A raw #fragment becomes part of the converted content. With base64, it can silently add bytes.

from markitdown import MarkItDown

md = MarkItDown(enable_plugins=False)
print(repr(md.convert_uri('data:text/plain,Hello#section').text_content))
# Actual: 'Hello#section'; expected: 'Hello'

print(repr(md.convert_uri('data:text/plain;base64,SGVsbG8h#AAAA').text_content))
# Actual: 'Hello!\x00\x00\x00'; expected: 'Hello!'

Reproduced on main 4cc9fa17653d695d64fb9eee5b33d4de55ff84e8, macOS, Python 3.12.15. It also occurs on #2682, which fixes a different decoding problem.

parse_data_uri decodes everything after the comma, including the fragment. Strip the raw fragment before percent-decoding: %23 must still produce a literal #, and ? must remain in the payload. The WHATWG Fetch data URL processor excludes fragments; Node's fetch returns the expected text for both examples.

Parser and public API regression tests reproduced both failures. Escaped-hash and query controls passed. I searched existing issues and PRs and didn't find this failure covered.

Investigated with Hermes AI assistance.