Optional wolfxl reader for .xlsx files in ExcelExtractor #44848

Activity

1. crazywoola

   added

   1.18.0Bugs from 1.18.0Bugs from 1.18.0

   on Oct 30, 2026
2. xz567001-afk commented on Oct 30, 2026

   

   xz567001-afk

   on Oct 30, 2026 – with ChatGPT Codex Connector

   More actions

   One concrete compatibility gate before deciding the packaging policy: exercise an image-only row through the real installed reader, including locked 2.0.18.

   The draft's real-workbook test covers scalar cells; its image tests replace load_workbook with _FakeWorkbook. They don't verify the optional reader's drawing objects. The extractor treats a missing _images attribute as no images, so an otherwise blank image row could produce no document without raising.

   A small .xlsx fixture with an embedded PNG, real reader selection, and test storage/database could assert the markdown image, one upload, and retry reuse in both dependency environments. This would complement the disclosed 2.0.17 corpus results; I haven't run either backend or observed image loss.

   dify/api/tests/unit_tests/core/rag/extractor/test_excel_extractor.py

   Lines 93 to 271 in 2542aaf

   	def test_installed_xlsx_reader(self, tmp_path):
   	import openpyxl
   	if importlib.util.find_spec("wolfxl") is None:
   	assert excel_module.load_workbook is openpyxl.load_workbook
   	else:
   	assert excel_module.load_workbook is importlib.import_module("wolfxl").load_workbook
   	path = tmp_path / "reader.xlsx"
   	workbook = openpyxl.Workbook()
   	sheet = workbook.active
   	assert sheet is not None
   	sheet.append(["Name", "Count"])
   	sheet.append(["Alice", 3])
   	workbook.save(path)
   	workbook.close()
   	documents = ExcelExtractor(str(path)).extract()
   	assert [document.page_content for document in documents] == ['"Name":"Alice";"Count":"3"']
   	def test_extract_xlsx_with_hyperlinks_and_sheet_skip(self, monkeypatch: pytest.MonkeyPatch):
   	sheet_with_data = _FakeSheet(
   	header_rows=[("Name", "Link")],
   	data_rows=[
   	(_FakeCell("Alice"), _FakeCell("Doc", hyperlink=SimpleNamespace(target="https://example.com/doc"))),
   	(_FakeCell(None), _FakeCell(123)),
   	(_FakeCell(None), _FakeCell(None)),
   	],
   	)
   	empty_sheet = _FakeSheet(header_rows=[(None, None)], data_rows=[])
   	workbook = _FakeWorkbook({"Data": sheet_with_data, "Empty": empty_sheet})
   	monkeypatch.setattr(excel_module, "load_workbook", lambda *args, **kwargs: workbook)
   	extractor = ExcelExtractor("/tmp/sample.xlsx")
   	docs = extractor.extract()
   	assert workbook.closed is True
   	assert len(docs) == 2
   	assert docs[0].page_content == '"Name":"Alice";"Link":"[Doc](https://example.com/doc)"'
   	assert docs[1].page_content == '"Name":"";"Link":"123"'
   	assert all(doc.metadata["source"] == "/tmp/sample.xlsx" for doc in docs)
   	def test_extract_xlsx_turns_embedded_images_into_markdown_links(
   	self,
   	monkeypatch: pytest.MonkeyPatch,
   	database_session_maker: sessionmaker[Session],
   	):
   	image_bytes = b"\x89PNG\r\n\x1a\nexcel-image"
   	sheet = _FakeSheet(
   	header_rows=[("Question", "Answer", "Image")],
   	data_rows=[
   	(_FakeCell("Q1"), _FakeCell("A1"), _FakeCell(None)),
   	(_FakeCell("Q2"), _FakeCell("A2"), _FakeCell(None)),
   	],
   	images=[
   	_FakeImage(image_bytes, row=1, col=2),
   	_FakeImage(image_bytes, row=1, col=2),
   	],
   	)
   	workbook = _FakeWorkbook({"Data": sheet})
   	monkeypatch.setattr(excel_module, "load_workbook", lambda *args, **kwargs: workbook)
   	saves = _patch_image_persistence(monkeypatch)
   	extractor = ExcelExtractor(
   	"/tmp/sample.xlsx",
   	tenant_id="tenant-1",
   	user_id="user-1",
   	source_file_id="source-file-1",
   	)
   	docs = extractor.extract()
   	upload_files = _get_upload_files(database_session_maker)
   	assert workbook.closed is True
   	assert len(docs) == 2
   	assert len(upload_files) == 1
   	assert docs[0].page_content == (
   	'"Question":"Q1";"Answer":"A1";'
   	f'"Image":" '
   	f'"'
   	)
   	assert docs[1].page_content == '"Question":"Q2";"Answer":"A2";"Image":""'
   	assert len(saves) == 1
   	assert saves[0][0].startswith("image_files/tenant-1/source-file-1/")
   	assert saves[0][0].endswith(".png")
   	assert saves[0][1] == image_bytes
   	assert upload_files[0].tenant_id == "tenant-1"
   	assert upload_files[0].key == saves[0][0]
   	assert upload_files[0].used is True
   	def test_extract_xlsx_keeps_rows_with_only_embedded_images(
   	self,
   	monkeypatch: pytest.MonkeyPatch,
   	database_session_maker: sessionmaker[Session],
   	):
   	image_bytes = b"\x89PNG\r\n\x1a\nimage-only-row"
   	sheet = _FakeSheet(
   	header_rows=[("Question", "Answer", "Image")],
   	data_rows=[
   	(_FakeCell(None), _FakeCell(None), _FakeCell(None)),
   	(_FakeCell(None), _FakeCell(None), _FakeCell(None)),
   	],
   	images=[_FakeImage(image_bytes, row=1, col=2)],
   	)
   	workbook = _FakeWorkbook({"Data": sheet})
   	monkeypatch.setattr(excel_module, "load_workbook", lambda *args, **kwargs: workbook)
   	saves = _patch_image_persistence(monkeypatch)
   	extractor = ExcelExtractor(
   	"/tmp/sample.xlsx",
   	tenant_id="tenant-1",
   	user_id="user-1",
   	source_file_id="source-file-1",
   	)
   	docs = extractor.extract()
   	upload_files = _get_upload_files(database_session_maker)
   	assert workbook.closed is True
   	assert len(docs) == 1
   	assert len(upload_files) == 1
   	assert docs[0].page_content == (
   	f'"Question":"";"Answer":"";"Image":""'
   	)
   	assert len(saves) == 1
   	def test_extract_xlsx_reuses_existing_embedded_image_uploads_on_retry(
   	self,
   	monkeypatch: pytest.MonkeyPatch,
   	database_session_maker: sessionmaker[Session],
   	):
   	image_bytes = b"\x89PNG\r\n\x1a\nretry-safe-image"
   	workbooks = [
   	_FakeWorkbook(
   	{
   	"Data": _FakeSheet(
   	header_rows=[("Question", "Answer", "Image")],
   	data_rows=[(_FakeCell("Q1"), _FakeCell("A1"), _FakeCell(None))],
   	images=[_FakeImage(image_bytes, row=1, col=2)],
   	)
   	}
   	),
   	_FakeWorkbook(
   	{
   	"Data": _FakeSheet(
   	header_rows=[("Question", "Answer", "Image")],
   	data_rows=[(_FakeCell("Q1"), _FakeCell("A1"), _FakeCell(None))],
   	images=[_FakeImage(image_bytes, row=1, col=2)],
   	)
   	}
   	),
   	]
   	monkeypatch.setattr(excel_module, "load_workbook", lambda *args, **kwargs: workbooks.pop(0))
   	saves = _patch_image_persistence(monkeypatch)
   	extractor = ExcelExtractor(
   	"/tmp/sample.xlsx",
   	tenant_id="tenant-1",
   	user_id="user-1",
   	source_file_id="source-file-1",
   	)
   	first_docs = extractor.extract()
   	second_docs = extractor.extract()
   	upload_files = _get_upload_files(database_session_maker)
   	assert len(upload_files) == 1
   	expected_page_content = (
   	'"Question":"Q1";"Answer":"A1";'
   	f'"Image":""'
   	)
   	assert first_docs[0].page_content == expected_page_content
   	assert second_docs[0].page_content == expected_page_content
   	assert len(saves) == 1
   	def test_extract_xls_path(self, monkeypatch: pytest.MonkeyPatch):
   	class FakeExcelFile:
   	sheet_names = ["Sheet1"]
   	def parse(self, sheet_name):

   dify/api/core/rag/extractor/excel_extractor.py

   Lines 129 to 176 in 91a94a9

   	if row_has_content and page_content:
   	documents.append(
   	Document(page_content=";".join(page_content), metadata={"source": self._file_path})
   	)
   	finally:
   	wb.close()
   	elif file_extension == ".xls":
   	excel_file = pd.ExcelFile(self._file_path, engine="xlrd")
   	for excel_sheet_name in excel_file.sheet_names:
   	df = excel_file.parse(sheet_name=excel_sheet_name)
   	df.dropna(how="all", inplace=True)
   	for _, series_row in df.iterrows():
   	page_content = []
   	for k, v in series_row.items():
   	if pd.notna(v):
   	# Escape embedded double quotes like the .xlsx branch
   	# does, so quoted cell values do not corrupt the row.
   	value = str(v).replace('"', '\\"')
   	page_content.append(f'"{k}":"{value}"')
   	documents.append(
   	Document(page_content=";".join(page_content), metadata={"source": self._file_path})
   	)
   	else:
   	raise ValueError(f"Unsupported file extension: {file_extension}")
   	return documents
   	def _extract_images_from_sheet(
   	self, sheet_name: str, sheet, valid_columns: set[int], min_row: int
   	) -> dict[tuple[int, int], list[str]]:
   	"""
   	Extract embedded worksheet images and map them to their anchor cell.
   	Images are stored with deterministic keys derived from the source upload file,
   	sheet, anchor cell, and content hash so retried tasks can reuse the same
   	UploadFile rows and storage objects.
   	"""
   	if not self._tenant_id or not self._user_id or not self._source_file_id:
   	return {}
   	images = getattr(sheet, "_images", None) or []
   	image_candidates: list[SheetImageCandidate] = []
   	for image in images:
   	marker = getattr(getattr(image, "anchor", None), "_from", None)
   	row_idx = getattr(marker, "row", None)
3. wolfiesch commented on Oct 31, 2026

   

   wolfiesch

   on Oct 31, 2026

   Author

   More actions

   Thanks, the test you described found a real gap. The cause is row iteration, not image loading.

   Both readers return the embedded image with the same anchor, format and bytes. ExcelExtractor only iterated rows up to sheet.max_row, and a row that holds only an image has no cells, so it was skipped:

   • openpyxl: the 10-row header scan (iter_rows(max_row=10)) creates empty cells and pads max_row to 10, so image-only rows after row 10 are dropped on current main.
   • wolfxl 2.0.18: the scan does not pad max_row, so an image-only row directly below the data was dropped as well.

   #44849 now runs the row loop to the last image anchor (sheets without images iterate as before) and adds a real-workbook test: embedded PNGs anchored at C3 and C12, the installed reader, test storage and database. It asserts the markdown links, one upload per image, and upload reuse on a second extract(). It fails without the fix in both environments (openpyxl misses row 12; wolfxl misses rows 3 and 12) and passes with it.