From 2dcf54293c03c65333d230078858ec80248fa2e6 Mon Sep 17 00:00:00 2001 From: Afloat Date: Sat, 21 Nov 2026 11:20:53 +0800 Subject: [PATCH] fix(html): read table row groups in rendered order --- .../html/test_table_rendered_order.py | 70 +++++++++++++++++++ unstructured/common/html_table.py | 22 ++++++ unstructured/partition/html/parser.py | 4 ++ 3 files changed, 96 insertions(+) create mode 100644 test_unstructured/partition/html/test_table_rendered_order.py diff --git a/test_unstructured/partition/html/test_table_rendered_order.py b/test_unstructured/partition/html/test_table_rendered_order.py new file mode 100644 index 0000000000..80c2af88be --- /dev/null +++ b/test_unstructured/partition/html/test_table_rendered_order.py @@ -0,0 +1,70 @@ +import pytest + +from unstructured.common.html_table import HtmlTable +from unstructured.partition.html import partition_html + + +@pytest.mark.parametrize( + ("body", "expected"), + [ + ( + "Footer" + "Body" + "Header", + "Header Body Footer", + ), + ( + "Footer" + "FirstSecond", + "First Second Footer", + ), + ( + "Footer" + "Extra footer" + "Body" + "Header" + "Extra header", + "Header Extra footer Body Extra header Footer", + ), + ("FirstSecond", "First Second"), + ], +) +def test_partition_and_chunking_table_reader_follow_rendered_row_group_order(body, expected): + html = f"{body}
" + + element = partition_html(text=html, languages=["eng"])[0] + compact = HtmlTable.from_html_text(html) + + assert element.text == expected + assert compact.text == expected + assert " ".join(" ".join(row.iter_cell_texts()) for row in compact.iter_rows()) == expected + + +def test_source_rows_and_header_flags_remain_aligned_after_ordering(): + table = HtmlTable.from_html_text( + '' + '' + '
Foot
Body
Head
' + ) + rows = list(table.iter_rows()) + + assert [list(row.iter_cell_texts()) for row in rows] == [["Head"], ["Body"], ["Foot"]] + assert [row.is_header for row in rows] == [True, False, False] + assert [row.row_group_key.tag for row in rows] == ["thead", "tbody", "tfoot"] + assert 'scope="col"' in rows[0].source_html + assert 'rowspan="0"' in rows[1].html + assert 'data-id="f"' in rows[2].source_html + + +def test_nested_table_order_is_normalized_without_moving_rows_between_tables(): + html = ( + "
Outer " + "" + "" + "
Inner footer
Inner body
" + ) + table = HtmlTable.from_html_text(html) + + assert table.text == "Outer Inner body Inner footer" + assert len(list(table.iter_rows())) == 1 + assert partition_html(text=html, languages=["eng"])[0].text == table.text diff --git a/unstructured/common/html_table.py b/unstructured/common/html_table.py index 1c3b90304e..ca46224797 100644 --- a/unstructured/common/html_table.py +++ b/unstructured/common/html_table.py @@ -21,6 +21,26 @@ SpannedCell: TypeAlias = "tuple[str, int, int]" +def iter_table_children_in_render_order(table: etree._Element) -> Iterator[etree._Element]: + """Yield direct children with the first header first and first footer last among rows. + + Under the default HTML table display rules, the first ``thead``/``tfoot`` is laid out + before/after other row groups regardless of source order. Additional headers/footers + behave as ordinary row groups. Non-row children retain their positions. + """ + row_tags = {"thead", "tbody", "tfoot", "tr"} + rows = [child for child in table if child.tag in row_tags] + header = next((child for child in rows if child.tag == "thead"), None) + footer = next((child for child in rows if child.tag == "tfoot"), None) + ordered = iter( + ([header] if header is not None else []) + + [child for child in rows if child is not header and child is not footer] + + ([footer] if footer is not None else []) + ) + for child in table: + yield next(ordered) if child.tag in row_tags else child + + def normalize_html_cell_text(cell: HtmlElement) -> str: """Return normalized cell text, treating a ``
`` as a word boundary. @@ -178,6 +198,8 @@ def from_html_text(cls, html_text: str) -> HtmlTable: if not tables: raise ValueError("`html_text` contains no `` element") table = tables[0] + for nested_table in table.iter("table"): + nested_table[:] = list(iter_table_children_in_render_order(nested_table)) # -- capture header semantics, source row HTML, and row-group identity before # -- compactification strips those details -- diff --git a/unstructured/partition/html/parser.py b/unstructured/partition/html/parser.py index 30c45e836e..d8cc63da43 100644 --- a/unstructured/partition/html/parser.py +++ b/unstructured/partition/html/parser.py @@ -86,6 +86,7 @@ from typing_extensions import TypeAlias from unstructured.cleaners.core import clean_bullets +from unstructured.common.html_table import iter_table_children_in_render_order from unstructured.documents.elements import ( Address, CodeSnippet, @@ -568,6 +569,9 @@ class TableBlock(Flow): def iter_elements(self) -> Iterator[Table]: """Generate paragraph string for each block item within.""" + for table in self.iter("table"): + table[:] = list(iter_table_children_in_render_order(table)) + # -- NOTE this algorithm handles a nested-table by parsing all of its text into the text # -- for the _cell_ containing the table (and this is recursive, so a table nested within # -- a cell within a table within a cell too.)