 = 10
assert abs(bbox1_tl.y_union_with(bbox5_tl) - 10.0) < 1.0e-3
# BOTTOMLEFT origin
bbox1_bl = BoundingBox(l=0, b=0, r=10, t=10, coord_origin=CoordOrigin.BOTTOMLEFT)
bbox2_bl = BoundingBox(l=0, b=5, r=10, t=15, coord_origin=CoordOrigin.BOTTOMLEFT)
# y_union = max(10, 15) - min(0, 5) = 15 - 0 = 15
assert abs(bbox1_bl.y_union_with(bbox2_bl) - 15.0) < 1.0e-3
# No overlap (disjoint above)
bbox3_bl = BoundingBox(l=0, b=20, r=10, t=30, coord_origin=CoordOrigin.BOTTOMLEFT)
# y_union = max(10, 30) - min(0, 20) = 30 - 0 = 30
assert abs(bbox1_bl.y_union_with(bbox3_bl) - 30.0) < 1.0e-3
# Touching edges
bbox4_bl = BoundingBox(l=0, b=10, r=10, t=20, coord_origin=CoordOrigin.BOTTOMLEFT)
# y_union = max(10, 20) - min(0, 10) = 20 - 0 = 20
assert abs(bbox1_bl.y_union_with(bbox4_bl) - 20.0) < 1.0e-3
# Full containment
bbox5_bl = BoundingBox(l=0, b=2, r=10, t=8, coord_origin=CoordOrigin.BOTTOMLEFT)
# y_union = max(10, 8) - min(0, 2) = 10 - 0 = 10
assert abs(bbox1_bl.y_union_with(bbox5_bl) - 10.0) < 1.0e-3
# Different CoordOrigin
with pytest.raises(ValueError):
bbox1_tl.y_union_with(bbox1_bl)
def test_orientation():
page_height = 300
# Same CoordOrigin (TOPLEFT)
bbox1 = BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT)
bbox2 = BoundingBox(l=5, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT)
bbox3 = BoundingBox(l=11, t=5, r=15, b=15, coord_origin=CoordOrigin.TOPLEFT)
bbox4 = BoundingBox(l=0, t=11, r=10, b=15, coord_origin=CoordOrigin.TOPLEFT)
assert bbox1.is_left_of(bbox2) is True
assert bbox1.is_strictly_left_of(bbox2) is False
assert bbox1.is_strictly_left_of(bbox3) is True
bbox1_ = bbox1.to_bottom_left_origin(page_height=page_height)
bbox2_ = bbox2.to_bottom_left_origin(page_height=page_height)
bbox3.to_bottom_left_origin(page_height=page_height)
bbox4_ = bbox4.to_bottom_left_origin(page_height=page_height)
assert bbox1.is_above(bbox2) is True
assert bbox1_.is_above(bbox2_) is True
assert bbox1.is_strictly_above(bbox4) is True
assert bbox1_.is_strictly_above(bbox4_) is True
def test_docitems():
# Iterative function to find all subclasses
def find_all_subclasses_iterative(base_class):
subclasses = deque([base_class]) # Use a deque for efficient popping from the front
all_subclasses = []
while subclasses:
current_class = subclasses.popleft() # Get the next class to process
for subclass in current_class.__subclasses__():
all_subclasses.append(subclass)
subclasses.append(subclass) # Add the subclass for further exploration
return all_subclasses
def serialise(obj):
return yaml.safe_dump(obj.model_dump(mode="json", by_alias=True))
def write(name: str, serialisation: str):
with open(f"./tests/data/docling_document/unit/{name}.yaml", "w", encoding="utf-8") as fw:
fw.write(serialisation)
def read(name: str):
with open(f"./tests/data/docling_document/unit/{name}.yaml", encoding="utf-8") as fr:
gold = fr.read()
return yaml.safe_load(gold)
def verify(dc, obj):
pred_text = serialise(obj).strip()
pred = yaml.safe_load(pred_text)
if GEN_TEST_DATA:
# Rewrite only when the content genuinely differs. `safe_dump`
# sorts keys while the committed fixtures preserve insertion
# order, so writing unconditionally would reorder every file
# without changing what any of them mean.
try:
outdated = read(dc.__name__) != pred
except (FileNotFoundError, yaml.YAMLError):
# Missing or unparseable: regenerating is the whole point of
# the flag, so do not let a corrupt fixture block it.
outdated = True
if outdated:
write(dc.__name__, pred_text)
gold = read(dc.__name__)
assert pred == gold, f"pred!=gold for {dc.__name__}"
# Iterate over the derived classes of the BaseClass
derived_classes = find_all_subclasses_iterative(DocItem)
for dc in derived_classes:
if dc is TextItem:
obj = dc(
text="whatever",
orig="whatever",
label=DocItemLabel.TEXT,
self_ref="#",
)
verify(dc, obj)
elif dc is ListItem:
obj = dc(
text="whatever",
orig="whatever",
marker="(1)",
enumerated=True,
self_ref="#",
)
verify(dc, obj)
elif dc is FloatingItem:
obj = dc(
label=DocItemLabel.TEXT,
self_ref="#",
)
verify(dc, obj)
elif dc is KeyValueItem:
graph = GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=0,
text="number",
orig="#",
),
GraphCell(
label=GraphCellLabel.VALUE,
cell_id=1,
text="1",
orig="1",
),
],
links=[
GraphLink(
label=GraphLinkLabel.TO_VALUE,
source_cell_id=0,
target_cell_id=1,
),
GraphLink(label=GraphLinkLabel.TO_KEY, source_cell_id=1, target_cell_id=0),
],
)
obj = dc(
label=DocItemLabel.KEY_VALUE_REGION,
graph=graph,
self_ref="#",
)
verify(dc, obj)
elif dc is FormItem:
graph = GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=0,
text="number",
orig="#",
),
GraphCell(
label=GraphCellLabel.VALUE,
cell_id=1,
text="1",
orig="1",
),
],
links=[
GraphLink(
label=GraphLinkLabel.TO_VALUE,
source_cell_id=0,
target_cell_id=1,
),
GraphLink(label=GraphLinkLabel.TO_KEY, source_cell_id=1, target_cell_id=0),
],
)
obj = dc(
label=DocItemLabel.FORM,
graph=graph,
self_ref="#",
)
verify(dc, obj)
elif dc is TitleItem:
obj = dc(
text="whatever",
orig="whatever",
label=DocItemLabel.TITLE,
self_ref="#",
)
verify(dc, obj)
elif dc is SectionHeaderItem:
obj = dc(
text="whatever",
orig="whatever",
label=DocItemLabel.SECTION_HEADER,
self_ref="#",
level=2,
)
verify(dc, obj)
elif dc is PictureItem:
obj = dc(
self_ref="#",
)
verify(dc, obj)
elif dc is TableItem:
obj = dc(
self_ref="#",
data=TableData(num_rows=3, num_cols=5, table_cells=[]),
)
verify(dc, obj)
elif dc is CodeItem:
obj = dc(
self_ref="#",
orig="whatever",
text="print(Hello World!)",
code_language="Python",
)
verify(dc, obj)
elif dc is FormulaItem:
obj = dc(
self_ref="#",
orig="whatever",
text="E=mc^2",
)
verify(dc, obj)
elif dc is FieldRegionItem:
obj = dc(
self_ref="#",
)
verify(dc, obj)
elif dc is FieldItem:
obj = dc(
self_ref="#",
)
verify(dc, obj)
elif dc is FieldValueItem:
obj = dc(
self_ref="#",
orig="whatever",
text="whatever",
kind="fillable",
)
verify(dc, obj)
elif dc is FieldHeadingItem:
obj = dc(
text="whatever",
orig="whatever",
label=DocItemLabel.FIELD_HEADING,
self_ref="#",
level=2,
)
verify(dc, obj)
elif dc is GraphData: # we skip this on purpose
continue
else:
raise RuntimeError(f"New derived class detected {dc.__name__}")
def test_reference_doc():
filename = "tests/data/doc/dummy_doc.yaml"
# Read YAML file of manual reference doc
with open(filename, encoding="utf-8") as fp:
dict_from_yaml = yaml.safe_load(fp)
doc = DoclingDocument.model_validate(dict_from_yaml)
# Objects can be accessed
text_item = doc.texts[0]
# access members
text_item.text
text_item.prov[0].page_no
# Objects that are references need explicit resolution for now:
obj = doc.texts[2] # Text item with parent
parent = obj.parent.resolve(doc=doc) # it is a figure
obj2 = parent.children[0].resolve(doc=doc) # Child of figure must be the same as obj
assert obj == obj2
assert obj is obj2
# Iterate all elements
for item, level in doc.iterate_items():
_ = f"Item: {item} at level {level}"
# print(f"Item: {item} at level {level}")
# Serialize and reload
_test_serialize_and_reload(doc)
# Call Export methods
_test_export_methods(doc, filename=filename)
def test_parse_doc():
filename = "tests/data/doc/2206.01062.yaml"
with open(filename, encoding="utf-8") as fp:
dict_from_yaml = yaml.safe_load(fp)
doc = DoclingDocument.model_validate(dict_from_yaml)
page_break = "<!-- page break -->"
_test_export_methods(doc, filename=filename, page_break_placeholder=page_break)
_test_serialize_and_reload(doc)
def test_construct_doc(sample_doc):
filename = "tests/data/doc/constructed_document.yaml"
assert sample_doc.validate_tree(sample_doc.body)
# check that deprecation warning for furniture has been raised.
with pytest.warns(DeprecationWarning, match="deprecated"):
assert sample_doc.validate_tree(sample_doc.furniture)
_test_export_methods(sample_doc, filename=filename)
_test_serialize_and_reload(sample_doc)
def test_construct_bad_doc():
filename = "tests/data/doc/bad_doc.yaml"
doc = _construct_bad_doc()
with pytest.raises(ValueError):
doc.validate_tree(doc.body, raise_on_error=True)
with pytest.raises(ValueError):
_test_export_methods(doc, filename=filename)
with pytest.raises(ValueError):
_test_serialize_and_reload(doc)
def _test_serialize_and_reload(doc):
### Serialize and deserialize stuff
yaml_dump = yaml.safe_dump(doc.model_dump(mode="json", by_alias=True))
# print(f"\n\n{yaml_dump}")
doc_reload = DoclingDocument.model_validate(yaml.safe_load(yaml_dump))
yaml_dump_reload = yaml.safe_dump(doc_reload.model_dump(mode="json", by_alias=True))
assert yaml_dump == yaml_dump_reload, "yaml_dump!=yaml_dump_reload"
"""
for item, level in doc.iterate_items():
if isinstance(item, PictureItem):
_ = item.get_image(doc)
assert doc_reload == doc # must be equal
"""
assert doc_reload is not doc # can't be identical
def _verify_regression_test(pred: str, filename: str, ext: str):
if os.path.exists(filename + f".{ext}") and not GEN_TEST_DATA:
with open(filename + f".{ext}", encoding="utf-8") as fr:
gt_true = fr.read().rstrip()
assert gt_true == pred, f"Does not pass regression-test for {filename}.{ext}\n\n{gt_true}\n\n{pred}"
else:
with open(filename + f".{ext}", "w", encoding="utf-8") as fw:
fw.write(f"{pred}\n")
def _test_export_methods(doc: DoclingDocument, filename: str, page_break_placeholder: str | None = None):
# Iterate all elements
et_pred = doc.export_to_element_tree()
_verify_regression_test(et_pred, filename=filename, ext="et")
# Export stuff
md_pred = doc.export_to_markdown()
_verify_regression_test(md_pred, filename=filename, ext="md")
if page_break_placeholder is not None:
md_pred = doc.export_to_markdown(page_break_placeholder=page_break_placeholder)
_verify_regression_test(md_pred, filename=filename, ext="paged.md")
# Test sHTML export ...
html_pred = doc.export_to_html()
_verify_regression_test(html_pred, filename=filename, ext="html")
# Test DocTags export ...
dt_pred = doc.export_to_doctags()
_verify_regression_test(dt_pred, filename=filename, ext="dt")
dt_min_pred = doc.export_to_doctags(minified=True)
_verify_regression_test(dt_min_pred, filename=filename, ext="min.dt")
# Test pages parameter in DocTags export
if doc.pages: # Only test if document has pages
first_page = min(doc.pages.keys())
second_page = first_page + 1
if second_page in doc.pages: # Only test if document has at least 2 pages
dt_pages_pred = doc.export_to_doctags(pages={first_page, second_page})
# print(dt_pages_pred)
_verify_regression_test(dt_pages_pred, filename=filename, ext="pages.dt")
# Test WebVTT export ...
# Note: Documents without TrackSource will result in empty WebVTT, but this is valid
vtt_pred = doc.export_to_vtt()
_verify_regression_test(vtt_pred, filename=filename, ext="vtt")
parsed = WebVTTFile.parse(vtt_pred)
assert isinstance(parsed, WebVTTFile)
assert not parsed.cue_blocks
# Test Tables export ...
for table in doc.tables:
table.export_to_markdown()
table.export_to_html(doc)
table.export_to_dataframe(doc)
table.export_to_doctags(doc)
# Test Images export ...
for fig in doc.pictures:
fig.export_to_doctags(doc)
def _construct_bad_doc():
doc = DoclingDocument(name="Bad doc")
title = doc.add_text(label=DocItemLabel.TITLE, text="This is the title")
group = doc.add_group(parent=title, name="chapter 1")
text = doc.add_text(
parent=group,
label=DocItemLabel.SECTION_HEADER,
text="This is the first section",
)
# Bend the parent of an element to be another.
text.parent = title.get_ref()
return doc
def test_pil_image():
doc = DoclingDocument(name="Untitled 1")
fig_image = PILImage.new(mode="RGB", size=(2, 2), color=(0, 0, 0))
doc.add_picture(image=ImageRef.from_pil(image=fig_image, dpi=72))
### Serialize and deserialize the document
yaml_dump = yaml.safe_dump(doc.model_dump(mode="json", by_alias=True))
doc_reload = DoclingDocument.model_validate(yaml.safe_load(yaml_dump))
reloaded_fig = doc_reload.pictures[0]
reloaded_image = reloaded_fig.image.pil_image
assert isinstance(reloaded_image, PILImage.Image)
assert reloaded_image.size == fig_image.size
assert reloaded_image.mode == fig_image.mode
assert reloaded_image.tobytes() == fig_image.tobytes()
def test_image_ref():
data_uri = {
"dpi": 72,
"mimetype": "image/png",
"size": {"width": 10, "height": 11},
"uri": "file:///tests/data/image.png",
}
image = ImageRef.model_validate(data_uri)
assert isinstance(image.uri, AnyUrl)
assert image.uri.scheme == "file"
assert image.uri.path == "/tests/data/image.png"
data_path = {
"dpi": 72,
"mimetype": "image/png",
"size": {"width": 10, "height": 11},
"uri": "./tests/data/image.png",
}
image = ImageRef.model_validate(data_path)
assert isinstance(image.uri, Path)
def test_image_ref_from_pil_png_compression():
"""PNGs from from_pil must not be larger than PIL's default compression level."""
from PIL import ImageDraw
image = PILImage.new("RGB", (600, 800), "white")
draw = ImageDraw.Draw(image)
for i in range(20):
draw.rectangle([40, 40 + i * 35, 560, 60 + i * 35], fill=(30, 30, 30))
draw.ellipse([100, 600, 400, 780], fill=(200, 40, 40))
image_ref = ImageRef.from_pil(image=image, dpi=72)
encoded_size = len(base64.b64decode(str(image_ref.uri).split(",")[1]))
buffer = BytesIO()
image.save(buffer, format="PNG")
assert encoded_size <= len(buffer.getvalue()) * 1.2
def test_image_ref_blocks_file_scheme():
"""Test that file:// URI scheme is blocked."""
fig_image = PILImage.new(mode="RGB", size=(2, 2), color=(0, 0, 0))
image_ref = ImageRef.from_pil(image=fig_image, dpi=72)
image_ref.uri = AnyUrl("file:///tmp/test.png")
with pytest.raises(ValueError, match="Reading local image files"):
_ = image_ref.pil_image
def test_image_ref_blocks_oversized_base64():
"""Test that oversized base64 data URIs are blocked."""
import base64
large_bytes = b"X" * (28 * 1024 * 1024)
large_data = base64.b64encode(large_bytes).decode("ascii")
data_uri = f"data:image/png;base64,{large_data}"
image_ref = ImageRef(dpi=72, mimetype="image/png", size=Size(width=100, height=100), uri=AnyUrl(data_uri))
with pytest.raises(ValueError, match="exceeds size limit"):
_ = image_ref.pil_image
def test_image_ref_accepts_valid_base64():
"""Test that valid base64 data URIs within size limit work correctly."""
import base64
from io import BytesIO
fig_image = PILImage.new(mode="RGB", size=(1, 1), color=(255, 0, 0))
# Convert to base64 data URI
buffer = BytesIO()
fig_image.save(buffer, format="PNG")
img_bytes = buffer.getvalue()
img_base64 = base64.b64encode(img_bytes).decode("ascii")
data_uri = f"data:image/png;base64,{img_base64}"
# Create ImageRef with data URI
image_ref = ImageRef(dpi=72, mimetype="image/png", size=Size(width=1, height=1), uri=AnyUrl(data_uri))
# Should successfully decode the image
decoded_image = image_ref.pil_image
assert isinstance(decoded_image, PILImage.Image)
assert decoded_image.size == (1, 1)
assert decoded_image.mode == "RGB"
def test_file_uri_allowed_with_env_var():
"""Test that file:// URIs work when enabled via settings."""
test_img_path = Path("/tmp/test_docling_env.png")
img = PILImage.new("RGB", (100, 100), color="red")
img.save(test_img_path)
orig_allow_image_file_uri = settings.allow_image_file_uri
try:
settings.allow_image_file_uri = True
image_ref = ImageRef(
dpi=72,
mimetype="image/png",
size=Size(width=100, height=100),
uri=AnyUrl(f"file://{test_img_path}"),
)
pil_img = image_ref.pil_image
assert pil_img is not None
assert pil_img.size == (100, 100)
assert pil_img.mode == "RGB"
finally:
test_img_path.unlink(missing_ok=True)
settings.allow_image_file_uri = orig_allow_image_file_uri
def test_file_uri_blocked_by_default():
"""Test that file:// URIs are blocked by default."""
image_ref = ImageRef(
dpi=72,
mimetype="image/png",
size=Size(width=100, height=100),
uri=AnyUrl("file:///tmp/test.png"),
)
with pytest.raises(ValueError, match="Reading local image files"):
_ = image_ref.pil_image
def test_max_decoded_size_custom():
"""Test that oversized images are rejected based on custom limit."""
orig_max_image_decoded_size = settings.max_image_decoded_size
try:
settings.max_image_decoded_size = 100 # 100 bytes limit
# Create image that will exceed 100 bytes when base64 decoded
# A 50x50 RGB image is 50*50*3 = 7500 bytes uncompressed
img = PILImage.new("RGB", (50, 50), color="green")
buffer = BytesIO()
img.save(buffer, format="PNG")
img_bytes = buffer.getvalue()
# Verify the decoded size will exceed our limit
assert len(img_bytes) > 100, f"Test image is only {len(img_bytes)} bytes, need > 100"
encoded = base64.b64encode(img_bytes).decode("utf-8")
data_uri = f"data:image/png;base64,{encoded}"
image_ref = ImageRef(
dpi=72,
mimetype="image/png",
size=Size(width=50, height=50),
uri=AnyUrl(data_uri),
)
with pytest.raises(ValueError, match="Decoded image exceeds size limit"):
_ = image_ref.pil_image
finally:
settings.max_image_decoded_size = orig_max_image_decoded_size
def test_max_decoded_size_default():
"""Test that small images work with default 20MB limit."""
img = PILImage.new("RGB", (100, 100), color="blue")
buffer = BytesIO()
img.save(buffer, format="PNG")
img_bytes = buffer.getvalue()
encoded = base64.b64encode(img_bytes).decode("utf-8")
data_uri = f"data:image/png;base64,{encoded}"
image_ref = ImageRef(
dpi=72,
mimetype="image/png",
size=Size(width=100, height=100),
uri=AnyUrl(data_uri),
)
pil_img = image_ref.pil_image
assert pil_img is not None
assert pil_img.size == (100, 100)
def test_image_ref_path_rejects_oversize_file(tmp_path: Path):
"""Path-based ImageRef must enforce ``max_image_decoded_size`` like data URIs."""
oversize = tmp_path / "big.png"
oversize.write_bytes(b"\x89PNG\r\n\x1a\n" + (b"\x00" * 256))
orig_max = settings.max_image_decoded_size
orig_allow = settings.allow_image_file_uri
try:
settings.max_image_decoded_size = 64
settings.allow_image_file_uri = True
image_ref = ImageRef(
dpi=72,
mimetype="image/png",
size=Size(width=1, height=1),
uri=oversize,
)
with pytest.raises(ValueError, match="exceeds size limit"):
_ = image_ref.pil_image
finally:
settings.max_image_decoded_size = orig_max
settings.allow_image_file_uri = orig_allow
def _doc_with_local_picture(uri: str) -> DoclingDocument:
"""Build a document as loaded from untrusted JSON, with a picture pointing at ``uri``."""
doc = DoclingDocument(name="local_picture")
pic = doc.add_picture()
pic.image = ImageRef.from_pil(PILImage.new("RGB", (1, 1)), dpi=72)
data = doc.export_to_dict()
data["pictures"][0]["image"]["uri"] = uri
return DoclingDocument.model_validate(data)
def _embedded_exports(doc: DoclingDocument, out_dir: Path) -> dict[str, Callable[[], str]]:
def _save(fn, name: str) -> str:
fn(out_dir / name, image_mode=ImageRefMode.EMBEDDED)
return (out_dir / name).read_text()
return {
"export_to_markdown": lambda: doc.export_to_markdown(image_mode=ImageRefMode.EMBEDDED),
"export_to_html": lambda: doc.export_to_html(image_mode=ImageRefMode.EMBEDDED),
"save_as_markdown": lambda: _save(doc.save_as_markdown, "out.md"),
"save_as_html": lambda: _save(doc.save_as_html, "out.html"),
}
@pytest.mark.parametrize("as_file_uri", [False, True])
@pytest.mark.parametrize("export", ["export_to_markdown", "export_to_html", "save_as_markdown", "save_as_html"])
def test_local_image_blocked_in_embedded_exports(tmp_path: Path, as_file_uri: bool, export: str):
"""Plain paths and file:// URIs must not be read on embedding unless enabled."""
secret = tmp_path / "secret.png"
PILImage.new("RGB", (4, 4), color=(11, 222, 33)).save(secret)
doc = _doc_with_local_picture(secret.as_uri() if as_file_uri else str(secret))
with pytest.raises(ValueError, match="Reading local image files"):
_embedded_exports(doc, tmp_path)[export]()
@pytest.mark.parametrize("as_file_uri", [False, True])
@pytest.mark.parametrize("export", ["export_to_markdown", "export_to_html", "save_as_markdown", "save_as_html"])
def test_local_image_allowed_in_embedded_exports_when_enabled(
tmp_path: Path, monkeypatch: pytest.MonkeyPatch, as_file_uri: bool, export: str
):
"""With the opt-in, local images are embedded as before."""
monkeypatch.setattr(settings, "allow_image_file_uri", True)
img = tmp_path / "img.png"
PILImage.new("RGB", (4, 4), color=(11, 222, 33)).save(img)
doc = _doc_with_local_picture(img.as_uri() if as_file_uri else str(img))
assert "data:image/png;base64," in _embedded_exports(doc, tmp_path)[export]()
def test_local_image_blocked_in_referenced_export(tmp_path: Path):
"""Referenced exports must not copy local files into the artifacts directory."""
secret = tmp_path / "secret.png"
PILImage.new("RGB", (4, 4)).save(secret)
doc = _doc_with_local_picture(str(secret))
with pytest.raises(ValueError, match="Reading local image files"):
doc.save_as_markdown(tmp_path / "out.md", image_mode=ImageRefMode.REFERENCED)
def test_local_page_image_blocked(tmp_path: Path):
"""Page images with a local path must be gated when cropping pictures."""
secret = tmp_path / "page.png"
PILImage.new("RGB", (10, 10)).save(secret)
doc = DoclingDocument(name="local_page")
doc.add_page(
page_no=1,
size=Size(width=10, height=10),
image=ImageRef(mimetype="image/png", dpi=72, size=Size(width=10, height=10), uri=secret),
)
doc.add_picture(
prov=ProvenanceItem(
page_no=1, bbox=BoundingBox(l=0, t=10, r=5, b=5, coord_origin=CoordOrigin.BOTTOMLEFT), charspan=(0, 0)
)
)
with pytest.raises(ValueError, match="Reading local image files"):
doc.export_to_markdown(image_mode=ImageRefMode.EMBEDDED)
def _save_referenced_doc(out_dir: Path, filename: str) -> Path:
"""Save a document with a picture and a page image in referenced mode."""
doc = DoclingDocument(name="referenced")
page_image = PILImage.new("RGB", (10, 10), color=(11, 222, 33))
doc.add_page(page_no=1, size=Size(width=10, height=10), image=ImageRef.from_pil(page_image, dpi=72))
doc.add_picture(image=ImageRef.from_pil(PILImage.new("RGB", (4, 4), color=(1, 2, 3)), dpi=72))
path = out_dir / filename
if path.suffix == ".json":
doc.save_as_json(path, image_mode=ImageRefMode.REFERENCED)
else:
doc.save_as_yaml(path, image_mode=ImageRefMode.REFERENCED)
return path
@pytest.mark.parametrize("filename", ["doc.json", "doc.yaml"])
def test_load_with_allow_local_images(tmp_path: Path, monkeypatch: pytest.MonkeyPatch, filename: str):
"""Loaders read local images on explicit opt-in, relative to the file's directory."""
path = _save_referenced_doc(tmp_path, filename)
other_dir = tmp_path / "elsewhere"
other_dir.mkdir()
monkeypatch.chdir(other_dir)
load = DoclingDocument.load_from_json if filename.endswith(".json") else DoclingDocument.load_from_yaml
with pytest.raises(ValueError, match="Reading local image files"):
load(path).export_to_markdown(image_mode=ImageRefMode.EMBEDDED)
doc = load(path, allow_local_images=True)
assert "data:image/png;base64," in doc.export_to_markdown(image_mode=ImageRefMode.EMBEDDED)
assert doc.pictures[0].image is not None
assert (picture_pil := doc.pictures[0].image.pil_image) is not None
assert picture_pil.getpixel((0, 0)) == (1, 2, 3)
assert doc.pages[1].image is not None
assert (page_pil := doc.pages[1].image.pil_image) is not None
assert page_pil.getpixel((0, 0)) == (11, 222, 33)
def test_load_local_images_ignores_later_uri_changes(tmp_path: Path):
"""Only the images present when opting in are read; later paths are not."""
doc = DoclingDocument.load_from_json(_save_referenced_doc(tmp_path, "doc.json"), allow_local_images=True)
secret = tmp_path / "secret.png"
PILImage.new("RGB", (4, 4), color=(255, 0, 0)).save(secret)
image = doc.pictures[0].image
assert image is not None
image.uri = secret
assert (pil := image.pil_image) is not None
assert pil.getpixel((0, 0)) == (1, 2, 3)
doc.add_picture(image=ImageRef(mimetype="image/png", dpi=72, size=Size(width=4, height=4), uri=secret))
with pytest.raises(ValueError, match="Reading local image files"):
doc.export_to_markdown(image_mode=ImageRefMode.EMBEDDED)
def test_load_local_images_rejects_non_regular_file(tmp_path: Path):
doc = _doc_with_local_picture(str(tmp_path))
with pytest.raises(ValueError, match="not a regular file"):
doc.load_local_images()
def test_load_local_images_rejects_oversize_file(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
big = tmp_path / "big.png"
PILImage.new("RGB", (50, 50), color=(1, 2, 3)).save(big)
monkeypatch.setattr(settings, "max_image_decoded_size", 64)
doc = _doc_with_local_picture(str(big))
with pytest.raises(ValueError, match="exceeds size limit"):
doc.load_local_images()
def test_upgrade_content_layer_from_1_0_0() -> None:
doc = DoclingDocument.load_from_json("tests/data/doc/2206.01062-1.0.0.json")
assert doc.version == CURRENT_VERSION
assert doc.texts[0].content_layer == ContentLayer.FURNITURE
# test that transform_to_content_layer model validator can handle any data type
class ContentOutput(BaseModel):
content: str | DoclingDocument
co = ContentOutput.model_validate_json('{"content": "Random string with version"}')
assert co
assert isinstance(co.content, str)
def test_version_doc():
# default version
doc = DoclingDocument(name="Untitled 1")
assert doc.version == CURRENT_VERSION
with open("tests/data/doc/dummy_doc.yaml", encoding="utf-8") as fp:
dict_from_yaml = yaml.safe_load(fp)
doc = DoclingDocument.model_validate(dict_from_yaml)
assert doc.version == CURRENT_VERSION
# invalid version
with pytest.raises(ValidationError, match="NoneType"):
DoclingDocument(name="Untitled 1", version=None)
with pytest.raises(ValidationError, match="pattern"):
DoclingDocument(name="Untitled 1", version="abc")
# incompatible version (major)
major_split = CURRENT_VERSION.split(".", 1)
new_version = f"{int(major_split[0]) + 1}.{major_split[1]}"
with pytest.raises(ValidationError, match="incompatible"):
DoclingDocument(name="Untitled 1", version=new_version)
# incompatible version (minor)
minor_split = major_split[1].split(".", 1)
new_version = f"{major_split[0]}.{int(minor_split[0]) + 1}.{minor_split[1]}"
with pytest.raises(ValidationError, match="incompatible"):
DoclingDocument(name="Untitled 1", version=new_version)
# compatible version (equal or lower minor)
patch_split = minor_split[1].split(".", 1)
comp_version = f"{major_split[0]}.{minor_split[0]}.{int(patch_split[0]) + 1}"
doc = DoclingDocument(name="Untitled 1", version=comp_version)
assert doc.version == CURRENT_VERSION
def test_formula_mathml():
doc = DoclingDocument(name="Dummy")
equation = "\\frac{1}{x}"
doc.add_text(label=DocItemLabel.FORMULA, text=equation)
doc_html = doc.export_to_html(formula_to_mathml=True, html_head="")
file = "tests/data/docling_document/export/formula_mathml.html"
if GEN_TEST_DATA:
with open(file, mode="w", encoding="utf8") as f:
f.write(f"{doc_html}\n")
else:
with open(file, encoding="utf8") as f:
gt_html = f.read().rstrip()
assert doc_html == gt_html
def test_formula_with_missing_fallback():
doc = DoclingDocument(name="Dummy")
bbox = BoundingBox.from_tuple((1, 2, 3, 4), origin=CoordOrigin.BOTTOMLEFT)
prov = ProvenanceItem(page_no=1, bbox=bbox, charspan=(0, 2))
doc.add_text(label=DocItemLabel.FORMULA, text="", orig="(II.24) 2 Imar", prov=prov)
doc.export_to_html(formula_to_mathml=True, html_head="", image_mode=ImageRefMode.EMBEDDED)
expected = """<!DOCTYPE html>
<html lang="en">
<div class="formula-not-decoded">Formula not decoded</div>
</html>"""
assert '<div class="formula-not-decoded">Formula not decoded</div>' in expected
def test_docitem_get_image():
# Prepare the document
doc = DoclingDocument(name="Dummy")
page1_image = PILImage.new(mode="RGB", size=(200, 400), color=(0, 0, 0))
doc_item_image = PILImage.new(mode="RGB", size=(20, 40), color=(255, 0, 0))
page1_image.paste(doc_item_image, box=(20, 40))
doc.add_page( # With image
page_no=1,
size=Size(width=20, height=40),
image=ImageRef.from_pil(page1_image, dpi=72),
)
doc.add_page(page_no=2, size=Size(width=20, height=40), image=None) # Without image
# DocItem with no provenance
doc_item = DocItem(self_ref="#", label=DocItemLabel.TEXT, prov=[])
assert doc_item.get_image(doc=doc) is None
# DocItem on an invalid page
doc_item = DocItem(
self_ref="#",
label=DocItemLabel.TEXT,
prov=[ProvenanceItem(page_no=3, bbox=Mock(spec=BoundingBox), charspan=(1, 2))],
)
assert doc_item.get_image(doc=doc) is None
# DocItem on a page without page image
doc_item = DocItem(
self_ref="#",
label=DocItemLabel.TEXT,
prov=[ProvenanceItem(page_no=2, bbox=Mock(spec=BoundingBox), charspan=(1, 2))],
)
assert doc_item.get_image(doc=doc) is None
# DocItem on a page with valid page image
doc_item = DocItem(
self_ref="#",
label=DocItemLabel.TEXT,
prov=[ProvenanceItem(page_no=1, bbox=BoundingBox(l=2, t=4, r=4, b=8), charspan=(1, 2))],
)
returned_doc_item_image = doc_item.get_image(doc=doc)
assert returned_doc_item_image is not None and returned_doc_item_image.tobytes() == doc_item_image.tobytes()
def test_floatingitem_get_image():
# Prepare the document
doc = DoclingDocument(name="Dummy")
page1_image = PILImage.new(mode="RGB", size=(200, 400), color=(0, 0, 0))
floating_item_image = PILImage.new(mode="RGB", size=(20, 40), color=(255, 0, 0))
page1_image.paste(floating_item_image, box=(20, 40))
doc.add_page( # With image
page_no=1,
size=Size(width=20, height=40),
image=ImageRef.from_pil(page1_image, dpi=72),
)
doc.add_page(page_no=2, size=Size(width=20, height=40), image=None) # Without image
# FloatingItem with explicit image different from image based on provenance
new_image = PILImage.new(mode="RGB", size=(40, 80), color=(0, 255, 0))
floating_item = FloatingItem(
self_ref="#",
label=DocItemLabel.PICTURE,
prov=[ProvenanceItem(page_no=1, bbox=BoundingBox(l=2, t=4, r=6, b=12), charspan=(1, 2))],
image=ImageRef.from_pil(image=new_image, dpi=72),
)
retured_image = floating_item.get_image(doc=doc)
assert retured_image is not None and retured_image.tobytes() == new_image.tobytes()
# FloatingItem without explicit image and no provenance
floating_item = FloatingItem(self_ref="#", label=DocItemLabel.PICTURE, prov=[], image=None)
assert floating_item.get_image(doc=doc) is None
# FloatingItem without explicit image on invalid page
floating_item = FloatingItem(
self_ref="#",
label=DocItemLabel.PICTURE,
prov=[ProvenanceItem(page_no=3, bbox=Mock(spec=BoundingBox), charspan=(1, 2))],
image=None,
)
assert floating_item.get_image(doc=doc) is None
# FloatingItem without explicit image on a page without page image
floating_item = FloatingItem(
self_ref="#",
label=DocItemLabel.PICTURE,
prov=[ProvenanceItem(page_no=2, bbox=Mock(spec=BoundingBox), charspan=(1, 2))],
image=None,
)
assert floating_item.get_image(doc=doc) is None
# FloatingItem without explicit image on a page with page image
floating_item = FloatingItem(
self_ref="#",
label=DocItemLabel.PICTURE,
prov=[ProvenanceItem(page_no=1, bbox=BoundingBox(l=2, t=4, r=4, b=8), charspan=(1, 2))],
image=None,
)
retured_image = floating_item.get_image(doc=doc)
assert retured_image is not None and retured_image.tobytes() == floating_item_image.tobytes()
def test_save_pictures(sample_doc):
new_doc = sample_doc._with_pictures_refs(image_dir=Path("./tests/data/constructed_images/"), page_no=None)
img_paths = new_doc._list_images_on_disk()
assert len(img_paths) == 1, "len(img_paths)!=1"
def test_save_pictures_with_page():
# Given
doc = DoclingDocument(name="Dummy")
doc.add_page(page_no=1, size=Size(width=2000, height=4000), image=None)
doc.add_page(
page_no=2,
size=Size(width=2000, height=4000),
)
image = PILImage.new(mode="RGB", size=(200, 400), color=(0, 0, 0))
doc.add_picture(
image=ImageRef.from_pil(image=image, dpi=72),
prov=ProvenanceItem(
page_no=2,
bbox=BoundingBox(b=0, l=0, r=200, t=400, coord_origin=CoordOrigin.BOTTOMLEFT),
charspan=(1, 2),
),
)
# When
with_ref = doc._with_pictures_refs(image_dir=Path("./tests/data/constructed_images/"), page_no=1)
# Then
n_images = len(with_ref._list_images_on_disk())
assert n_images == 0
# When
with_ref = with_ref._with_pictures_refs(image_dir=Path("./tests/data/constructed_images/"), page_no=2)
n_images = len(with_ref._list_images_on_disk())
# Then
assert n_images == 1
def _normalise_string_wrt_filepaths(instr: str, paths: list[Path]):
for p in paths:
instr = instr.replace(str(p), str(p.name))
return instr
def _verify_saved_output(filename: str | Path, paths: list[Path]):
pred = ""
with open(filename, encoding="utf-8") as fr:
pred = fr.read()
pred = _normalise_string_wrt_filepaths(pred, paths=paths)
if GEN_TEST_DATA:
with open(str(filename) + ".gt", "w", encoding="utf-8") as fw:
fw.write(pred)
else:
gt = ""
with open(str(filename) + ".gt", encoding="utf-8") as fr:
gt = fr.read()
assert pred == gt, f"pred!=gt for {filename}"
def _gt_filename(filename: Path) -> Path:
return Path(str(filename) + ".gt")
def _verify_loaded_output(filename: Path, pred=None):
# gt = DoclingDocument.load_from_json(Path(str(filename) + ".gt"))
gt = DoclingDocument.load_from_json(_gt_filename(filename=filename))
pred = pred or DoclingDocument.load_from_json(Path(filename))
assert isinstance(pred, DoclingDocument)
assert pred.export_to_dict() == gt.export_to_dict(), f"pred.export_to_dict() != gt.export_to_dict() for {filename}"
assert pred == gt, f"pred!=gt for {filename}"
def test_save_to_disk(sample_doc):
test_dir = Path("./tests/data/doc")
image_dir = Path("constructed_images/") # will be relative to test_dir
doc_with_references = sample_doc._with_pictures_refs(
image_dir=(test_dir / image_dir),
page_no=None,
)
# paths will be different on different machines, so needs to be kept!
paths = doc_with_references._list_images_on_disk()
assert len(paths) == 1, "len(paths)!=1"
### MarkDown
filename: Path = test_dir / "constructed_doc.placeholder.md"
sample_doc.save_as_markdown(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.PLACEHOLDER)
_verify_saved_output(filename=filename, paths=paths)
filename = test_dir / "constructed_doc.embedded.md"
sample_doc.save_as_markdown(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.EMBEDDED)
_verify_saved_output(filename=filename, paths=paths)
filename = test_dir / "constructed_doc.referenced.md"
sample_doc.save_as_markdown(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.REFERENCED)
_verify_saved_output(filename=filename, paths=paths)
### HTML
filename = test_dir / "constructed_doc.placeholder.html"
sample_doc.save_as_html(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.PLACEHOLDER)
_verify_saved_output(filename=filename, paths=paths)
filename = test_dir / "constructed_doc.embedded.html"
sample_doc.save_as_html(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.EMBEDDED)
_verify_saved_output(filename=filename, paths=paths)
filename = test_dir / "constructed_doc.referenced.html"
sample_doc.save_as_html(filename=filename, artifacts_dir=image_dir, image_mode=ImageRefMode.REFERENCED)
_verify_saved_output(filename=filename, paths=paths)
### Document Tokens
filename = test_dir / "constructed_doc.dt"
sample_doc.save_as_doctags(filename=filename)
_verify_saved_output(filename=filename, paths=paths)
### JSON
filename = test_dir / "constructed_doc.embedded.json"
sample_doc.save_as_json(
filename=filename,
artifacts_dir=image_dir,
image_mode=ImageRefMode.EMBEDDED,
)
_verify_saved_output(filename=filename, paths=paths)
doc_emb_loaded = DoclingDocument.load_from_json(filename)
_verify_loaded_output(filename=filename, pred=doc_emb_loaded)
filename = test_dir / "constructed_doc.referenced.json"
sample_doc.save_as_json(
filename=filename,
artifacts_dir=image_dir,
image_mode=ImageRefMode.REFERENCED,
)
_verify_saved_output(filename=filename, paths=paths)
doc_ref_loaded = DoclingDocument.load_from_json(filename)
_verify_loaded_output(filename=filename, pred=doc_ref_loaded)
### YAML
filename = test_dir / "constructed_doc.embedded.yaml"
sample_doc.save_as_yaml(
filename=filename,
artifacts_dir=image_dir,
image_mode=ImageRefMode.EMBEDDED,
)
_verify_saved_output(filename=filename, paths=paths)
filename = test_dir / "constructed_doc.referenced.yaml"
sample_doc.save_as_yaml(
filename=filename,
artifacts_dir=image_dir,
image_mode=ImageRefMode.REFERENCED,
)
_verify_saved_output(filename=filename, paths=paths)
### WebVTT
# Note: Documents without TrackSource will result in empty WebVTT, but this is valid
filename = test_dir / "constructed_doc.vtt"
sample_doc.save_as_vtt(filename=filename)
_verify_saved_output(filename=filename, paths=paths)
assert True
def test_save_as_json_encoding_options(tmp_path: Path):
polish_text = "Należy wówczas zetrzeć warstwę"
doc = DoclingDocument(name="non_ascii")
doc.add_text(label=DocItemLabel.TEXT, text=polish_text)
unescaped_file = tmp_path / "unescaped.json"
doc.save_as_json(filename=unescaped_file, ensure_ascii=False)
unescaped = unescaped_file.read_text(encoding="utf-8")
assert polish_text in unescaped
assert "Nale\\u017cy" not in unescaped
# Without the new arguments the stdlib defaults still apply, so existing callers
# are unaffected.
escaped_file = tmp_path / "escaped.json"
doc.save_as_json(filename=escaped_file)
escaped = escaped_file.read_text(encoding="utf-8")
assert polish_text not in escaped
assert "Nale\\u017cy" in escaped
sorted_file = tmp_path / "sorted.json"
doc.save_as_json(filename=sorted_file, sort_keys=True)
reloaded = DoclingDocument.load_from_json(sorted_file)
assert reloaded.texts[0].text == polish_text
def test_document_stack_operations(sample_doc):
# _print(document=doc)
ref = RefItem(cref="#/texts/12")
success, stack = sample_doc._get_stack_of_refitem(ref=ref)
assert success
assert stack == [
2,
2,
2,
0,
2,
0,
0,
], f"stack==[2, 2, 2, 0, 2, 0, 0] for stack: {stack}"
def test_document_manipulation(sample_doc: DoclingDocument) -> None:
def _resolve(document: DoclingDocument, cref: str) -> NodeItem:
ref = RefItem(cref=cref)
return ref.resolve(doc=document)
def _verify(
filename: Path,
document: DoclingDocument,
generate: bool = False,
):
if generate or (not os.path.exists(_gt_filename(filename=filename))):
document.save_as_json(
filename=_gt_filename(filename=filename),
artifacts_dir=image_dir,
image_mode=ImageRefMode.EMBEDDED,
)
# test if the document is still model-validating
DoclingDocument.load_from_json(filename=_gt_filename(filename=filename))
# test if the document is the same as the stored GT
_verify_loaded_output(filename=filename, pred=DoclingDocument.model_validate(document))
image_dir = Path("./tests/data/doc/constructed_images/")
text_item_1 = ListItem(
self_ref="#",
text="new list item (before)",
orig="new list item (before)",
)
text_item_2 = ListItem(
self_ref="#",
text="new list item (after)",
orig="new list item (after)",
)
node = _resolve(document=sample_doc, cref="#/texts/10")
sample_doc.insert_item_before_sibling(new_item=text_item_1, sibling=node)
sample_doc.insert_item_after_sibling(new_item=text_item_2, sibling=node)
filename = Path("tests/data/doc/constructed_doc.inserted_text.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
items = [_resolve(document=sample_doc, cref="#/texts/10")]
sample_doc.delete_items(node_items=items)
filename = Path("tests/data/doc/constructed_doc.deleted_text.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
items = [_resolve(document=sample_doc, cref="#/groups/1")]
sample_doc.delete_items(node_items=items)
filename = Path("tests/data/doc/constructed_doc.deleted_group.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
items = [_resolve(document=sample_doc, cref="#/pictures/1")]
sample_doc.delete_items(node_items=items)
filename = Path("tests/data/doc/constructed_doc.deleted_picture.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
text_item_3 = TextItem(
self_ref="#",
text="child text appended at body",
orig="child text appended at body",
label=DocItemLabel.TEXT,
)
sample_doc.append_child_item(child=text_item_3)
text_item_4 = ListItem(
self_ref="#",
text="child text appended at body",
orig="child text appended at body",
label=DocItemLabel.LIST_ITEM,
)
parent = _resolve(document=sample_doc, cref="#/groups/11")
sample_doc.append_child_item(child=text_item_4, parent=parent)
# try to add a sibling to the root:
with pytest.raises(ValueError):
sample_doc.insert_item_before_sibling(
new_item=TextItem(
self_ref="#",
label=DocItemLabel.TEXT,
text="foo",
orig="foo",
),
sibling=sample_doc.body,
)
# try to append a child with children of its own:
with pytest.raises(ValueError):
sample_doc.append_child_item(
child=TextItem(
self_ref="#",
label=DocItemLabel.TEXT,
text="foo",
orig="foo",
children=[_resolve(document=deepcopy(sample_doc), cref=text_item_4.self_ref).get_ref()],
),
parent=sample_doc.body,
)
filename = Path("tests/data/doc/constructed_doc.appended_child.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
text_item_5 = TextItem(
self_ref="#",
text="new child",
orig="new child",
label=DocItemLabel.TEXT,
)
sample_doc.replace_item(old_item=text_item_3, new_item=text_item_5)
filename = Path("tests/data/doc/constructed_doc.replaced_item.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test insert_* methods with mixed insertion before/after sibling
node = _resolve(document=sample_doc, cref="#/texts/45")
last_node = sample_doc.insert_list_group(sibling=node, name="Inserted List Group", after=True)
group_node = sample_doc.insert_inline_group(sibling=node, name="Inserted Inline Group", after=False)
sample_doc.insert_group(
sibling=node,
label=GroupLabel.LIST,
name="Inserted Group w/ LIST Label",
after=True,
)
sample_doc.insert_group(
sibling=node,
label=GroupLabel.ORDERED_LIST,
name="Inserted Group w/ ORDERED_LIST Label",
after=False,
)
sample_doc.insert_group(
sibling=node,
label=GroupLabel.INLINE,
name="Inserted Group w/ INLINE Label",
after=True,
)
sample_doc.insert_group(
sibling=node,
label=GroupLabel.UNSPECIFIED,
name="Inserted Group w/ UNSPECIFIED Label",
after=False,
)
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.TITLE,
text="Inserted Text w/ TITLE Label",
after=True,
)
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.SECTION_HEADER,
text="Inserted Text w/ SECTION_HEADER Label",
after=False,
)
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.CODE,
text="Inserted Text w/ CODE Label",
after=True,
)
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.FORMULA,
text="Inserted Text w/ FORMULA Label",
after=False,
)
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.TEXT,
text="Inserted Text w/ TEXT Label",
after=True,
)
num_rows = 3
num_cols = 3
table_cells = []
for i in range(num_rows):
for j in range(num_cols):
table_cells.append(
TableCell(
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
text=str(i * num_rows + j),
)
)
table_data = TableData(table_cells=table_cells, num_rows=num_rows, num_cols=num_cols)
sample_doc.insert_table(sibling=node, data=table_data, after=False)
size = (64, 64)
img = PILImage.new("RGB", size, "black")
sample_doc.insert_picture(sibling=node, image=ImageRef.from_pil(image=img, dpi=72), after=True)
sample_doc.insert_title(sibling=node, text="Inserted Title", after=False)
sample_doc.insert_code(sibling=node, text="Inserted Code", after=True)
sample_doc.insert_formula(sibling=node, text="Inserted Formula", after=False)
sample_doc.insert_heading(sibling=node, text="Inserted Heading", after=True)
graph = GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=0,
text="number",
orig="#",
),
GraphCell(
label=GraphCellLabel.VALUE,
cell_id=1,
text="1",
orig="1",
),
],
links=[
GraphLink(
label=GraphLinkLabel.TO_VALUE,
source_cell_id=0,
target_cell_id=1,
),
GraphLink(label=GraphLinkLabel.TO_KEY, source_cell_id=1, target_cell_id=0),
],
)
sample_doc.insert_key_values(sibling=node, graph=graph, after=False)
sample_doc.insert_form(sibling=node, graph=graph, after=True)
filename = Path("tests/data/doc/constructed_doc.inserted_items.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test the handling of list items in insert_* methods, both with and without parent groups
with pytest.warns(DeprecationWarning, match="ListItem parent must be a ListGroup"):
li_sibling = sample_doc.insert_list_item(sibling=node, text="Inserted List Item, Incorrect Parent", after=False)
sample_doc.insert_list_item(sibling=li_sibling, text="Inserted List Item, Correct Parent", after=True)
sample_doc.insert_text(
sibling=li_sibling,
label=DocItemLabel.LIST_ITEM,
text="Inserted Text with LIST_ITEM Label, Correct Parent",
after=False,
)
with pytest.warns(DeprecationWarning, match="ListItem parent must be a ListGroup"):
sample_doc.insert_text(
sibling=node,
label=DocItemLabel.LIST_ITEM,
text="Inserted Text with LIST_ITEM Label, Incorrect Parent",
after=True,
)
filename = Path("tests/data/doc/constructed_doc.inserted_list_items.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test the bulk addition of node items
text_item_6 = TextItem(
self_ref="#",
text="Bulk Addition 1",
orig="Bulk Addition 1",
label=DocItemLabel.TEXT,
)
text_item_7 = TextItem(
self_ref="#",
text="Bulk Addition 2",
orig="Bulk Addition 2",
label=DocItemLabel.TEXT,
)
sample_doc.add_node_items(node_items=[text_item_6, text_item_7], doc=sample_doc, parent=group_node)
filename = Path("tests/data/doc/constructed_doc.bulk_item_addition.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test the bulk insertion of node items
text_item_8 = TextItem(
self_ref="#",
text="Bulk Insertion 1",
orig="Bulk Insertion 1",
label=DocItemLabel.TEXT,
)
text_item_9 = TextItem(
self_ref="#",
text="Bulk Insertion 2",
orig="Bulk Insertion 2",
label=DocItemLabel.TEXT,
)
sample_doc.insert_node_items(sibling=node, node_items=[text_item_8, text_item_9], doc=sample_doc, after=False)
filename = Path("tests/data/doc/constructed_doc.bulk_item_insertion.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test table data manipulation methods
table_data.add_row(["*"] * 3)
table_data.add_rows([["a", "b", "c"], ["d", "e", "f"]])
table_data.insert_row(1, ["*"] * 3)
table_data.insert_rows(1, [["a", "b", "c"], ["d", "e", "f"]], after=True)
table_data.pop_row()
table_data.remove_row(3)
table_data.remove_rows([1, 2, 5])
# Try to remove a nonexistent row
with pytest.raises(IndexError):
table_data.remove_row(100)
filename = Path("tests/data/doc/constructed_doc.manipulated_table.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
# Test range manipulation methods
# Try to extract a range where start is after end
with pytest.raises(ValueError):
extracted_doc = sample_doc.extract_items_range(start=node, end=group_node)
# Try to extract a range where start and end have different parents
with pytest.raises(ValueError):
extracted_doc = sample_doc.extract_items_range(start=li_sibling, end=node)
extracted_doc = sample_doc.extract_items_range(start=group_node, end=node, end_inclusive=False, delete=True)
filename = Path("tests/data/doc/constructed_doc.extracted_with_deletion.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
sample_doc.add_document(doc=extracted_doc, parent=last_node)
filename = Path("tests/data/doc/constructed_doc.added_extracted_doc.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
sample_doc.insert_document(doc=extracted_doc, sibling=last_node, after=False)
filename = Path("tests/data/doc/constructed_doc.inserted_extracted_doc.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
sample_doc.delete_items_range(start=node, end=last_node, start_inclusive=False)
filename = Path("tests/data/doc/constructed_doc.deleted_items_range.json")
_verify(filename=filename, document=sample_doc, generate=GEN_TEST_DATA)
def test_table_remove_rows_ignores_duplicate_indices() -> None:
table_cells = [
TableCell(
text=f"r{i}c{j}",
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
)
for i in range(4)
for j in range(2)
]
data = TableData(num_rows=4, num_cols=2, table_cells=table_cells)
removed = data.remove_rows([1, 1])
assert [[c.text for c in row] for row in removed] == [["r1c0", "r1c1"]]
assert data.num_rows == 3
assert [c.text for row in data.grid for c in row] == ["r0c0", "r0c1", "r2c0", "r2c1", "r3c0", "r3c1"]
def test_misplaced_list_items():
filename = Path("tests/data/doc/misplaced_list_items.yaml")
doc = DoclingDocument.load_from_yaml(filename)
dt_pred = doc.export_to_doctags()
_verify_regression_test(dt_pred, filename=str(filename), ext="dt")
exp_file = filename.parent / f"{filename.stem}.out.yaml"
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
else:
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
doc._normalize_references()
exp_file = filename.parent / f"{filename.stem}.norm.out.yaml"
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
else:
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
def test_delete_items_renumbers_orphaned_items():
# The caption is referenced by the picture but is not one of its children, so it is not
# reachable from body via children. Deleting an item before it must still renumber it.
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "orphaned_items",
"body": {
"self_ref": "#/body",
"children": [{"$ref": "#/texts/0"}, {"$ref": "#/pictures/0"}],
"name": "_root_",
"label": "unspecified",
},
"texts": [
{
"self_ref": "#/texts/0",
"parent": {"$ref": "#/body"},
"label": "list_item",
"orig": "item",
"text": "item",
"enumerated": False,
"marker": "-",
},
{
"self_ref": "#/texts/1",
"parent": {"$ref": "#/pictures/0"},
"label": "caption",
"orig": "caption",
"text": "caption",
},
],
"pictures": [
{
"self_ref": "#/pictures/0",
"parent": {"$ref": "#/body"},
"label": "picture",
"captions": [{"$ref": "#/texts/1"}],
}
],
}
doc = DoclingDocument.model_validate(doc_dict)
# loading moves the misplaced list item into a new list, without renumbering anything
assert [it.text for it in doc.texts] == ["item", "caption"]
assert isinstance(doc.texts[0].parent.resolve(doc), ListGroup)
# the caption is an orphan, which the rules detect
with pytest.raises(ValueError, match="#/texts/1 is not a child of its parent #/pictures/0"):
doc._validate_rules()
with pytest.warns(UserWarning, match="#/texts/1 is not a child of its parent #/pictures/0"):
doc._validate_rules(raise_on_error=False)
# deleting the list (and so the list item) renumbers the orphaned caption too
doc.delete_items(node_items=[doc.groups[0]])
doc._validate_unique_refs()
for item in doc.texts:
assert item.get_ref().resolve(doc) is item
assert [it.text for it in doc.texts] == ["caption"]
assert doc.pictures[0].captions[0].resolve(doc).text == "caption"
assert doc.texts[0].parent.cref == "#/pictures/0"
# a dangling parent is reported, not crashed on
doc.texts[0].parent = RefItem(cref="#/pictures/1")
with pytest.raises(ValueError, match="#/texts/0 has non-existent parent #/pictures/1"):
doc._validate_rules()
def test_validate_rules_key_value_and_form_items():
doc = DoclingDocument(name="")
doc.add_key_values(graph=GraphData(cells=[], links=[]))
doc.add_form(graph=GraphData(cells=[], links=[]))
with pytest.raises(ValueError, match="Key-value item #/key_value_items/0 is to be migrated to a field region"):
doc._validate_rules()
with pytest.warns(UserWarning) as record:
doc._validate_rules(raise_on_error=False)
messages = [str(w.message) for w in record]
assert "Key-value item #/key_value_items/0 is to be migrated to a field region" in messages
assert "Form item #/form_items/0 is to be migrated to a field region" in messages
doc._migrate_to_field_regions()
doc._validate_rules()
def test_field_item_with_several_keys_is_invalid():
doc = DoclingDocument(name="")
region = doc.add_field_region()
item = doc.add_field_item(parent=region)
doc.add_field_key(text="a", parent=item)
doc.add_field_value(text="b", parent=item)
doc.add_field_value(text="c", parent=item) # several values are fine
DoclingDocument.model_validate(doc.model_dump(mode="json"))
doc.add_field_key(text="d", parent=item) # building through the API does not validate
with pytest.raises(ValueError, match="Field item #/field_items/0 has 2 keys, at most one is allowed"):
DoclingDocument.model_validate(doc.model_dump(mode="json"))
def test_graph_cell_item_refs_follow_renumbering():
def build() -> tuple[DoclingDocument, KeyValueItem]:
doc = DoclingDocument(name="")
a = doc.add_text(label=DocItemLabel.TEXT, text="a")
b = doc.add_text(label=DocItemLabel.TEXT, text="b")
cells = [
GraphCell(label=GraphCellLabel.KEY, cell_id=0, text="b", orig="b", item_ref=b.get_ref()),
GraphCell(label=GraphCellLabel.VALUE, cell_id=1, text="a", orig="a", item_ref=a.get_ref()),
]
return doc, doc.add_key_values(graph=GraphData(cells=cells, links=[]))
# deletion: refs are renumbered, refs to deleted items dropped
doc, kv = build()
doc.delete_items(node_items=[doc.texts[0]])
assert [c.item_ref.cref if c.item_ref else None for c in kv.graph.cells] == ["#/texts/0", None]
assert kv.graph.cells[0].item_ref.resolve(doc).text == "b"
# normalization: refs follow the reordering
doc, _ = build()
doc.body.children = [doc.body.children[1], doc.body.children[0], doc.body.children[2]] # b, a, kv
doc._normalize_references()
assert [t.text for t in doc.texts] == ["b", "a"]
cells = doc.key_value_items[0].graph.cells
assert [c.item_ref.resolve(doc).text for c in cells] == ["b", "a"]
def test_delete_items_renumbers_furniture():
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "legacy_furniture",
"furniture": {
"self_ref": "#/furniture",
"children": [{"$ref": "#/texts/1"}],
"content_layer": "furniture",
"name": "_root_",
"label": "unspecified",
},
"body": {"self_ref": "#/body", "children": [{"$ref": "#/texts/0"}], "name": "_root_", "label": "unspecified"},
"texts": [
{"self_ref": "#/texts/0", "parent": {"$ref": "#/body"}, "label": "text", "orig": "text", "text": "text"},
{
"self_ref": "#/texts/1",
"parent": {"$ref": "#/furniture"},
"content_layer": "furniture",
"label": "page_header",
"orig": "header",
"text": "header",
},
],
}
doc = DoclingDocument.model_validate(doc_dict)
doc.delete_items(node_items=[doc.texts[0]])
assert doc.texts[0].self_ref == "#/texts/0"
with warnings.catch_warnings():
warnings.simplefilter("ignore", DeprecationWarning)
assert [ref.resolve(doc).text for ref in doc.furniture.children] == ["header"]
def test_migrate_furniture_to_body():
def text(idx: int, parent: str, label: str, top: float | None = None) -> dict:
item: dict = {
"self_ref": f"#/texts/{idx}",
"parent": {"$ref": parent},
"label": label,
"orig": label,
"text": label,
}
if top is not None:
item["prov"] = [
{
"page_no": 1,
"bbox": {"l": 10, "t": top, "r": 90, "b": top + 5, "coord_origin": "TOPLEFT"},
"charspan": [0, len(label)],
}
]
return item
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "legacy_furniture",
"furniture": {
"self_ref": "#/furniture",
"children": [{"$ref": "#/texts/1"}, {"$ref": "#/texts/2"}],
"content_layer": "furniture",
"name": "_root_",
"label": "unspecified",
},
"body": {
"self_ref": "#/body",
"children": [{"$ref": "#/texts/0"}],
"name": "_root_",
"label": "unspecified",
},
"texts": [
text(0, "#/body", "text", top=50),
text(1, "#/furniture", "page_footer", top=90),
text(2, "#/furniture", "page_header", top=5),
],
"pages": {"1": {"page_no": 1, "size": {"width": 100, "height": 100}}},
}
doc = DoclingDocument.model_validate(doc_dict)
assert [ref.cref for ref in doc.body.children] == ["#/texts/0"] # not migrated on load
assert doc._migrate_furniture_to_body() == 2
with warnings.catch_warnings():
warnings.simplefilter("ignore", DeprecationWarning)
assert doc.furniture.children == []
assert [ref.resolve(doc).text for ref in doc.body.children] == ["page_header", "text", "page_footer"]
assert all(it.parent.cref == "#/body" for it in doc.texts)
assert [it.content_layer for it in doc.texts] == [
ContentLayer.BODY,
ContentLayer.FURNITURE,
ContentLayer.FURNITURE,
]
def test_misplaced_list_items_keep_children_and_refs():
# a list item outside a list, with a picture as child and a graph cell pointing to it
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "misplaced_with_children",
"body": {
"self_ref": "#/body",
"children": [{"$ref": "#/texts/0"}, {"$ref": "#/texts/1"}, {"$ref": "#/key_value_items/0"}],
"name": "_root_",
"label": "unspecified",
},
"texts": [
{
"self_ref": "#/texts/0",
"parent": {"$ref": "#/body"},
"children": [{"$ref": "#/pictures/0"}],
"label": "list_item",
"orig": "item",
"text": "item",
"enumerated": False,
"marker": "-",
},
{"self_ref": "#/texts/1", "parent": {"$ref": "#/body"}, "label": "text", "orig": "text", "text": "text"},
],
"pictures": [{"self_ref": "#/pictures/0", "parent": {"$ref": "#/texts/0"}, "label": "picture"}],
"key_value_items": [
{
"self_ref": "#/key_value_items/0",
"parent": {"$ref": "#/body"},
"label": "key_value_region",
"graph": {
"cells": [
{
"label": "key",
"cell_id": 0,
"text": "item",
"orig": "item",
"item_ref": {"$ref": "#/texts/0"},
}
],
"links": [],
},
}
],
}
doc = DoclingDocument.model_validate(doc_dict)
# the list item is moved into a new list, keeping its ref, its child and the refs to it
item = doc.texts[0]
assert item.text == "item" and isinstance(item.parent.resolve(doc), ListGroup)
assert [ref.cref for ref in doc.body.children] == ["#/groups/0", "#/texts/1", "#/key_value_items/0"]
assert len(doc.pictures) == 1 and doc.pictures[0].parent == item.get_ref()
assert [ref.cref for ref in item.children] == ["#/pictures/0"]
assert doc.key_value_items[0].graph.cells[0].item_ref.resolve(doc) is item
def test_moving_within_same_parent():
doc = DoclingDocument(name="")
doc.add_text(label=DocItemLabel.TEXT, text="bar")
foo = doc.add_text(label=DocItemLabel.TEXT, text="foo")
assert foo.parent is not None
doc._move_subtree(old_subroot=foo, new_subroot=foo.parent.resolve(doc), pos=0)
assert [it.text for it, _ in doc.iterate_items() if isinstance(it, TextItem)] == ["foo", "bar"]
def test_export_with_precision():
doc = DoclingDocument.load_from_yaml(filename="tests/data/doc/dummy_doc_2.yaml")
act_data = doc.export_to_dict(coord_precision=2, confid_precision=1)
exp_file = Path("tests/data/doc/dummy_doc_2_prec.yaml")
if GEN_TEST_DATA:
with open(exp_file, "w", encoding="utf-8") as f:
yaml.dump(act_data, f, default_flow_style=False)
else:
with open(exp_file, encoding="utf-8") as f:
exp_data = yaml.load(f, Loader=yaml.SafeLoader)
assert act_data == exp_data
def test_concatenate():
files = [
"tests/data/doc/2501.17887v1.json",
"tests/data/doc/constructed_doc.embedded.json",
"tests/data/doc/2311.18481v1.json",
]
docs = [DoclingDocument.load_from_json(filename=f) for f in files]
doc = DoclingDocument.concatenate(docs=docs)
html_data = doc.export_to_html(image_mode=ImageRefMode.EMBEDDED, split_page_view=True)
exp_json_file = Path("tests/data/doc/concatenated.json")
exp_html_file = exp_json_file.with_suffix(".html")
if GEN_TEST_DATA:
doc.save_as_json(exp_json_file)
with open(exp_html_file, "w", encoding="utf-8") as f:
f.write(html_data)
else:
exp_doc = DoclingDocument.load_from_json(exp_json_file)
assert doc == exp_doc
with open(exp_html_file, encoding="utf-8") as f:
exp_html_data = f.read()
assert html_data == exp_html_data
def test_concatenate_shifts_graph_cell_pages_for_keyvalue_and_form():
def _make_graph(page_no: int, value_text: str) -> GraphData:
return GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=1,
text="k",
orig="k",
prov=ProvenanceItem(
page_no=page_no,
charspan=(0, 1),
bbox=BoundingBox(
l=10,
t=40,
r=30,
b=10,
coord_origin=CoordOrigin.BOTTOMLEFT,
),
),
),
GraphCell(
label=GraphCellLabel.VALUE,
cell_id=2,
text=value_text,
orig=value_text,
prov=ProvenanceItem(
page_no=page_no,
charspan=(0, len(value_text)),
bbox=BoundingBox(
l=35,
t=40,
r=80,
b=10,
coord_origin=CoordOrigin.BOTTOMLEFT,
),
),
),
],
links=[
GraphLink(
label=GraphLinkLabel.TO_VALUE,
source_cell_id=1,
target_cell_id=2,
)
],
)
def _make_doc(name: str, value_text: str) -> DoclingDocument:
doc = DoclingDocument(name=name)
doc.add_page(page_no=1, size=Size(width=100, height=100))
doc.add_key_values(graph=_make_graph(page_no=1, value_text=value_text))
doc.add_form(graph=_make_graph(page_no=1, value_text=f"{value_text}-form"))
return doc
doc1 = _make_doc(name="doc1", value_text="v1")
doc2 = _make_doc(name="doc2", value_text="v2")
merged = DoclingDocument.concatenate(docs=[doc1, doc2])
assert sorted(merged.pages.keys()) == [1, 2]
assert len(merged.key_value_items) == 2
assert len(merged.form_items) == 2
kv_item_pages = [
sorted({cell.prov.page_no for cell in item.graph.cells if cell.prov}) for item in merged.key_value_items
]
form_item_pages = [
sorted({cell.prov.page_no for cell in item.graph.cells if cell.prov}) for item in merged.form_items
]
assert kv_item_pages == [[1], [2]]
assert form_item_pages == [[1], [2]]
def test_concatenate_squeezes_successive_duplicate_names():
def _make_doc(name: str) -> DoclingDocument:
doc = DoclingDocument(name=name)
doc.add_page(page_no=1, size=Size(width=100, height=100))
return doc
assert DoclingDocument.concatenate([_make_doc("a"), _make_doc("a"), _make_doc("a")]).name == "a"
assert (
DoclingDocument.concatenate(
[_make_doc("a"), _make_doc("a"), _make_doc("b"), _make_doc("b"), _make_doc("b"), _make_doc("a")]
).name
== "a + b + a"
)
def test_export_markdown_compact_tables():
"""Test compact_tables parameter for markdown export."""
doc = DoclingDocument(name="Compact Table Test")
table = doc.add_table(data=TableData(num_rows=3, num_cols=3))
# Add cells with varying lengths to demonstrate padding difference
cells_data = [
["item", "qty", "description"],
["spam", "42", "A canned meat product"],
["eggs", "451", "Fresh farm eggs"],
]
for row in range(3):
for col in range(3):
doc.add_table_cell(
table,
TableCell(
start_row_offset_idx=row,
end_row_offset_idx=row + 1,
start_col_offset_idx=col,
end_col_offset_idx=col + 1,
text=cells_data[row][col],
),
)
# Test default (padded) format
md_padded = doc.export_to_markdown()
assert "| item |" in md_padded # Has padding
assert "|--------|" in md_padded # Long separator
# Test compact format
md_compact = doc.export_to_markdown(compact_tables=True)
assert "| item |" in md_compact # No padding
assert "| - |" in md_compact # Minimal separator
# Verify compact is shorter
assert len(md_compact) < len(md_padded)
def test_export_traverse_pictures_ocr_scanned_pdf():
"""Test that OCR text nested under a PictureItem is included when traverse_pictures=True."""
doc = DoclingDocument(name="Scanned Doc")
picture = doc.add_picture()
ocr_item_1 = TextItem(
self_ref=f"#/texts/{len(doc.texts)}",
parent=RefItem(cref=picture.self_ref),
label=DocItemLabel.TEXT,
text="SOCIAL SECURITY",
orig="SOCIAL SECURITY",
)
doc.texts.append(ocr_item_1)
picture.children.append(RefItem(cref=ocr_item_1.self_ref))
ocr_item_2 = TextItem(
self_ref=f"#/texts/{len(doc.texts)}",
parent=RefItem(cref=picture.self_ref),
label=DocItemLabel.TEXT,
text="000-00-0000",
orig="000-00-0000",
)
doc.texts.append(ocr_item_2)
picture.children.append(RefItem(cref=ocr_item_2.self_ref))
result_no_traverse_md = doc.export_to_markdown()
result_no_traverse_text = doc.export_to_text()
assert "SOCIAL SECURITY" not in result_no_traverse_md
assert "000-00-0000" not in result_no_traverse_md
assert "<!-- image -->" in result_no_traverse_md
assert "SOCIAL SECURITY" not in result_no_traverse_text
assert "000-00-0000" not in result_no_traverse_text
result_with_traverse_md = doc.export_to_markdown(traverse_pictures=True)
result_with_traverse_text = doc.export_to_text(traverse_pictures=True)
assert "SOCIAL SECURITY" in result_with_traverse_md
assert "000-00-0000" in result_with_traverse_md
assert "<!-- image -->" in result_with_traverse_md
assert "SOCIAL SECURITY" in result_with_traverse_text
assert "000-00-0000" in result_with_traverse_text
def test_list_group_with_list_items():
good_doc = DoclingDocument(name="")
l1 = good_doc.add_list_group()
good_doc.add_list_item(text="ListItem 1", parent=l1)
good_doc.add_list_item(text="ListItem 2", parent=l1)
good_doc._validate_rules()
def test_list_group_with_non_list_items():
bad_doc = DoclingDocument(name="")
l1 = bad_doc.add_list_group()
bad_doc.add_list_item(text="ListItem 1", parent=l1)
bad_doc.add_text(text="non-ListItem in ListGroup", label=DocItemLabel.TEXT, parent=l1)
with pytest.raises(ValueError):
bad_doc._validate_rules()
def test_list_item_outside_list_group():
def unsafe_add_list_item(
doc: DoclingDocument,
text: str,
enumerated: bool = False,
marker: str | None = None,
orig: str | None = None,
prov: ProvenanceItem | None = None,
parent: NodeItem | None = None,
content_layer: ContentLayer | None = None,
formatting: Formatting | None = None,
hyperlink: AnyUrl | Path | None = None,
):
if not parent:
parent = doc.body
if not orig:
orig = text
text_index = len(doc.texts)
cref = f"#/texts/{text_index}"
list_item = ListItem(
text=text,
orig=orig,
self_ref=cref,
parent=parent.get_ref(),
enumerated=enumerated,
marker=marker or "",
formatting=formatting,
hyperlink=hyperlink,
)
if prov:
list_item.prov.append(prov)
if content_layer:
list_item.content_layer = content_layer
doc.texts.append(list_item)
parent.children.append(RefItem(cref=cref))
return list_item
bad_doc = DoclingDocument(name="")
unsafe_add_list_item(doc=bad_doc, text="ListItem outside ListGroup")
with pytest.raises(ValueError):
bad_doc._validate_rules()
def test_list_item_inside_list_group():
doc = DoclingDocument(name="")
l1 = doc.add_list_group()
doc.add_list_item(text="ListItem inside ListGroup", parent=l1)
doc._validate_rules()
def test_group_with_children():
good_doc = DoclingDocument(name="")
grp = good_doc.add_group()
good_doc.add_text(text="Text in group", label=DocItemLabel.TEXT, parent=grp)
good_doc._validate_rules()
def test_group_without_children():
bad_doc = DoclingDocument(name="")
bad_doc.add_group()
with pytest.raises(ValueError):
bad_doc._validate_rules()
def test_rich_tables(rich_table_doc):
exp_file = Path("tests/data/doc/rich_table.out.yaml")
if GEN_TEST_DATA:
rich_table_doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert rich_table_doc == exp_doc
def test_doc_manipulation_with_rich_tables(rich_table_doc):
rich_table_doc.delete_items(node_items=[rich_table_doc.texts[0]])
exp_file = Path("tests/data/doc/rich_table_post_text_del.out.yaml")
if GEN_TEST_DATA:
rich_table_doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert rich_table_doc == exp_doc
def test_rich_table_cell_text_reuses_given_serializer(rich_table_doc, monkeypatch):
"""A rich cell must not build a fallback serializer when the caller passes one.
Building a serializer validates the whole document, so doing it for every rich
cell made serializing documents with many rich tables quadratic in the document
size (docling-project/docling#3222).
"""
serializer = MarkdownDocSerializer(doc=rich_table_doc)
rich_cells = [
cell for table in rich_table_doc.tables for cell in table.data.table_cells if isinstance(cell, RichTableCell)
]
expected = [serializer.serialize(item=cell.ref.resolve(doc=rich_table_doc)).text for cell in rich_cells]
def fail_on_fallback(self, *args, **kwargs):
raise AssertionError("RichTableCell built its own serializer although one was passed")
monkeypatch.setattr(MarkdownDocSerializer, "__init__", fail_on_fallback)
assert rich_cells
assert [cell._get_text(doc=rich_table_doc, doc_serializer=serializer) for cell in rich_cells] == expected
def test_invalid_rich_table_doc():
doc = DoclingDocument(name="")
table_item = doc.add_table(data=TableData(num_rows=2, num_cols=2))
rich_item = doc.add_text(
text="rich item",
label=DocItemLabel.TEXT,
parent=doc.body, # not the table item
)
for i in range(table_item.data.num_rows):
for j in range(table_item.data.num_cols):
if i == 1 and j == 1:
table_cell = RichTableCell(
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
ref=rich_item.get_ref(),
)
# ensure add_table_cell() raises:
with pytest.raises(ValueError):
doc.add_table_cell(table_item=table_item, cell=table_cell)
else:
table_cell = TableCell(
text=f"cell {i},{j}",
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
)
# discouraged but technically possible:
table_item.data.table_cells.append(table_cell)
# ensure validate_document() raises:
with pytest.raises(ValueError):
DoclingDocument.validate_document(doc)
def test_invalid_single_linked_rich_table_doc():
doc = DoclingDocument(name="")
table_item = doc.add_table(data=TableData(num_rows=2, num_cols=2))
rich_item = doc.add_text(
text="rich item",
label=DocItemLabel.TEXT,
parent=table_item,
)
for i in range(table_item.data.num_rows):
for j in range(table_item.data.num_cols):
if i == 1 and j == 1:
table_cell = RichTableCell(
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
ref=rich_item.get_ref(),
)
else:
table_cell = TableCell(
text=f"cell {i},{j}",
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
)
doc.add_table_cell(table_item=table_item, cell=table_cell)
# delete child reference from table item
del table_item.children[0]
# ensure validate_document() raises:
with pytest.raises(ValueError):
DoclingDocument.validate_document(doc)
def test_rich_table_item_insertion_normalization():
doc = DoclingDocument(name="")
doc.add_text(label=DocItemLabel.TITLE, text="Rich tables")
table_item = doc.add_table(
data=TableData(
num_rows=4,
num_cols=2,
),
)
rich_item = doc.add_text(
parent=table_item,
text="text in italic",
label=DocItemLabel.TEXT,
formatting=Formatting(italic=True),
)
for i in range(table_item.data.num_rows):
for j in range(table_item.data.num_cols):
if i == 1 and j == 1:
cell = RichTableCell(
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
ref=rich_item.get_ref(),
)
else:
cell = TableCell(
start_row_offset_idx=i,
end_row_offset_idx=i + 1,
start_col_offset_idx=j,
end_col_offset_idx=j + 1,
text=f"cell {i},{j}",
)
doc.add_table_cell(table_item=table_item, cell=cell)
# state before insert:
exp_file = Path("tests/data/doc/rich_table_item_ins_norm_1.out.yaml")
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
doc.insert_item_before_sibling(
new_item=TextItem(
self_ref="#",
text="text before",
orig="text before",
label=DocItemLabel.TEXT,
),
sibling=table_item,
)
# state after insert (prior to normalization):
exp_file = Path("tests/data/doc/rich_table_item_ins_norm_2.out.yaml")
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
doc._normalize_references()
# state after insert (after normalization):
exp_file = Path("tests/data/doc/rich_table_item_ins_norm_3.out.yaml")
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
def test_filter_pages():
src = Path("./tests/data/doc/2408.09869v3_enriched.json")
orig_doc = DoclingDocument.load_from_json(src)
doc = orig_doc.filter(page_nrs={2, 3, 5})
html_data = doc.export_to_html(image_mode=ImageRefMode.EMBEDDED, split_page_view=True)
exp_json_file = src.with_name(f"{src.stem}_p2_p3_p5.gt.json")
exp_html_file = exp_json_file.with_suffix(".html")
if GEN_TEST_DATA:
doc.save_as_json(exp_json_file)
with open(exp_html_file, "w", encoding="utf-8") as f:
f.write(html_data)
else:
exp_doc = DoclingDocument.load_from_json(exp_json_file)
assert doc == exp_doc
with open(exp_html_file, encoding="utf-8") as f:
exp_html_data = f.read()
assert html_data == exp_html_data
def _create_doc_for_filtering():
doc = DoclingDocument(
name="",
pages={i: PageItem(page_no=i, size=Size(width=100, height=100), image=None) for i in range(1, 3)},
)
p1_text = doc.add_text(
text="Text 1",
parent=doc.body,
label=DocItemLabel.TEXT,
prov=ProvenanceItem(page_no=1, bbox=BoundingBox(l=0, t=0, r=100, b=100), charspan=(0, 1)),
)
doc.add_group(parent=p1_text)
doc.add_text(
text="Text 2",
parent=doc.body,
label=DocItemLabel.TEXT,
prov=ProvenanceItem(page_no=2, bbox=BoundingBox(l=0, t=0, r=100, b=100), charspan=(0, 1)),
)
return doc
def test_filter_pages_filtered_out_parent():
doc = _create_doc_for_filtering()
with pytest.warns(
UserWarning,
match="Parent #/texts/0 not found in indexed nodes, using ancestor #/body instead",
):
doc.filter(page_nrs={2})
def test_filter_invalid_pages():
doc = _create_doc_for_filtering()
with pytest.raises(
ValueError,
match=re.escape("The following page numbers are not present in the document: {3}"),
):
doc.filter(page_nrs={3})
def test_validate_rules():
doc = _create_doc_for_filtering()
message = "Group #/groups/0 has no children"
with pytest.raises(ValueError, match=message):
doc._validate_rules()
with pytest.warns(UserWarning, match=message):
doc._validate_rules(raise_on_error=False)
def test_meta_migration_warnings():
# the following should not raise any warnings
with warnings.catch_warnings():
warnings.simplefilter("error")
doc = DoclingDocument.load_from_yaml("tests/data/doc/dummy_doc_2.yaml")
# the following should raise a deprecation warning
with pytest.warns(DeprecationWarning):
_ = doc.pictures[0].annotations
with pytest.warns(DeprecationWarning):
_ = doc.tables[0].annotations
def test_migrate_non_list_item_list_children():
doc = DoclingDocument(name="")
outer = doc.add_list_group()
doc.add_list_item(text="first", parent=outer)
inner = doc.add_list_group(parent=outer)
doc.add_list_item(text="nested", parent=inner)
text = doc.add_text(label=DocItemLabel.TEXT, text="stray", parent=outer)
with pytest.raises(ValueError, match="contains non-ListItem"):
doc._validate_rules()
assert doc._migrate_non_list_item_list_children() == 2
# the wrapped children keep their positions, each under a new, empty list item
first, wrap_inner, wrap_text = (ref.resolve(doc) for ref in outer.children)
assert first.text == "first"
for wrapper, child in ((wrap_inner, inner), (wrap_text, text)):
assert isinstance(wrapper, ListItem) and wrapper.text == ""
assert wrapper.parent == outer.get_ref()
assert [ref.resolve(doc) for ref in wrapper.children] == [child]
assert child.parent == wrapper.get_ref()
doc._validate_rules()
assert doc._migrate_non_list_item_list_children() == 0 # idempotent
def test_empty_groups():
def group(idx: int, label: str, parent: str, children: list[str] | None = None) -> dict:
return {
"self_ref": f"#/groups/{idx}",
"parent": {"$ref": parent},
"children": [{"$ref": c} for c in children or []],
"label": label,
"name": "group",
}
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "empty_groups",
"body": {
"self_ref": "#/body",
"children": [{"$ref": "#/groups/0"}, {"$ref": "#/groups/1"}, {"$ref": "#/groups/4"}],
"name": "_root_",
"label": "unspecified",
},
"groups": [
group(0, "list", "#/body"), # empty, linked
group(1, "list", "#/body", ["#/texts/0"]), # not empty
group(2, "list", "#/body"), # empty, orphaned
group(3, "list", "#/groups/2"), # empty, orphaned, claiming #/groups/2 as parent
group(4, "inline", "#/body"), # empty, linked
group(5, "unspecified", "#/body"), # empty, orphaned, but claimed by an orphan with content
],
"texts": [
{"self_ref": "#/texts/0", "parent": {"$ref": "#/groups/1"}, "label": "list_item", "orig": "a", "text": "a"},
{"self_ref": "#/texts/1", "parent": {"$ref": "#/groups/5"}, "label": "text", "orig": "b", "text": "b"},
],
}
doc = DoclingDocument.model_validate(doc_dict)
# detected: every empty group, also lists and orphaned ones
with pytest.warns(UserWarning) as record:
doc._validate_rules(raise_on_error=False)
empty = sorted(str(w.message) for w in record if str(w.message).endswith("has no children"))
assert empty == [f"Group #/groups/{i} has no children" for i in (0, 2, 3, 4, 5)]
assert doc._remove_empty_groups() == 4
assert [g.label.value for g in doc.groups] == ["list", "unspecified"]
assert [ref.cref for ref in doc.body.children] == ["#/groups/0"]
assert [ref.cref for ref in doc.groups[0].children] == ["#/texts/0"]
assert doc.texts[0].parent.cref == "#/groups/0"
assert doc.texts[1].parent.cref == "#/groups/1" # the claimed group is kept (and renumbered)
doc._validate_unique_refs()
assert doc._remove_empty_groups() == 0 # idempotent
def test_repair_referenced_orphans():
def text(idx: int, label: str) -> dict:
return {
"self_ref": f"#/texts/{idx}",
"parent": {"$ref": "#/pictures/0"},
"label": label,
"orig": label,
"text": label,
}
doc_dict = {
"schema_name": "DoclingDocument",
"version": CURRENT_VERSION,
"name": "referenced_orphans",
"body": {
"self_ref": "#/body",
"children": [{"$ref": "#/pictures/0"}],
"name": "_root_",
"label": "unspecified",
},
# all three are orphans; only the caption and the footnote are referenced by the picture
"texts": [text(0, "caption"), text(1, "footnote"), text(2, "text")],
"pictures": [
{
"self_ref": "#/pictures/0",
"parent": {"$ref": "#/body"},
"label": "picture",
"captions": [{"$ref": "#/texts/0"}],
"footnotes": [{"$ref": "#/texts/1"}],
}
],
}
doc = DoclingDocument.model_validate(doc_dict)
assert doc._repair_referenced_orphans() == 2
assert [ref.cref for ref in doc.pictures[0].children] == ["#/texts/0", "#/texts/1"]
assert doc._repair_referenced_orphans() == 0 # idempotent
# the unreferenced orphan is left as is
with pytest.raises(ValueError, match="#/texts/2 is not a child of its parent #/pictures/0"):
doc._validate_rules()
@pytest.mark.parametrize(
"example_num",
[1, 2, 3, 4, 5],
)
def test_webvtt_export(example_num):
"""Test WebVTT export with example files that contain TrackSource data."""
json_file = Path(f"tests/data/doc/webvtt_example_{example_num:02d}.json")
gt_vtt_file = json_file.with_suffix(".gt.vtt")
# Load the document
doc = DoclingDocument.load_from_json(json_file)
# Export to WebVTT
vtt_output = doc.export_to_vtt()
# Verify against ground truth
if GEN_TEST_DATA:
with open(gt_vtt_file, "w", encoding="utf-8") as fw:
fw.write(f"{vtt_output}\n")
else:
with open(gt_vtt_file, encoding="utf-8") as fr:
gt_vtt = fr.read().rstrip()
assert vtt_output == gt_vtt, f"WebVTT output does not match ground truth for example {example_num:02d}"
def test_validate_dupl_refs():
doc = DoclingDocument(name="")
t1 = doc.add_text(label=DocItemLabel.TEXT, text="foo")
t2 = doc.add_text(label=DocItemLabel.TEXT, text="bar")
t2.self_ref = t1.self_ref # duplicating the self_ref
t1.parent.resolve(doc).children = [t1.get_ref()] # removing the dangling pointer
with pytest.raises(ValidationError) as valid_err_info:
DoclingDocument.model_validate(doc)
error_str = str(valid_err_info.value)
assert "Duplicate ref" in error_str
def test_docitem_comments_field():
"""Test that DocItem has a comments field that can hold RefItem references."""
doc = DoclingDocument(name="test_comments")
doc.add_text(label=DocItemLabel.TEXT, text="Normal text without comment.")
# Add a paragraph (which is a DocItem)
text = doc.add_text(
label=DocItemLabel.TEXT,
text="This text has a comment attached.",
)
# Add a comment
doc.add_comment(
text="[John Reviewer]: This is a reviewer comment.",
targets=[text],
)
exp_file = Path("tests/data/doc/docitem_comments_field.out.yaml")
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
def test_docitem_comments_multiple():
"""Test that a DocItem can have multiple comments attached."""
doc = DoclingDocument(name="test_multiple_comments")
text1 = doc.add_text(
label=DocItemLabel.TEXT,
text="Text 1.",
)
text2 = doc.add_text(
label=DocItemLabel.TEXT,
text="Text 2.",
)
text3 = doc.add_text(
label=DocItemLabel.TEXT,
text="Text 3.",
)
# Add comments on overlapping scopes:
doc.add_comment(
text="[Reviewer A]: This is a comment on texts 1 and 2.",
targets=[
text1,
text2,
],
)
doc.add_comment(
text="[Reviewer B]: This is a comment on texts 2 (range [0,6)) and 3.",
targets=[
(text2, (0, 6)),
text3,
],
)
exp_file = Path("tests/data/doc/docitem_comments_multiple.out.yaml")
if GEN_TEST_DATA:
doc.save_as_yaml(exp_file)
exp_doc = DoclingDocument.load_from_yaml(exp_file)
assert doc == exp_doc
def test_docitem_comments_delete_updates_refs():
"""Test that deleting items properly updates comment references."""
doc = DoclingDocument(name="test_comments_delete")
# Add two paragraphs
para1 = doc.add_text(
label=DocItemLabel.PARAGRAPH,
text="First paragraph.",
)
para2 = doc.add_text(
label=DocItemLabel.PARAGRAPH,
text="Second paragraph with comment.",
)
# Add a comment group for the second paragraph
doc.add_comment(
text="Comment on second paragraph.",
targets=[para2],
)
# Delete the first paragraph - this should update indices
doc.delete_items(node_items=[para1])
# The second paragraph is now the first text item
assert len(doc.texts) >= 1
# The comment reference should still be valid
updated_para = doc.texts[0]
assert len(updated_para.comments) == 1
# The resolved comment should still work
resolved = updated_para.comments[0].resolve(doc)
assert resolved.text == "Comment on second paragraph."
def test_table_data_vertical_bounding_boxes():
"""Vertical table: rows are vertical stripes, columns are horizontal stripes.
When `horizontal=False` and `minimal=False`, rows should share a common
vertical (t/b) extent and columns should share a common horizontal (l/r) extent.
"""
# 2 rows x 3 cols vertical table (logical rows run top-to-bottom on page).
# Row 0 cells sit on the page-left (l=0..10), row 1 on the page-right (l=10..20).
# Col 0 cells sit at the page-top, col 2 at the page-bottom.
cells = [
TableCell(
text="r0c0",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0, t=2, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c0",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=10, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r0c1",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=1,
end_col_offset_idx=2,
bbox=BoundingBox(l=0, t=10, r=10, b=20, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c1",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=1,
end_col_offset_idx=2,
bbox=BoundingBox(l=10, t=10, r=18, b=20, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r0c2",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=2,
end_col_offset_idx=3,
bbox=BoundingBox(l=0, t=20, r=10, b=28, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c2",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=2,
end_col_offset_idx=3,
bbox=BoundingBox(l=10, t=20, r=20, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
]
table = TableData(num_rows=2, num_cols=3, table_cells=cells, orientation=Orientation.ROT_90)
# Minimal mode — sanity-check the per-row/col enclosing bbox.
rows_min = table.get_row_bounding_boxes(minimal=True)
assert rows_min[0].l == 0 and rows_min[0].r == 10
assert rows_min[0].t == 2 and rows_min[0].b == 28
assert rows_min[1].l == 10 and rows_min[1].r == 20
assert rows_min[1].t == 0 and rows_min[1].b == 30
cols_min = table.get_column_bounding_boxes(minimal=True)
assert cols_min[0].t == 0 and cols_min[0].b == 10
assert cols_min[1].t == 10 and cols_min[1].b == 20
assert cols_min[2].t == 20 and cols_min[2].b == 30
# Non-minimal + vertical: rows share t/b, columns share l/r.
rows = table.get_row_bounding_boxes(minimal=False)
assert rows[0].t == 0 and rows[0].b == 30
assert rows[1].t == 0 and rows[1].b == 30
# Per-row l/r extents must remain distinct (rows are vertical stripes).
assert rows[0].l == 0 and rows[0].r == 10
assert rows[1].l == 10 and rows[1].r == 20
cols = table.get_column_bounding_boxes(minimal=False)
for c in cols.values():
assert c.l == 0 and c.r == 20
# Per-col t/b extents must remain distinct (cols are horizontal stripes).
assert (cols[0].t, cols[0].b) == (0, 10)
assert (cols[1].t, cols[1].b) == (10, 20)
assert (cols[2].t, cols[2].b) == (20, 30)
# Sanity: with ROT_0 orientation the equalized axes flip back.
table.orientation = Orientation.ROT_0
rows_h = table.get_row_bounding_boxes(minimal=False)
for r in rows_h.values():
assert r.l == 0 and r.r == 20
cols_h = table.get_column_bounding_boxes(minimal=False)
for c in cols_h.values():
assert c.t == 0 and c.b == 30
def test_table_data_vertical_bounding_boxes_with_spans():
"""Vertical table with spanning cells.
In a vertical table, a col_span=2 cell occupies two consecutive horizontal
column stripes on the page, so it extends across two t/b ranges. Its presence
must extend the spanned columns along l/r (the column's natural axis), NOT
along t/b — otherwise column N's bbox bleeds into column N+1.
Symmetric for row_span=2: extends along t/b, not l/r.
"""
# Layout (TOPLEFT, 2 rows x 3 cols, vertical):
# row 0 cells live in l=0..10, row 1 cells in l=10..20
# col 0 stripe is at t=0..10, col 1 at t=10..20, col 2 at t=20..30
# Cell (r0, c1+c2) is a col_span=2 cell in row 0 spanning cols 1 and 2:
# it lives at l=0..10 (row 0) and t=10..30 (both col 1 and col 2 stripes).
# Cell (r0+r1, c0) is a row_span=2 cell in col 0 spanning rows 0 and 1:
# it lives at t=0..10 (col 0) and l=0..20 (both row 0 and row 1 stripes).
cells = [
TableCell( # row_span=2 in col 0
text="r0+r1,c0",
start_row_offset_idx=0,
end_row_offset_idx=2,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0, t=0, r=20, b=10, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r0c1+c2", # col_span=2 in row 0
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=1,
end_col_offset_idx=3,
bbox=BoundingBox(l=0, t=10, r=10, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c1",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=1,
end_col_offset_idx=2,
bbox=BoundingBox(l=10, t=10, r=20, b=20, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c2",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=2,
end_col_offset_idx=3,
bbox=BoundingBox(l=10, t=20, r=20, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
]
table = TableData(num_rows=2, num_cols=3, table_cells=cells, orientation=Orientation.ROT_90)
# COLUMNS — col_span=2 cell must NOT stretch col 1 / col 2 along t/b.
cols = table.get_column_bounding_boxes(minimal=True)
# col 1 stripe on page is t=10..20; the col_span cell adds l-extent only.
assert (cols[1].t, cols[1].b) == (10, 20)
assert cols[1].l == 0 and cols[1].r == 20
# col 2 stripe on page is t=20..30; col_span cell again adds only l-extent.
assert (cols[2].t, cols[2].b) == (20, 30)
assert cols[2].l == 0 and cols[2].r == 20
# ROWS — row_span=2 cell must NOT stretch row 0 / row 1 along l/r.
rows = table.get_row_bounding_boxes(minimal=True)
# row 0 stripe on page is l=0..10; the row_span cell adds t-extent only.
assert (rows[0].l, rows[0].r) == (0, 10)
assert rows[0].t == 0 and rows[0].b == 30
# row 1 stripe on page is l=10..20; row_span cell again adds only t-extent.
assert (rows[1].l, rows[1].r) == (10, 20)
assert rows[1].t == 0 and rows[1].b == 30
# Non-overlap property: row bboxes (and col bboxes) must be pairwise
# disjoint when cell bboxes are sane and separated.
for a, b in itertools.combinations(rows.values(), 2):
assert a.intersection_area_with(b) == 0
for a, b in itertools.combinations(cols.values(), 2):
assert a.intersection_area_with(b) == 0
# Same check in minimal=False mode.
rows_nm = table.get_row_bounding_boxes(minimal=False)
cols_nm = table.get_column_bounding_boxes(minimal=False)
for a, b in itertools.combinations(rows_nm.values(), 2):
assert a.intersection_area_with(b) == 0
for a, b in itertools.combinations(cols_nm.values(), 2):
assert a.intersection_area_with(b) == 0
def test_table_data_horizontal_bounding_boxes_with_spans_no_overlap():
"""Horizontal table with spanning cells: row/col bboxes must not overlap."""
# 3 rows x 3 cols horizontal table with one col_span=2 and one row_span=2 cell.
cells = [
TableCell(
text="r0c0",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0, t=0, r=10, b=10, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell( # col_span=2
text="r0c1+c2",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=1,
end_col_offset_idx=3,
bbox=BoundingBox(l=10, t=0, r=30, b=10, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell( # row_span=2
text="r1+r2,c0",
start_row_offset_idx=1,
end_row_offset_idx=3,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0, t=10, r=10, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c1",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=1,
end_col_offset_idx=2,
bbox=BoundingBox(l=10, t=10, r=20, b=20, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r1c2",
start_row_offset_idx=1,
end_row_offset_idx=2,
start_col_offset_idx=2,
end_col_offset_idx=3,
bbox=BoundingBox(l=20, t=10, r=30, b=20, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r2c1",
start_row_offset_idx=2,
end_row_offset_idx=3,
start_col_offset_idx=1,
end_col_offset_idx=2,
bbox=BoundingBox(l=10, t=20, r=20, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
TableCell(
text="r2c2",
start_row_offset_idx=2,
end_row_offset_idx=3,
start_col_offset_idx=2,
end_col_offset_idx=3,
bbox=BoundingBox(l=20, t=20, r=30, b=30, coord_origin=CoordOrigin.TOPLEFT),
),
]
table = TableData(num_rows=3, num_cols=3, table_cells=cells)
for minimal in (True, False):
rows = table.get_row_bounding_boxes(minimal=minimal)
cols = table.get_column_bounding_boxes(minimal=minimal)
for a, b in itertools.combinations(rows.values(), 2):
assert a.intersection_area_with(b) == 0, f"row overlap (minimal={minimal})"
for a, b in itertools.combinations(cols.values(), 2):
assert a.intersection_area_with(b) == 0, f"col overlap (minimal={minimal})"
def _validate_doc(doc: DoclingDocument) -> DoclingDocument:
return DoclingDocument.model_validate(doc.model_dump(mode="json"))
def test_document_validation_clamps_provenance_bbox_without_warning_within_tolerance() -> None:
doc = DoclingDocument(name="test")
doc.add_page(page_no=1, size=Size(width=100.0, height=200.0))
doc.add_text(
label=DocItemLabel.TEXT,
text="Hello",
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=-1.0, t=0.0, r=101.0, b=201.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 5),
),
)
with warnings.catch_warnings(record=True) as caught:
warnings.simplefilter("always")
validated = _validate_doc(doc)
bbox = validated.texts[0].prov[0].bbox
assert (bbox.l, bbox.t, bbox.r, bbox.b) == (0.0, 0.0, 100.0, 200.0)
assert not [warning for warning in caught if "outside page bounds" in str(warning.message)]
def test_document_validation_warns_when_provenance_bbox_exceeds_tolerance() -> None:
doc = DoclingDocument(name="test")
doc.add_page(page_no=1, size=Size(width=100.0, height=200.0))
doc.add_text(
label=DocItemLabel.TEXT,
text="Hello",
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=-1.1, t=0.0, r=100.0, b=202.1, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 5),
),
)
with warnings.catch_warnings(record=True) as caught:
warnings.simplefilter("always")
validated = _validate_doc(doc)
bbox = validated.texts[0].prov[0].bbox
assert (bbox.l, bbox.t, bbox.r, bbox.b) == (0.0, 0.0, 100.0, 200.0)
messages = [str(warning.message) for warning in caught]
assert any("coordinate l on page 1 is outside page bounds" in message for message in messages)
assert any("coordinate b on page 1 is outside page bounds" in message for message in messages)
def test_document_validation_clamps_graph_cell_provenance_bbox() -> None:
doc = DoclingDocument(name="test")
doc.add_page(page_no=1, size=Size(width=100.0, height=200.0))
doc.add_key_values(
graph=GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=0,
text="Key",
orig="Key",
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=0.0, t=-0.5, r=10.0, b=10.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 3),
),
)
]
)
)
validated = _validate_doc(doc)
prov = validated.key_value_items[0].graph.cells[0].prov
assert prov and prov.bbox.t == 0.0
def test_document_validation_clamps_table_cell_bbox() -> None:
doc = DoclingDocument(name="test")
doc.add_page(page_no=1, size=Size(width=100.0, height=200.0))
doc.add_table(
data=TableData(
num_rows=1,
num_cols=1,
table_cells=[
TableCell(
text="cell",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=-1.1, t=0.0, r=101.1, b=10.0, coord_origin=CoordOrigin.TOPLEFT),
)
],
),
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=0.0, t=0.0, r=100.0, b=10.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 4),
),
)
with warnings.catch_warnings(record=True) as caught:
warnings.simplefilter("always")
validated = _validate_doc(doc)
bbox = validated.tables[0].data.table_cells[0].bbox
assert bbox and (bbox.l, bbox.t, bbox.r, bbox.b) == (0.0, 0.0, 100.0, 10.0)
messages = [str(warning.message) for warning in caught]
assert any("Table cell bbox coordinate l on page 1 is outside page bounds" in message for message in messages)
assert any("Table cell bbox coordinate r on page 1 is outside page bounds" in message for message in messages)
def test_doc_item_lists_cover_all_item_fields() -> None:
"""A new item list on DoclingDocument must be registered in `_doc_item_lists`.
Otherwise bbox clamping and `get_min_page_sizes` would silently skip its items.
Groups carry no provenance and are the only intended exception.
"""
doc = DoclingDocument(name="test")
registered = {id(item_list) for item_list in doc._doc_item_lists()}
for field_name, value in doc: # pydantic models iterate as (field name, value)
if isinstance(value, list) and field_name != "groups":
assert id(value) in registered, f"{field_name} is missing in DoclingDocument._doc_item_lists"
def test_get_min_page_size_empty_page() -> None:
doc = DoclingDocument(name="test")
assert doc.get_min_page_sizes() == {}
assert doc.get_min_page_size(1) == Size(width=0.0, height=0.0)
def test_get_min_page_size_covers_content_and_survives_validation() -> None:
doc = DoclingDocument(name="test")
doc.add_text(
label=DocItemLabel.TEXT,
text="a",
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=1.0, t=1.0, r=3.0, b=2.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 1),
),
)
# furniture layer is included
doc.add_text(
label=DocItemLabel.PAGE_FOOTER,
text="f",
content_layer=ContentLayer.FURNITURE,
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=0.0, t=8.0, r=2.0, b=10.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 1),
),
)
# other pages are ignored
doc.add_text(
label=DocItemLabel.TEXT,
text="p2",
prov=ProvenanceItem(
page_no=2,
bbox=BoundingBox(l=0.0, t=0.0, r=50.0, b=60.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 2),
),
)
# table cell bboxes count as well
doc.add_table(
data=TableData(
num_rows=1,
num_cols=1,
table_cells=[
TableCell(
text="cell",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0.0, t=0.0, r=7.0, b=4.0, coord_origin=CoordOrigin.TOPLEFT),
)
],
),
prov=ProvenanceItem(
page_no=1,
bbox=BoundingBox(l=0.0, t=0.0, r=5.0, b=4.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 4),
),
)
# cell bboxes of multi-page tables are not anchored on a page
doc.add_table(
data=TableData(
num_rows=1,
num_cols=1,
table_cells=[
TableCell(
text="cell",
start_row_offset_idx=0,
end_row_offset_idx=1,
start_col_offset_idx=0,
end_col_offset_idx=1,
bbox=BoundingBox(l=0.0, t=0.0, r=99.0, b=99.0, coord_origin=CoordOrigin.TOPLEFT),
)
],
),
prov=ProvenanceItem(
page_no=2,
bbox=BoundingBox(l=0.0, t=0.0, r=1.0, b=1.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 4),
),
).prov.append(
ProvenanceItem(
page_no=3,
bbox=BoundingBox(l=0.0, t=0.0, r=1.0, b=1.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 4),
)
)
# key-value graph cell provenances count as well
doc.add_key_values(
graph=GraphData(
cells=[
GraphCell(
label=GraphCellLabel.KEY,
cell_id=0,
text="k",
orig="k",
prov=ProvenanceItem(
page_no=3,
bbox=BoundingBox(l=0.0, t=0.0, r=20.0, b=30.0, coord_origin=CoordOrigin.TOPLEFT),
charspan=(0, 1),
),
)
],
links=[],
)
)
sizes = doc.get_min_page_sizes()
assert sizes == {
1: Size(width=7.0, height=10.0),
2: Size(width=50.0, height=60.0),
3: Size(width=20.0, height=30.0),
}
size = doc.get_min_page_size(1)
assert size == sizes[1]
# a page of exactly this size loses no content to the bbox clamping
doc.add_page(page_no=1, size=size)
validated = _validate_doc(doc)
bbox = validated.texts[0].prov[0].bbox
assert (bbox.l, bbox.t, bbox.r, bbox.b) == (1.0, 1.0, 3.0, 2.0)
cell_bbox = validated.tables[0].data.table_cells[0].bbox
assert cell_bbox and cell_bbox.r == 7.0
footer_bbox = validated.texts[1].prov[0].bbox
assert footer_bbox.b == 10.0
def _nested_doc_dict(depth: int, *, kind: str) -> dict:
"""Build a document dict whose nodes reference their single child more than once.
kind="groups": every group lists its only child twice (a DAG, not a tree).
kind="rich_cells": every table holds the next table in a single 2x2-spanning rich cell.
"""
node_list = "groups" if kind == "groups" else "tables"
nodes = []
for i in range(depth):
last = i == depth - 1
parent = {"$ref": "#/body" if i == 0 else f"#/{node_list}/{i - 1}"}
nxt = {"$ref": f"#/{node_list}/{i + 1}"}
node: dict = {
"self_ref": f"#/{node_list}/{i}",
"parent": parent,
"content_layer": "body",
}
if kind == "groups":
node |= {"name": "g", "label": "unspecified", "children": [] if last else [nxt, nxt]}
else:
cells = [] if last else [_span_rich_cell(nxt["$ref"])]
node |= {
"label": "table",
"children": [] if last else [nxt],
"data": {"num_rows": 2, "num_cols": 2, "table_cells": cells},
}
nodes.append(node)
return {
"schema_name": "DoclingDocument",
"version": "1.10.0",
"name": "nested",
"body": {
"self_ref": "#/body",
"name": "_root_",
"label": "unspecified",
"content_layer": "body",
"children": [{"$ref": f"#/{node_list}/0"}],
},
node_list: nodes,
}
def _span_rich_cell(ref: str) -> dict:
return {
"text": "",
"ref": {"$ref": ref},
"start_row_offset_idx": 0,
"end_row_offset_idx": 2,
"start_col_offset_idx": 0,
"end_col_offset_idx": 2,
}
def test_validate_rejects_repeated_child_references():
"""A node listed more than once under its parent makes the tree a DAG with exponential traversal."""
data = _nested_doc_dict(60, kind="groups") # 2**60 paths if walked once per reference
with pytest.raises(ValidationError, match="already reachable"):
DoclingDocument.model_validate(data)
def test_validate_rejects_rich_cell_sharing_one_subtree():
"""Two rich cells pointing at the same item would render its subtree twice."""
data = _nested_doc_dict(2, kind="rich_cells")
cells = data["tables"][0]["data"]["table_cells"]
cells.append(cells[0] | {"start_col_offset_idx": 1, "end_col_offset_idx": 2})
with pytest.raises(ValidationError):
DoclingDocument.model_validate(data)
@pytest.mark.parametrize(
"export",
[
lambda doc: doc.export_to_markdown(),
lambda doc: doc.export_to_doctags(),
lambda doc: doc.export_to_doclang(),
],
ids=["markdown", "doctags", "doclang"],
)
def test_export_nested_spanning_rich_cells_is_not_exponential(export):
"""A rich cell spanning N slots must be rendered once, not once per covered slot.
With 4 slots per cell and 25 nesting levels, per-slot rendering would take ~4**25 steps.
"""
doc = DoclingDocument.model_validate(_nested_doc_dict(25, kind="rich_cells"))
assert export(doc)
@pytest.mark.parametrize(
("num_rows", "num_cols"),
[(2**40, 2**40), (2**40, 0), (0, 2**40), (1001, 1000)],
)
def test_table_data_rejects_oversized_grid(num_rows, num_cols):
"""Rows and columns are bounded separately too: many rows with no columns still allocate one list each."""
with pytest.raises(ValidationError, match="exceeds the limit"):
TableData(num_rows=num_rows, num_cols=num_cols)
def test_table_data_grid_limit_is_configurable(monkeypatch):
monkeypatch.setattr(settings, "max_table_cells", 4)
TableData(num_rows=2, num_cols=2)
with pytest.raises(ValidationError):
TableData(num_rows=2, num_cols=3)
def _tables_doc_dict(body: int, furniture: int = 0, orphans: int = 0) -> dict:
"""Build a document dict with 2x2 tables under body, under furniture, and outside the tree."""
placement = ["#/body"] * body + ["#/furniture"] * furniture + ["#/body"] * orphans
tables = [
{
"self_ref": f"#/tables/{i}",
"parent": {"$ref": parent},
"label": "table",
"data": {"num_rows": 2, "num_cols": 2, "table_cells": []},
}
for i, parent in enumerate(placement)
]
def root(name: str, start: int, count: int) -> dict:
return {
"self_ref": f"#/{name}",
"name": "_root_",
"label": "unspecified",
"children": [{"$ref": f"#/tables/{i}"} for i in range(start, start + count)],
}
return {
"schema_name": "DoclingDocument",
"version": "1.10.0",
"name": "tables",
"body": root("body", 0, body),
"furniture": root("furniture", body, furniture),
"tables": tables,
}
@pytest.mark.parametrize(
("body", "furniture", "orphans"),
[(3, 0, 0), (1, 2, 0), (1, 0, 2)],
ids=["body", "body_and_furniture", "outside_tree"],
)
def test_validate_rejects_total_table_cells_over_limit(monkeypatch, body, furniture, orphans):
"""The budget covers all tables, wherever they sit: every grid in `tables` is serialized."""
monkeypatch.setattr(settings, "max_total_table_cells", 8)
DoclingDocument.model_validate(_tables_doc_dict(body=2))
with pytest.raises(ValidationError, match="exceeding the limit of 8"):
DoclingDocument.model_validate(_tables_doc_dict(body=body, furniture=furniture, orphans=orphans))