agentfootprint/dist/esm/patterns/SelfConsistency.js

The explainable agent framework — backtrack a wrong answer to the exact context that caused it (evidence, not guesses). Built on footprintjs.

/**
 * SelfConsistency — sample N answers, pick the majority.
 *
 * Paper: "Self-Consistency Improves Chain of Thought Reasoning in
 * Language Models" — Wang et al., 2022 (https://arxiv.org/abs/2203.11171).
 *
 * Pattern: Factory (GoF) → produces a `Runner` that composes `Parallel`
 *          of N `LLMCall` branches with a majority-vote merge function.
 * Role:    patterns/ layer. Pure composition of core primitives +
 *          core-flow compositions — no new abstractions.
 * Emits:   Everything `Parallel` + `LLMCall` emit (stream / composition /
 *          context). No pattern-specific event domain needed; consumers
 *          observe via the standard typed listeners.
 */
import { LLMCall } from '../core/LLMCall.js';
import { Parallel } from '../core-flow/Parallel.js';
/**
 * Build a SelfConsistency Runner. Given a system prompt, the Runner
 * runs `samples` parallel LLMCalls with the same input, extracts each
 * response's vote token, then returns the most-frequent token. Ties
 * are broken by the first response's extract.
 */
export function selfConsistency(opts) {
    if (opts.samples < 1) {
        throw new Error('SelfConsistency: samples must be >= 1');
    }
    if (opts.samples < 2) {
        throw new Error('SelfConsistency: samples must be >= 2 to have anything to vote on (use LLMCall for 1 sample)');
    }
    const extract = opts.extract ?? ((s) => s.trim());
    let builder = Parallel.create({
        name: opts.name ?? 'SelfConsistency',
        id: opts.id ?? 'self-consistency',
    });
    for (let i = 0; i < opts.samples; i++) {
        const branch = LLMCall.create({
            provider: opts.provider,
            model: opts.model,
            // Higher default temperature for sampling diversity — the whole
            // point of SelfConsistency is to get DIFFERENT samples.
            temperature: opts.temperature ?? 0.7,
            ...(opts.maxTokens !== undefined && { maxTokens: opts.maxTokens }),
        })
            .system(opts.systemPrompt)
            .build();
        builder = builder.branch(`sample-${i}`, branch);
    }
    return builder
        .mergeWithFn((results) => {
        const tallies = new Map();
        const order = [];
        for (const id of Object.keys(results).sort()) {
            // Object.keys() guarantees the index hits, so results[id] is defined.
            const value = results[id];
            if (value === undefined)
                continue;
            const vote = extract(value);
            if (!tallies.has(vote))
                order.push(vote);
            tallies.set(vote, (tallies.get(vote) ?? 0) + 1);
        }
        // SelfConsistency runs N branches and merges; if N === 0 there's
        // nothing to vote on. Throw rather than silently returning empty.
        if (order.length === 0) {
            throw new Error('SelfConsistency: no branch results to vote on');
        }
        let best = order[0];
        let bestCount = tallies.get(best) ?? 0;
        for (const vote of order) {
            const count = tallies.get(vote) ?? 0;
            if (count > bestCount) {
                best = vote;
                bestCount = count;
            }
        }
        return best;
    })
        .build();
}
//# sourceMappingURL=SelfConsistency.js.map