guidellm.data.preprocessors.tool_calling

@PreprocessorRegistry.register("tool_calling_message_extractor")
class ToolCallingMessageExtractor(DatasetPreprocessor):
    """Extract user prompts, system prompts, and tool responses from messages.

    Many tool calling datasets (e.g. ``madroid/glaive-function-calling-openai``)
    store conversations as a ``messages`` column containing an array of
    message dicts. This preprocessor supports both OpenAI format
    (``role``/``content``) and ShareGPT format (``from``/``value``).

    It replaces the ``text_column`` value with the extracted user content,
    populates ``prefix_column`` with the system prompt when present, and
    populates ``tool_response_column`` with tool response content.

    Usage::

        guidellm benchmark run \\
            --data '{"kind": "hf", "source": "..."}' \\
            --data-column-mapper '{"kind": "generative_column_mapper",
                "column_mappings": {"text_column": "messages"}}' \\
            --data-preprocessor kind=tool_calling_message_extractor
    """

    def __init__(self, config: ToolCallingMessageExtractorArgs, **_: Any) -> None:
        pass

    def __call__(  # noqa: C901
        self, items: list[dict[str, Any]]
    ) -> list[dict[str, Any]]:
        for item in items:
            text_values = item.get("text_column")
            if not text_values or not isinstance(text_values, list):
                continue

            new_texts: list[str] = []
            prefixes: list[str] = []
            tool_responses: list[str] = []

            for value in text_values:
                if isinstance(value, list):
                    user_parts, system_parts, tool_parts = _extract_from_messages(value)
                    if user_parts:
                        new_texts.append(" ".join(user_parts))
                    if system_parts:
                        prefixes.append(" ".join(system_parts))
                    tool_responses.extend(tool_parts)
                elif isinstance(value, str):
                    new_texts.append(value)

            if new_texts:
                item["text_column"] = new_texts
            if prefixes:
                item.setdefault("prefix_column", []).extend(prefixes)
            if tool_responses:
                item.setdefault("tool_response_column", []).extend(tool_responses)

        return items