import type { ILLMProvider } from '../llm/ILLMProvider.ts';
/**
 * Context a judge needs to score a single evaluation task. `expected` is the
 * suite's ground truth; `output` is what the member actually produced.
 */
export interface JudgeContext {
    input: string;
    output: string;
    expected: string;
}
export interface EvalJudge {
    /** Returns a score in [0, 1], or null when the metric cannot be scored. */
    score(metric: string, context: JudgeContext): Promise<number | null>;
}
export declare const AB_METRICS: string[];
export declare const JUDGE_SYSTEM_PROMPT = "You are an evaluation judge. Score the answer on a single metric. Respond with ONLY one decimal number between 0 and 1, where 0 is worst and 1 is perfect. Do not explain.";
/**
 * Default judge. LLM-judged metrics (faithfulness, relevance, context_recall)
 * ask the provider for a bare score; answer_correctness compares embeddings
 * with cosine similarity so it costs no extra LLM calls.
 */
export declare class LLMJudge implements EvalJudge {
    private readonly llm;
    constructor(llm: ILLMProvider);
    score(metric: string, context: JudgeContext): Promise<number | null>;
    private llmScore;
    private embeddingScore;
}
/** Extracts the first number from a judge reply and clamps it to [0, 1]. */
export declare function parseJudgeScore(text: string): number | null;
//# sourceMappingURL=EvalJudge.d.ts.map