/**
 * Take clustering — finds segments that are likely re-takes of the same line.
 *
 * v0 uses Jaccard similarity over normalized token sets. This is:
 *   - Free (no API call)
 *   - Deterministic
 *   - Fast: O(N * windowSize) where windowSize caps the comparison range
 *
 * Why Jaccard over embeddings? Re-takes typically share most words verbatim
 * (the speaker says the same thing 2-3 times to nail the delivery). Token
 * overlap catches this reliably. Embeddings would catch paraphrased re-takes
 * too, but at the cost of an API call per segment — moved to v1 if needed.
 */
export interface Segment {
    start: number;
    end: number;
    text: string;
}
export interface Cluster {
    id: number;
    /** Indexes into the original segments array (preserves source order). */
    memberIndexes: number[];
    members: Segment[];
}
export interface ClusterOptions {
    /** Jaccard threshold above which two segments are considered the same take. Default 0.6. */
    threshold?: number;
    /** Only compare segments within this many positions of each other. Default 100. */
    window?: number;
    /** Skip segments shorter than this many tokens (too short to be meaningful re-takes). Default 4. */
    minTokens?: number;
}
/**
 * Normalize a segment to a token Set:
 *   - Lowercase
 *   - Strip punctuation
 *   - Drop stopwords + filler ("um", "uh", "like", "yeah", ...)
 *   - Drop tokens shorter than 2 chars
 */
export declare function tokenize(text: string): Set<string>;
export declare function jaccard(a: Set<string>, b: Set<string>): number;
/**
 * Group segments into clusters of likely re-takes.
 *
 * Returns clusters with size >= 2 only by default (singletons are not
 * interesting — they're just "things said once"). Use `includeSingletons`
 * for completeness.
 */
export declare function clusterSegments(segments: Segment[], opts?: ClusterOptions & {
    includeSingletons?: boolean;
}): Cluster[];
//# sourceMappingURL=clustering.d.ts.map