diff --git a/packages/markitdown/src/markitdown/converters/_markdownify.py b/packages/markitdown/src/markitdown/converters/_markdownify.py
index 19e8a2984..4631d8a78 100644
--- a/packages/markitdown/src/markitdown/converters/_markdownify.py
+++ b/packages/markitdown/src/markitdown/converters/_markdownify.py
@@ -2,7 +2,18 @@
import markdownify
from typing import Any, Optional
-from urllib.parse import quote, unquote, urlparse, urlunparse
+from urllib.parse import quote, urlparse, urlunparse
+
+
+def _quote_path_preserving_percent_encoded_octets(path: str) -> str:
+ parts = []
+ start = 0
+ for match in re.finditer(r"%[0-9A-Fa-f]{2}", path):
+ parts.append(quote(path[start : match.start()], safe="/"))
+ parts.append(match.group(0))
+ start = match.end()
+ parts.append(quote(path[start:], safe="/"))
+ return "".join(parts)
class _CustomMarkdownify(markdownify.MarkdownConverter):
@@ -60,7 +71,13 @@ def convert_a(
parsed_url = urlparse(href) # type: ignore
if parsed_url.scheme and parsed_url.scheme.lower() not in ["http", "https", "file"]: # type: ignore
return "%s%s%s" % (prefix, text, suffix)
- href = urlunparse(parsed_url._replace(path=quote(unquote(parsed_url.path)))) # type: ignore
+ href = urlunparse(
+ parsed_url._replace(
+ path=_quote_path_preserving_percent_encoded_octets(
+ parsed_url.path
+ )
+ )
+ ) # type: ignore
except ValueError: # It's not clear if this ever gets thrown
return "%s%s%s" % (prefix, text, suffix)
diff --git a/packages/markitdown/tests/test_module_misc.py b/packages/markitdown/tests/test_module_misc.py
index 4d62e4919..be20757ef 100644
--- a/packages/markitdown/tests/test_module_misc.py
+++ b/packages/markitdown/tests/test_module_misc.py
@@ -288,6 +288,30 @@ def test_input_as_strings() -> None:
assert "# Test" in result.text_content
+def test_html_preserves_percent_encoded_href_octets() -> None:
+ markitdown = MarkItDown()
+ href = "https://abc.com/hist/%a5%c8%a5%c3%a5%d7%a5%da%a1%bc%a5%b8"
+ input_data = f'example'
+
+ result = markitdown.convert_stream(
+ io.BytesIO(input_data.encode("utf-8")),
+ file_extension=".html",
+ )
+
+ assert f"[example]({href})" in result.text_content
+ assert "%EF%BF%BD" not in result.text_content
+
+ raw_percent_href = "https://abc.com/hist/100% ready.html"
+ input_data = f'raw'
+
+ result = markitdown.convert_stream(
+ io.BytesIO(input_data.encode("utf-8")),
+ file_extension=".html",
+ )
+
+ assert "[raw](https://abc.com/hist/100%25%20ready.html)" in result.text_content
+
+
def test_deeply_nested_html_fallback() -> None:
"""Large, deeply nested HTML should fall back to plain-text extraction
instead of silently returning unconverted HTML (issue #1636).