fix: normalize Nemotron formula delimiters before export

docling/utils/nemotron_parse_utils.py

@@ -31,6 +31,7 @@
 from docling.backend.md_backend import MarkdownDocumentBackend
 from docling.datamodel.base_models import InputFormat
 from docling.datamodel.document import InputDocument
+from docling.utils.vlm_utils import strip_math_delimiters
 
 _log = logging.getLogger(__name__)
 
@@ -321,7 +322,7 @@ def parse_nemotron_parse_v2(
         else:
             label = _TEXT_LABELS.get(region.label, DocItemLabel.TEXT)
             text = (
-                raw_text
+                strip_math_delimiters(raw_text)
                 if label == DocItemLabel.FORMULA
                 else _markdown_to_text(raw_text)
             )

docling/utils/vlm_utils.py

@@ -71,6 +71,19 @@ def parse_markdown_heading(text: str) -> tuple[str, int | None]:
     return match.group(2), len(match.group(1))
 
 
+def strip_math_delimiters(text: str) -> str:
+    """Remove one matching outer math pair without interpreting its LaTeX."""
+    text = text.strip()
+    for opening, closing in (("$$", "$$"), (r"\[", r"\]"), (r"\(", r"\)"), ("$", "$")):
+        if (
+            len(text) >= len(opening) + len(closing)
+            and text.startswith(opening)
+            and text.endswith(closing)
+        ):
+            return text[len(opening) : -len(closing)].strip()
+    return text
+
+
 def strip_stop_strings(texts: list[str], stop_strings: list[str]) -> list[str]:
     """Strip stop strings from decoded texts.
 

tach.toml

@@ -185,7 +185,7 @@ depends_on = []
 [[modules]]
 path = "docling.utils.nemotron_parse_utils"
 layer = "core"
-depends_on = ["docling.backend", "docling.datamodel"]
+depends_on = ["docling.backend", "docling.datamodel", "docling.utils"]
 
 [[modules]]
 path = "docling.utils"

tests/test_nemotron_formula_exports.py

@@ -0,0 +1,34 @@
+# SPDX-FileCopyrightText: The Docling Contributors
+# SPDX-License-Identifier: MIT
+
+import pytest
+from docling_core.types.doc import FormulaItem, Size
+
+from docling.utils.nemotron_parse_utils import parse_nemotron_parse_v2
+
+
+@pytest.mark.parametrize(
+    "opening, closing",
+    [("$", "$"), ("$$", "$$"), (r"\(", r"\)"), (r"\[", r"\]"), ("", "")],
+)
+def test_formula_export_has_one_delimiter_pair(opening: str, closing: str) -> None:
+    latex = r"\left\{x_1 * y_2 : x > 0\right\}"
+    raw = f"{opening}{latex}{closing}"
+    source = f"<x_0.2><y_0.2>{raw}<x_0.8><y_0.4><class_Formula>"
+    size = Size(width=600, height=800)
+    doc = parse_nemotron_parse_v2(source, size, size, page_no=1)
+    formula = doc.texts[0]
+    assert isinstance(formula, FormulaItem)
+    assert formula.text == latex
+    assert formula.orig == raw
+    assert formula.prov
+    assert doc.export_to_markdown() == f"$${latex}$$"
+
+
+def test_multiline_formula_keeps_latex_and_line_breaks() -> None:
+    latex = "\\begin{aligned}\nx &= y \\\\\nz &= w\n\\end{aligned}"
+    source = f"<x_0.2><y_0.2>$$\n{latex}\n$$<x_0.8><y_0.4><class_Formula>"
+    size = Size(width=600, height=800)
+    doc = parse_nemotron_parse_v2(source, size, size, page_no=1)
+    assert doc.texts[0].text == latex
+    assert doc.export_to_markdown() == f"$${latex}$$"

tests/test_nemotron_parse_v2.py

@@ -183,7 +183,7 @@ def test_parse_nemotron_markdown_and_semantic_classes() -> None:
         item.text for item in document.texts if item.label == DocItemLabel.LIST_ITEM
     ] == ["First item", "Second item"]
     assert any(
-        item.label == DocItemLabel.FORMULA and item.text == r"\(x^2\)"
+        item.label == DocItemLabel.FORMULA and item.text == r"x^2"
         for item in document.texts
     )
     assert len(document.pictures) == 2