DOCX math: a delimiter with several arguments keeps only the last one #2573

When a Word equation contains a delimiter (<m:d>) with more than one argument, MarkItDown keeps only the last argument and drops the rest without any warning. The separator character (<m:sepChr>) is never read either.

Word writes this structure for the bracket templates "with separators" (Insert > Equation > Bracket), for example (a|b), {x|x>0} and bra-ket notation.

Repro

from xml.etree import ElementTree as ET
from markitdown.converter_utils.docx.math.omml import OMML_NS, oMath2Latex

M = f'xmlns:m="{OMML_NS[1:-1]}"'

def latex(xml):
    return oMath2Latex(ET.fromstring(f"<m:oMath {M}>{xml}</m:oMath>")).latex

r = lambda t: f"<m:r><m:t>{t}</m:t></m:r>"

# Explicit separator
print(latex(r("f") + '<m:d><m:dPr><m:sepChr m:val=","/></m:dPr>'
            f"<m:e>{r('x')}</m:e><m:e>{r('y')}</m:e></m:d>"))
# -> f\left(y\right)

# Set-builder braces, default separator (no sepChr)
print(latex('<m:d><m:dPr><m:begChr m:val="{"/><m:endChr m:val="}"/></m:dPr>'
            f"<m:e>{r('x')}</m:e><m:e>{r('x>0')}</m:e></m:d>"))
# -> \left\{x>0\right\}

End to end, adding the first equation to a copy of tests/test_files/equations.docx makes MarkItDown().convert() emit $$f\left(y\right)$$.

Expected vs actual

In the OMML schema (ECMA-376 / ISO/IEC 29500-1, CT_D), <m:e> has maxOccurs="unbounded". The sepChr element "specifies the character that separates base arguments e in the delimiter object d". When the element is omitted, the separator is a vertical bar (U+2502). When it is present without m:val, there is no separator. See the Open XML SDK SeparatorChar reference, which reproduces the spec text. Dropping arguments silently loses document content.

Cause

In packages/markitdown/src/markitdown/converter_utils/docx/math/omml.py, oMath2Latex.do_d reads its children with process_children_dict(). That function returns a dict keyed by tag name, so each <m:e> overwrites the previous one and c_dict["e"] holds only the last argument. Separately, Pr.__val_tags (and Pr.tag2meth) do not include sepChr, so pr.sepChr is always None.

Suggested fix

In do_d, collect every <m:e> in order with process_children_list(elm, include=("e",)). Have Pr read sepChr, with an empty value when m:val is missing. Then join the arguments with the separator, falling back to | when sepChr is absent. A single-argument delimiter gives the same output as before.

Environment

• markitdown main at 1f9530a (0.1.8)
• Python 3.13, macOS

AI assistance: drafted with an AI coding assistant (Claude). The reproduction above was run locally against the current default branch.