Skip to content

Build and run agents you can see, understand and trust.

fix(rag): preserve paragraph boundaries inside Word text boxes (#3415)

12kkqq0d5ea1ad5ee8
diff --git a/src/agentscope/rag/_parser/_word.py b/src/agentscope/rag/_parser/_word.py
index 762e76d59..2ec9dbc23 100644
--- a/src/agentscope/rag/_parser/_word.py
+++ b/src/agentscope/rag/_parser/_word.py
@@ -51,7 +51,8 @@ def _extract_text_from_paragraph(para: DocxParagraph) -> str:
Each ``mc:AlternateContent`` containing a text box uses the first branch
with readable Word text, retaining the fallback when its choices cannot
supply any text. Run-level tabs and soft line breaks are preserved in
- both the paragraph and its text boxes.
+ both the paragraph and its text boxes. Paragraph boundaries within each
+ text box are retained.
Args:
para (`DocxParagraph`):
@@ -64,6 +65,7 @@ def _extract_text_from_paragraph(para: DocxParagraph) -> str:
from docx.oxml.ns import qn
text_tag = qn("w:t")
+ paragraph_tag = qn("w:p")
text_box_tag = qn("w:txbxContent")
deleted_tag = qn("w:del")
run_tag = qn("w:r")
@@ -101,7 +103,7 @@ def _extract_text_from_paragraph(para: DocxParagraph) -> str:
yield from iter_text(child)
own_parts: list[str] = []
- text_boxes: dict[Any, list[str]] = {}
+ text_boxes: dict[Any, dict[Any, list[str]]] = {}
for element in iter_text(root):
if element.tag == text_tag:
value = element.text or ""
@@ -117,21 +119,35 @@ def _extract_text_from_paragraph(para: DocxParagraph) -> str:
if not value:
continue
parts = own_parts
+ paragraph = root
for ancestor in element.iterancestors():
if ancestor is root:
break
+ if ancestor.tag == paragraph_tag:
+ paragraph = ancestor
if ancestor.tag == text_box_tag:
- parts = text_boxes.setdefault(ancestor, [])
+ parts = text_boxes.setdefault(ancestor, {}).setdefault(
+ paragraph,
+ [],
+ )
break
parts.append(value)
own_text = "".join(own_parts)
if not own_text.strip() and not any(
- "".join(parts).strip() for parts in text_boxes.values()
+ "".join(parts).strip()
+ for paragraphs in text_boxes.values()
+ for parts in paragraphs.values()
):
# Separator-only extraction must not hide w:noBreakHyphen, etc.
own_text = para.text
- blocks = [own_text, *("".join(parts) for parts in text_boxes.values())]
+ blocks = [
+ own_text,
+ *(
+ "\n".join("".join(parts) for parts in paragraphs.values())
+ for paragraphs in text_boxes.values()
+ ),
+ ]
return "\n".join(block.strip() for block in blocks if block.strip())
diff --git a/tests/rag_parser_test.py b/tests/rag_parser_test.py
index e787b0ea1..5aff9d447 100644
--- a/tests/rag_parser_test.py
+++ b/tests/rag_parser_test.py
@@ -3119,6 +3119,67 @@ class WordParserParagraphTest(IsolatedAsyncioTestCase):
_expected_word_sections("Paragraph text\nCallout\nAfter"),
)
+ async def test_text_box_paragraph_boundaries_are_preserved(self) -> None:
+ """Separate text-box paragraphs while joining runs within each."""
+ paragraphs = (
+ "<w:p><w:r><w:t>First </w:t></w:r>"
+ "<w:r><w:t>paragraph</w:t></w:r></w:p>"
+ "<w:p><w:r><w:t>Second paragraph</w:t></w:r></w:p>"
+ )
+ for name, template in (
+ ("modern", _DOCX_MODERN_TEXT_BOX),
+ ("with_fallback", _DOCX_TEXT_BOX_WITH_VML_FALLBACK),
+ (
+ "fallback_only",
+ _DOCX_TEXT_BOX_WITH_VML_FALLBACK.replace(
+ "<w:t>{inside}</w:t>",
+ "<w:t/>",
+ 1,
+ ),
+ ),
+ ):
+ with self.subTest(shape=name):
+ template = template.replace(
+ "<w:p><w:r><w:t>{inside}</w:t></w:r></w:p>",
+ paragraphs,
+ )
+ sections = await WordParser(include_image=False).parse(
+ _make_docx_with_text_box(template),
+ "demo.docx",
+ )
+ self.assertEqual(
+ [section.model_dump() for section in sections],
+ _expected_word_sections(
+ "Paragraph text\nFirst paragraph\n"
+ "Second paragraph\nAfter",
+ ),
+ )
+
+ async def test_text_box_paragraphs_keep_inline_separators(self) -> None:
+ """Keep paragraph and inline boundaries with no surrounding text."""
+ paragraphs = (
+ "<w:p><w:r><w:t>A</w:t><w:tab/><w:t>B</w:t></w:r></w:p>"
+ "<w:p><w:r><w:t>C</w:t><w:br/><w:t>D</w:t>"
+ "<w:cr/><w:t>E</w:t></w:r></w:p>"
+ )
+ for name, template in (
+ ("modern", _DOCX_MODERN_TEXT_BOX),
+ ("with_fallback", _DOCX_TEXT_BOX_WITH_VML_FALLBACK),
+ ):
+ with self.subTest(shape=name):
+ template = template.replace(
+ "<w:p><w:r><w:t>{inside}</w:t></w:r></w:p>",
+ paragraphs,
+ )
+ sections = await WordParser(include_image=False).parse(
+ _make_docx_with_text_box(template, before=""),
+ "demo.docx",
+ )
+ self.assertEqual(
+ [section.model_dump() for section in sections],
+ _expected_word_sections("A\tB\nC\nD\nE\nAfter"),
+ )
+
async def test_inline_separators_with_text_box_branches(self) -> None:
"""Preserve inline separators while reading each text box once."""
for name, template in (