gepa-spo/dist/seeding.js

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

import { proposeNewSystem, summarizeTraces } from './reflection.js';
/** Generate seeded candidates with top-K strategies and keep the top few by average judge score */
export async function seedPopulation({ seed, screen, strategies, K, execute, muf, llm, budgetLeft, mufCosts = true }) {
    const out = [{ system: seed.system ?? '', _via: 'seed', _uplift: 0 }];
    let usedCalls = 0;
    for (const s of strategies.slice(0, K)) {
        if (budgetLeft !== undefined && budgetLeft <= 0)
            break;
        // Execute the seed system on screen items to get actual examples with traces
        const examples = [];
        for (const item of screen) {
            if (budgetLeft !== undefined && budgetLeft <= 0)
                break;
            const { output, traces } = await execute({ candidate: seed, item });
            const f = await muf({ item, output, traces: traces ?? null });
            const execTrace = summarizeTraces(traces, 1000);
            examples.push({
                user: item.user,
                output,
                feedback: f.feedbackText,
                ...(execTrace && { execTrace })
            });
            // Count execute always, plus muf when configured
            const delta = 1 + (mufCosts ? 1 : 0);
            usedCalls += delta;
            if (budgetLeft !== undefined)
                budgetLeft -= delta;
        }
        const sys2 = await proposeNewSystem(llm, seed.system ?? '', examples, s.hint);
        usedCalls += 1;
        if (budgetLeft !== undefined)
            budgetLeft -= 1; // propose call
        const scores = [];
        for (const item of screen) {
            if (budgetLeft !== undefined && budgetLeft <= 0)
                break;
            const { output, traces } = await execute({ candidate: { system: sys2 }, item });
            // Judge may or may not count to usedCalls
            const f = await muf({ item, output, traces: traces ?? null });
            scores.push(f.score);
            // Count execute always, plus muf when configured
            const delta = 1 + (mufCosts ? 1 : 0);
            usedCalls += delta;
            if (budgetLeft !== undefined)
                budgetLeft -= delta;
        }
        out.push({ system: sys2, _via: s.id, _uplift: avg(scores) });
    }
    out.sort((a, b) => b._uplift - a._uplift);
    return { candidates: out.slice(0, Math.min(5, out.length)), usedCalls };
}
const avg = (xs) => (xs.length ? xs.reduce((a, b) => a + b, 0) / xs.length : 0);