/**
 * BM25 over the turn's own messages.
 *
 * The corpus is small — a few hundred documents at most — so the index
 * is rebuilt from term frequencies each pass rather than maintained
 * incrementally; the arithmetic is the cost of a few string splits. The
 * constants are the textbook ones (k1 = 1.2, b = 0.75). What matters
 * here is not ranking quality against a web corpus but that a message
 * mentioning the goal's identifiers outscores one that does not, and
 * that a term every message contains (the turn's own file name, say)
 * stops counting.
 */
export interface Bm25Document {
    readonly tf: ReadonlyMap<string, number>;
    readonly length: number;
}
export interface Bm25Index {
    readonly documents: readonly Bm25Document[];
    readonly documentFrequency: ReadonlyMap<string, number>;
    readonly averageLength: number;
}
export declare function indexDocument(text: string): Bm25Document;
export declare function buildIndex(documents: readonly Bm25Document[]): Bm25Index;
/** BM25 score of one document for a query given as tokens. */
export declare function bm25Score(index: Bm25Index, doc: Bm25Document, queryTokens: readonly string[]): number;
//# sourceMappingURL=bm25.d.ts.map