XMLAgentOutputParser Class — langchain Architecture

Architecture documentation for the XMLAgentOutputParser class in xml.py from the langchain codebase.

Class python

Entity Profile

Dependency Diagram

graph TD
  647b2f55_255b_beae_316f_269d2e0e4bb5["XMLAgentOutputParser"]
  d9685a4b_26b0_bca9_f857_03bb2ffc9dd4["AgentOutputParser"]
  647b2f55_255b_beae_316f_269d2e0e4bb5 -->|extends| d9685a4b_26b0_bca9_f857_03bb2ffc9dd4
  f2ea7756_ae03_24f7_4d97_422e78f632d9["xml.py"]
  647b2f55_255b_beae_316f_269d2e0e4bb5 -->|defined in| f2ea7756_ae03_24f7_4d97_422e78f632d9
  e386861e_a7a7_9278_87ae_0d850022a200["parse()"]
  647b2f55_255b_beae_316f_269d2e0e4bb5 -->|method| e386861e_a7a7_9278_87ae_0d850022a200
  36b2a200_78a8_1292_4aa1_6a0fa119ef2c["get_format_instructions()"]
  647b2f55_255b_beae_316f_269d2e0e4bb5 -->|method| 36b2a200_78a8_1292_4aa1_6a0fa119ef2c
  c80ae0c3_91e5_9078_7dad_70a1d86186f2["_type()"]
  647b2f55_255b_beae_316f_269d2e0e4bb5 -->|method| c80ae0c3_91e5_9078_7dad_70a1d86186f2

Relationship Graph

Source Code

libs/langchain/langchain_classic/agents/output_parsers/xml.py lines 26–126

class XMLAgentOutputParser(AgentOutputParser):
    """Parses tool invocations and final answers from XML-formatted agent output.

    This parser extracts structured information from XML tags to determine whether
    an agent should perform a tool action or provide a final answer. It includes
    built-in escaping support to safely handle tool names and inputs
    containing XML special characters.

    Args:
        escape_format: The escaping format to use when parsing XML content.
            Supports 'minimal' which uses custom delimiters like [[tool]] to replace
            XML tags within content, preventing parsing conflicts.
            Use 'minimal' if using a corresponding encoding format that uses
            the _escape function when formatting the output (e.g., with format_xml).

    Expected formats:
        Tool invocation (returns AgentAction):
            <tool>search</tool>
            <tool_input>what is 2 + 2</tool_input>

        Final answer (returns AgentFinish):
            <final_answer>The answer is 4</final_answer>

    !!! note
        Minimal escaping allows tool names containing XML tags to be safely represented.
        For example, a tool named `search<tool>nested</tool>` would be escaped as
        `search[[tool]]nested[[/tool]]` in the XML and automatically unescaped during
        parsing.

    Raises:
        ValueError: If the input doesn't match either expected XML format or
            contains malformed XML structure.
    """

    escape_format: Literal["minimal"] | None = Field(default="minimal")
    """The format to use for escaping XML characters.

    minimal - uses custom delimiters to replace XML tags within content,
    preventing parsing conflicts. This is the only supported format currently.

    None - no escaping is applied, which may lead to parsing conflicts.
    """

    @override
    def parse(self, text: str) -> AgentAction | AgentFinish:
        # Check for tool invocation first
        tool_matches = re.findall(r"<tool>(.*?)</tool>", text, re.DOTALL)
        if tool_matches:
            if len(tool_matches) != 1:
                msg = (
                    f"Malformed tool invocation: expected exactly one <tool> block, "
                    f"but found {len(tool_matches)}."
                )
                raise ValueError(msg)
            _tool = tool_matches[0]

            # Match optional tool input
            input_matches = re.findall(
                r"<tool_input>(.*?)</tool_input>", text, re.DOTALL
            )
            if len(input_matches) > 1:
                msg = (
                    f"Malformed tool invocation: expected at most one <tool_input> "
                    f"block, but found {len(input_matches)}."
                )
                raise ValueError(msg)
            _tool_input = input_matches[0] if input_matches else ""

            # Unescape if minimal escape format is used
            if self.escape_format == "minimal":
                _tool = _unescape(_tool)
                _tool_input = _unescape(_tool_input)

            return AgentAction(tool=_tool, tool_input=_tool_input, log=text)
        # Check for final answer
        if "<final_answer>" in text and "</final_answer>" in text:
            matches = re.findall(r"<final_answer>(.*?)</final_answer>", text, re.DOTALL)
            if len(matches) != 1:
                msg = (
                    "Malformed output: expected exactly one "
                    "<final_answer>...</final_answer> block."