fix: drop processing instructions in HTML parser

assinscreedFC · assinscreedFC · commit 0157bb6254c0 · 2026-06-05T13:49:50.000+02:00
partition_html() crashed with 'lxml.etree._ProcessingInstruction object has no attribute is_phrasing' when the document contained a processing instruction such as '<?xml ...?>'. The HTML parser's element-class lookup only maps real elements, so a processing instruction reaches the tree as a plain lxml _ProcessingInstruction node without the parser's custom interface, then crashes the phrasing/flow iteration. Pass remove_pis=True to the HTMLParser so processing instructions are stripped during parsing, mirroring the existing remove_comments=True. They carry no rendered text, so dropping them matches browser behavior. Fixes #4358
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,3 +1,9 @@
+## 0.22.32
+
+### Fixes
+
+- **Fix crash on processing instructions in HTML**: `partition_html()` no longer raises `AttributeError: 'lxml.etree._ProcessingInstruction' object has no attribute 'is_phrasing'` when the document contains a processing instruction such as `<?xml ...?>`. Processing instructions are now dropped during parsing, consistent with how comments are already handled.
+
 ## 0.22.31
 
 ### Enhancements
diff --git a/test_unstructured/partition/html/test_partition.py b/test_unstructured/partition/html/test_partition.py
@@ -203,6 +203,15 @@ def test_partition_html_processes_chinese_chracters():
     assert elements[0].text == "每日新闻"
 
 
+def test_partition_html_ignores_processing_instructions():
+    """A processing-instruction like `<?xml ...?>` must not crash the parser (see #4358)."""
+    html_text = (
+        '<html><body><div>\n<?xml version="1.0" encoding="UTF-8"?>\n'
+        "<p>hello world</p></div></body></html>"
+    )
+    assert partition_html(text=html_text) == [Text("hello world")]
+
+
 def test_emoji_appears_with_emoji_utf8_code():
     assert partition_html(text='<html charset="utf-8"><p>Hello &#128512;</p></html>') == [
         Text("Hello 😀")
diff --git a/unstructured/__version__.py b/unstructured/__version__.py
@@ -1 +1 @@
-__version__ = "0.22.31"  # pragma: no cover
+__version__ = "0.22.32"  # pragma: no cover
diff --git a/unstructured/partition/html/parser.py b/unstructured/partition/html/parser.py
@@ -945,7 +945,10 @@ def derive_element_type_from_text(text: str) -> type[Text] | None:
 # ------------------------------------------------------------------------------------------------
 
 
-html_parser = etree.HTMLParser(remove_comments=True)
+# -- `remove_pis` drops processing-instructions (e.g. `<?xml ...?>`); like comments they
+# -- carry no rendered text and otherwise reach the parser as `_ProcessingInstruction` nodes
+# -- that lack the custom-element interface (`.is_phrasing`), crashing iteration.
+html_parser = etree.HTMLParser(remove_comments=True, remove_pis=True)
 # -- elements that don't have a registered class get DefaultElement --
 fallback = etree.ElementDefaultClassLookup(element=DefaultElement)
 # -- elements that do have a registered class are assigned that class via lookup --

Original file line number	Diff line number	Diff line change
`@@ -1 +1 @@`
`1`		`-__version__ = "0.22.31" # pragma: no cover`
	`1`	`+__version__ = "0.22.32" # pragma: no cover`