/**
* Utilities for parsing LLM output into FOL formulas.
*/
export interface ParsedOutput {
premises: string[];
conclusion?: string;
}
/**
* Parse raw LLM output to extract logical formulas.
* Handles code blocks, conversational filler, and conclusion markers.
*/
export function parseLLMOutput(rawOutput: string): ParsedOutput {
const codeBlockMatch = rawOutput.match(/```(?:prolog|logic|text)?\s*([\s\S]*?)```/);
const contentToParse = codeBlockMatch ? codeBlockMatch[1] : rawOutput;
const lines = contentToParse
.split('\n')
.map(l => l.trim())
.filter(l => l && !l.startsWith('%') && !l.startsWith('#') && !l.startsWith('//')); // comments
const premises: string[] = [];
let conclusion: string | undefined;
for (const line of lines) {
// Check for conclusion markers
if (/^(conclusion:|prove:|goal:)\s*/i.test(line)) {
conclusion = line.replace(/^(conclusion:|prove:|goal:)\s*/i, '');
continue;
}
// Heuristic filtering for conversational lines if we didn't find a code block
// If we found a code block, we assume the LLM put valid stuff in it.
// If we are parsing raw text, we need to be careful.
if (!codeBlockMatch) {
// Skip lines that end with ':' (often headers like "Here is the logic:")
if (line.endsWith(':')) continue;
// Skip lines that look like sentences but don't have logic symbols or parens
// (Very rough heuristic: must have '(', ')', '=', '->', '<->', '&', '|', 'all ', 'exists ', or be very short assignment)
const hasLogicChars = /[()=&|<>]|all\s|exists\s/.test(line);
// If it doesn't have logic chars, and has spaces (implies sentence), skip it.
// Allows "P(a)" but skips "Sure thing."
if (!hasLogicChars && line.includes(' ')) {
continue;
}
}
premises.push(line);
}
return {
premises,
conclusion
};
}
