diff --git a/packages/markitdown/src/markitdown/converters/_html_converter.py b/packages/markitdown/src/markitdown/converters/_html_converter.py
index 029b27f57..030cbed42 100644
--- a/packages/markitdown/src/markitdown/converters/_html_converter.py
+++ b/packages/markitdown/src/markitdown/converters/_html_converter.py
@@ -1,4 +1,5 @@
import io
+import re
import warnings
from typing import Any, BinaryIO, Optional
from bs4 import BeautifulSoup
@@ -39,6 +40,50 @@ def accepts(
return False
+ def _detect_html_encoding(self, file_stream: BinaryIO) -> Optional[str]:
+ """Peek at the HTML content to find charset declaration ().
+
+ Follows the HTML5 encoding sniffing algorithm: the tag
+ (or Content-Type http-equiv) is the authoritative source, taking
+ precedence over heuristic detection that is unreliable for non-Latin
+ scripts on some platforms.
+ """
+ cur_pos = file_stream.tell()
+ try:
+ # Read first 4 KB — enough to cover and in any
+ # reasonable HTML document (HTML5 spec uses 1024 bytes minimum).
+ raw = file_stream.read(4096)
+ # Decode as ASCII-superset (Latin-1) so we can scan for meta tags
+ # without committing to the real encoding yet.
+ try:
+ head = raw.decode("ascii")
+ except UnicodeDecodeError:
+ head = raw.decode("latin-1")
+
+ # Pattern 1: (HTML5)
+ m = re.search(
+ r']*\bcharset\s*=\s*["\']?\s*([\w\-\d]+)\s*["\']?',
+ head,
+ re.IGNORECASE,
+ )
+ if m:
+ return m.group(1)
+
+ # Pattern 2:
+ m = re.search(
+ r']*\bhttp-equiv\s*=\s*["\']?Content-Type["\']?[^>]*\b'
+ r'content\s*=\s*["\'][^"\']*\bcharset\s*=\s*([\w\-\d]+)',
+ head,
+ re.IGNORECASE,
+ )
+ if m:
+ return m.group(1)
+
+ return None
+ finally:
+ file_stream.seek(cur_pos)
+
def convert(
self,
file_stream: BinaryIO,
@@ -49,8 +94,21 @@ def convert(
# strict=True raises RecursionError instead of falling back to plain text.
strict: bool = kwargs.pop("strict", False)
+ # Determine encoding — prefer the HTML declaration,
+ # then the stream_info hint, then UTF-8:
+ # - is the HTML5-authoritative source.
+ # - stream_info.charset comes from charset_normalizer which can
+ # mis-detect non-Latin UTF-8 files (e.g. Chinese) on CJK-locale
+ # Windows systems.
+ meta_encoding = self._detect_html_encoding(file_stream)
+ if meta_encoding:
+ encoding = meta_encoding
+ elif stream_info.charset is not None:
+ encoding = stream_info.charset
+ else:
+ encoding = "utf-8"
+
# Parse the stream
- encoding = "utf-8" if stream_info.charset is None else stream_info.charset
soup = BeautifulSoup(file_stream, "html.parser", from_encoding=encoding)
# Remove javascript and style blocks