Parsers

Goal: parse structured model or tactic output without depending on a runtime.

Prerequisites

python -m pip install -e ".[dev]"

Files Used

lllm/
  parsers.py
tests/
  test_parsers.py

Parse

from lllm.parsers import DefaultTagParser


parser = DefaultTagParser(
    required_xml_tags=["answer"],
    required_md_tags=["json"],
    signal_tags=["DONE"],
)

parsed = parser.parse("<answer>Hello</answer>\n```json\n{}\n```\n<DONE>")
assert parsed["xml_tags"]["answer"] == ["Hello"]
assert parsed["signal_tags"]["DONE"] is True

Contract

Parser configuration is Pydantic-schema-compatible, so it can be serialized in runtime config or inspected by tools:

schema = DefaultTagParser.model_json_schema()
assert "required_xml_tags" in schema["properties"]

The parser fails fast with ParseError when required blocks are missing. Repair loops, retries, or partial-output recovery should live in the tactic or runtime adapter that owns the model call.

In A Tactic

The same parser can be attached to native prompts or used directly around Pydantic AI/plain-Python tactic outputs:

from lllm import Tactic
from lllm.parsers import DefaultTagParser


parser = DefaultTagParser(required_xml_tags=["answer"], signal_tags=["DONE"])


class ParseAnswerTactic(Tactic[str, dict[str, object]]):
    name = "parse_answer"
    input_type = str
    output_type = dict[str, object]

    def _run(self, input_value, *, context=None):
        return parser.parse(input_value)


parsed = ParseAnswerTactic().run("<answer>Hello</answer><DONE>")
assert parsed["xml_tags"]["answer"] == ["Hello"]

Verify

python -m pytest tests/test_parsers.py -q

Expected output:

17 passed

Next, attach the parser to a native prompt or wrap Pydantic AI output before returning through a Tactic.