import type { EvaluateOptions } from '../evaluate/run.js';
import type { Dataset, DatasetExample, Evaluator, Experiment } from '../types/evaluate.js';
import type { CompletionRequest } from '../types/messages.js';
import type { PromptDefinition, PromptVersion, RenderOptions } from '../types/prompts.js';
import type { NexusResponse } from '../types/response.js';
/** Options for `evaluatePrompt()`: everything `evaluate()` takes, plus the model client and how inputs become variables. */
export interface EvaluatePromptOptions extends EvaluateOptions {
    /** Runs each rendered request. */
    client: {
        complete(request: CompletionRequest): Promise<NexusResponse>;
    };
    /** Turns an example's inputs into template variables. Defaults to the inputs themselves, or `{ input }` for a non-object. */
    variables?: (inputs: unknown, example: DatasetExample) => Record<string, unknown>;
    /** Render options, such as a model override for comparing models on one prompt. */
    render?: RenderOptions;
}
/**
 * Runs a prompt version over a dataset and scores it: the headless playground.
 *
 * The experiment records the prompt's name and content version in `metadata.prompt`, which is what
 * `experimentGate()` looks for, so an experiment run here is what unlocks the promotion of the exact
 * version it ran. A prompt defined in code gets the same version it would get when committed.
 */
export declare function evaluatePrompt<I = unknown, O = unknown>(prompt: PromptVersion | PromptDefinition | {
    definition: PromptDefinition;
}, dataset: Dataset<I, O>, evaluators: Array<Evaluator<I, O>>, options: EvaluatePromptOptions): Promise<Experiment>;