/** Measures true, unclamped token lengths for each input text. */
export type TokenCounter = (texts: string[]) => Promise<number[]>;
/** The unit shape containment reads. `text` must be the exact slice `[sourceStart, sourceEnd)`. */
export interface ContainmentUnit {
    text: string;
    sourceStart: number;
    sourceEnd: number;
}
export interface ContainmentBudget {
    /** Resolved token cap: a piece measuring at or below it fits the model's window. */
    cap: number;
    countTokens: TokenCounter;
}
export interface RunContainmentBudget<T> extends ContainmentBudget {
    /**
     * Joins a run exactly as the chunk text is built, so the measured text is the
     * text that will be embedded.
     */
    joinUnits: (units: readonly T[]) => string;
}
/**
 * Divide one sentence unit into pieces that each measure at or below the cap.
 *
 * Halving, not a search for the longest fitting prefix: token count is not
 * monotonic in prefix length — the prefixes of `playingx` measure 3, 3, 4, 3,
 * 4, 4, 3, 4 here. A single grapheme over the cap is emitted anyway, leaving
 * the embedder's clamp to truncate it.
 */
export declare function splitUnitToFit<T extends ContainmentUnit>(unit: T, budget: ContainmentBudget): Promise<T[]>;
/**
 * Divide a group into consecutive runs of whole units whose joined text measures
 * at or below the cap.
 *
 * A one-unit run is final even above the cap: {@link splitUnitToFit} already
 * reduced that unit as far as a valid cut allows. Each division strictly reduces
 * the units per run, so progress does not depend on the text.
 */
export declare function splitUnitsIntoFittingRuns<T>(units: readonly T[], budget: RunContainmentBudget<T>): Promise<T[][]>;
//# sourceMappingURL=token-containment.d.ts.map