llm-response-schema-lite-py

Tiny schema validator for structured LLM responses. Two schema formats: a one-line shape spec ({"name": "str", "age": "int"}) and the JS-sibling rule format ({"name": {"type": "str", "required": True, "enum": [...]}}). Returns a (valid, errors) result you can hand straight back to the LLM as feedback. Zero runtime dependencies.

Python port of @mukundakatta/llm-response-schema-lite.

Install

pip install llm-response-schema-lite-py

Quick start

Shape format (Python-friendly, 1-liner)

from llm_response_schema_lite import validate

schema = {"name": "str", "age": "int", "tags": "list", "email": "str?"}

validate({"name": "Ada", "age": 36, "tags": ["scientist"]}, schema)
# ValidationResult(valid=True, value={...}, errors=[])

validate({"name": "Ada", "age": "thirty-six"}, schema)
# ValidationResult(valid=False, errors=[
#   {"path": "age", "message": "expected_int"},
#   {"path": "tags", "message": "required"}
# ])

Suffix a type with ? to mark it optional. Supported types: str, int, float, bool, list, dict (and JS aliases: string, number, boolean, array, object).

JS-sibling rule format (full feature parity)

schema = {
    "status": {"type": "str", "required": True, "enum": ["ok", "error"]},
    "code":   {"type": "int", "required": False},
}

validate({"status": "ok", "code": 200}, schema)        # valid
validate({"status": "weird"}, schema)                  # not_in_enum
validate({}, schema)                                   # required

Parse JSON + validate in one shot

from llm_response_schema_lite import parse_and_validate

result = parse_and_validate('{"name": "Ada"}', {"name": "str"})
result.valid   # True
result.value   # {"name": "Ada"}

If JSON parsing fails, you get back ValidationResult(valid=False, value=None, errors=[{"path": "$", "message": "invalid_json"}]).

API

• validate(value, schema) -> ValidationResult
• parse_and_validate(json_str, schema) -> ValidationResult
• validate_response(value, schema) -- JS-parity alias of validate.

ValidationResult is a dataclass with valid: bool, value: Any | None, errors: list[dict].

License

MIT