Bug

When exporting tables to markdown, tabulate is called without disable_numparse=True on the first attempt (markdown.py#L553, document.py#L2268). The disable_numparse=True fallback is only triggered when a ValueError is raised.

By default, tabulate auto-detects numeric strings, parses them as Python float, and reformats them — silently losing precision. This is particularly problematic for financial documents where exact decimal values matter.

Root cause: In docling_core/transforms/serializer/markdown.py (MarkdownTableSerializer.serialize), line 553:

try:
    table_text = tabulate(rows[1:], headers=rows[0], tablefmt="github")
except ValueError:
    table_text = tabulate(
        rows[1:],
        headers=rows[0],
        tablefmt="github",
        disable_numparse=True,
    )

The same pattern exists in the legacy path at docling_core/types/doc/document.py line 2268.

The problem is that tabulate successfully parses the numeric strings (no ValueError), but reformats them with reduced precision. The fix is straightforward: always pass disable_numparse=True to tabulate in both call sites, as a document converter should never silently alter source data. Optionally, expose a disable_numparse parameter on MarkdownParams (default True) for users who explicitly want tabulate's numeric alignment.

Steps to reproduce

1. Create or use a DOCX file containing a table with precise numeric values (e.g. 225.8183, 20896.7184)
2. Convert the document using docling:

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("document.docx")
md = result.document.export_to_markdown()
print(md)

3. Observe that numeric values in tables are silently rounded:

All values are truncated to approximately 6 significant figures, which is tabulate's default float formatting behavior.

Docling version

docling 2.81.0, docling-core 2.70.2

Python version

Python 3.14