API: llm

Three-stage pipeline for cleaning and repairing LLM-generated STL output.

Module: stl_parser.llm Import: from stl_parser import clean, repair, validate_llm_output, prompt_template, LLMValidationResult

validate_llm_output()

validate_llm_output(
    raw_text: str,
    schema: Optional[STLSchema] = None
) -> LLMValidationResult

Run the full 3-stage pipeline: clean → repair → parse (+ optional schema validation).

Parameters:

Returns: LLMValidationResult

Example:

from stl_parser import validate_llm_output
result = validate_llm_output("A => B mod(confience=1.5)")
print(result.is_valid, len(result.repairs))

clean()

clean(raw_text: str) -> Tuple[str, List[RepairAction]]

Stage 1: Text-level normalization.

Handles: Code fence extraction, prose stripping, arrow normalization (=> → ->), multi-line merge, markdown escape removal.

Returns: Tuple of (cleaned text, list of repair actions).

repair()

repair(text: str) -> Tuple[str, List[RepairAction]]

Stage 2: Structural repair.

Handles: Missing :: prefix, missing brackets, unquoted strings, out-of-range value clamping, modifier key typo correction.

Returns: Tuple of (repaired text, list of repair actions).

prompt_template()

prompt_template(schema: Optional[STLSchema] = None) -> str

Generate an instruction prompt for LLMs describing STL syntax. If a schema is provided, includes domain-specific constraints.

Returns: String suitable as a system message.

LLMValidationResult

class LLMValidationResult(BaseModel):
    statements: List[Statement] = []
    errors: List[ParseError] = []
    warnings: List[ParseWarning] = []
    is_valid: bool = True
    repairs: List[RepairAction] = []
    cleaned_text: str = ""
    original_text: str = ""
    schema_result: Optional[Any] = None

RepairAction

class RepairAction(BaseModel):
    type: str           # Repair type identifier
    line: Optional[int] # Line number
    original: str       # Text before repair
    repaired: str       # Text after repair
    description: str    # Human-readable description