gepa-spo/dist/strategy.js

Genetic-Pareto prompt optimizer to evolve system prompts from a few rollouts with modular support and intelligent crossover

import { clampNumber, objectOfNumberKeys, safeParseWithSchema } from './json.js';
/**
 * Ask the LLM to rate each strategy's applicability to the task corpus.
 * Returns strategies filtered by threshold and optional topK.
 */
export async function prefilterStrategies(llm, strategies, taskPreview, opts) {
    if (strategies.length === 0)
        return { kept: [], scores: {}, raw: '' };
    const ids = strategies.map(s => s.id);
    const prompt = [
        'You are rating strategy hints for applicability to the following representative tasks.',
        'Return STRICT JSON: a map from strategy id to a number in [0,1] where higher means more applicable.',
        'Unknowns should be 0. Base strictly on the provided tasks. No extra keys. No prose.',
        '',
        'Tasks preview:',
        ...taskPreview.slice(0, 8).map((t, i) => `${i + 1}. ${t.slice(0, 280)}`),
        '',
        'Strategies:',
        ...strategies.map(s => `- ${s.id}: ${s.hint.slice(0, 240)}`),
        '',
        'Output JSON only.'
    ].join('\n');
    const raw = await llm.complete(prompt);
    const schema = objectOfNumberKeys(ids);
    const parsed = safeParseWithSchema(raw, schema);
    const scores = {};
    for (const id of ids)
        scores[id] = clampNumber(parsed[id], 0, 1, 0);
    const filtered = strategies
        .map(s => ({ s, score: scores[s.id] ?? 0 }))
        .filter(x => x.score >= opts.threshold)
        .sort((a, b) => b.score - a.score);
    const topK = opts.topK && opts.topK > 0 ? filtered.slice(0, opts.topK) : filtered;
    return { kept: topK.map(x => x.s), scores, raw };
}