import { type ModelRuntime } from "../provider/index.js";
import type { EvalAnswerGroundednessMetricOptions } from "./types.js";
type GroundednessJudge = NonNullable<EvalAnswerGroundednessMetricOptions["judge"]>;
type RubricJudge = (input: {
    rubric: string;
    input: unknown;
    output: Record<string, unknown>;
    reference?: unknown;
    metadata: Record<string, unknown>;
}) => Promise<{
    score: number;
    pass?: boolean;
    explanation?: string;
}>;
/** Options for the built-in general-purpose LLM rubric judge. */
export interface EvalLlmRubricJudgeOptions {
    /** Model id or runtime used to judge answer quality. Defaults to the runtime auto model. */
    model?: string | ModelRuntime;
    /** Minimum score required for the judge to pass. Defaults to 0.8. */
    threshold?: number;
    /** Maximum judge response tokens. Defaults to 800. */
    maxOutputTokens?: number;
    /** Judge model temperature. Defaults to 0 for repeatability. */
    temperature?: number;
    /** Provider-specific options forwarded to the model runtime. */
    providerOptions?: Record<string, unknown>;
    /**
     * What the judge is being shown.
     *
     * `"answer"` (default) grades an agent's answer to a task, and sends the
     * task input alongside it. `"text"` grades a standing piece of text against
     * the rubric with no task premise -- use it when the graded value was not
     * produced in response to the input, for example a stored document or a
     * labelled corpus, where the answer framing would read as an agent that
     * echoed its prompt instead of doing the work.
     */
    framing?: "answer" | "text";
}
/** Options for the built-in LLM groundedness judge. */
export interface EvalLlmGroundednessJudgeOptions {
    /** Model id or runtime used to judge answer grounding. Defaults to the runtime auto model. */
    model?: string | ModelRuntime;
    /** Minimum score required for the judge to pass. Defaults to 0.8. */
    threshold?: number;
    /** Maximum retrieved evidence characters included in the judge prompt. Defaults to 12000. */
    maxEvidenceChars?: number;
    /** Maximum judge response tokens. Defaults to 800. */
    maxOutputTokens?: number;
    /** Judge model temperature. Defaults to 0 for repeatability. */
    temperature?: number;
    /** Provider-specific options forwarded to the model runtime. */
    providerOptions?: Record<string, unknown>;
}
declare function createLlmRubricJudge(options?: EvalLlmRubricJudgeOptions): RubricJudge;
declare function createLlmGroundednessJudge(options?: EvalLlmGroundednessJudgeOptions): GroundednessJudge;
/** Built-in judge factories for semantic eval metrics. */
export declare const judges: {
    readonly llm: {
        /** Create an LLM judge for `metrics.judge.rubric`. */
        readonly rubric: typeof createLlmRubricJudge;
        /** Create an LLM judge for `metrics.answer.groundedness`. */
        readonly groundedness: typeof createLlmGroundednessJudge;
    };
};
export {};
//# sourceMappingURL=judges.d.ts.map