docling-project/docling · error · NotImplementedError

Label must be either code or formula

Error message

Label must be either code or formula

What it means

The VLM-based code/formula model builds its prompt token from an element label; only 'code' and 'formula' are accepted, returning '<code>' or '<formula>'. Any other label raises NotImplementedError, signalling the caller routed an unsupported element type into this stage. It is a strict contract on the input labels, not a runtime/model failure.

Source

Thrown at docling/models/stages/code_formula/code_formula_vlm_model.py:152

    def _get_prompt(self, label: str) -> str:
        """Construct the prompt for the model based on the element type.

        Args:
            label: The type of input, either 'code' or 'formula'

        Returns:
            The prompt string

        Raises:
            NotImplementedError: If the label is not 'code' or 'formula'
        """
        if label == "code":
            return "<code>"
        elif label == "formula":
            return "<formula>"
        else:
            raise NotImplementedError("Label must be either code or formula")

    def _extract_code_language(self, input_string: str) -> Tuple[str, Optional[str]]:
        """Extract programming language from the beginning of a string.

        Checks if the input string starts with a pattern of the form
        ``<_some_language_>``. If it does, extracts the language string.

        Args:
            input_string: The input string, which may start with ``<_language_>``

        Returns:
            Tuple of (remainder, language) where:
            - remainder is the string after the language tag (or original if no match)
            - language is the extracted language if found, otherwise None
        """
        pattern = r"^<_([^_>]+)_>\s*(.*)"
        match = re.match(pattern, input_string, flags=re.DOTALL)
        if match:

View on GitHub (pinned to 61d76f1ff3)

Solutions

1. Filter element batches to labels 'code' and 'formula' before invoking the model (mirror what the standard pipeline stage does).
2. Normalize labels: assert/convert item.label to the expected lowercase enum value before building the prompt.
3. Add a unit check in your pipeline that the set of labels reaching this model is a subset of {'code','formula'}.

Example fix

# before
query = self._get_query(item.label)  # raises NotImplementedError for 'picture'

# after
if item.label in ("code", "formula"):
    query = self._get_query(item.label)
else:
    _log.debug(f"Skipping item with unsupported label {item.label!r}")
    return

SUPPORTED = {"code", "formula"}
items = [i for i in items if i.label in SUPPORTED]

def is_code_or_formula(label: str) -> bool:
    return label in ("code", "formula")

When it happens

Trigger: Passing elements to GraniteVlmCodeFormulaModel whose label is anything other than the exact strings 'code' or 'formula' — for example 'title', 'list_item', or a mistyped/uppercased label like 'Code'.

Common situations: Custom pipelines that skip the label filter; label enum mismatches after a docling upgrade; reusing items from another backend whose labels were never normalized to docling's enum values.

Related errors

• Label must be either code or formula
• Engine not initialized
• ThreadedDoclingParseDocumentBackend only supports iter_pages
• The parameters vlm_pipeline_model, vlm_pipeline_model_local
• Cannot specify both vlm_pipeline_preset and vlm_pipeline_cus

AI-assisted analysis of docling-project/docling@61d76f1ff3 (2026-08-14). Data as JSON: /api/errors/a59cd747ddc2de98. Report an issue: GitHub.