/** Tokenise text into lowercase content tokens (drops stopwords + length<3). */
export declare function tokenize(text: string): string[];
export interface CorpusStats {
    /** docFreq: how many documents contain each token */
    df: Map<string, number>;
    /** idf: log((N - df + 0.5) / (df + 0.5) + 1) — BM25's smoothed IDF */
    idf: Map<string, number>;
    /** averageDocumentLength in tokens */
    avgDocLen: number;
    /** total docs in corpus */
    N: number;
}
/** Build BM25 corpus statistics from a set of tokenised documents. */
export declare function buildCorpusStats(tokenisedDocs: string[][]): CorpusStats;
/**
 * Type penalty for meta-commits — multiplies the hybrid score by `factor`
 * when the document's name matches the meta-commit regex (release bumps,
 * merge commits, badge updates, etc.). These commits bundle every issue
 * number from a release window and steal top-1 from real work commits.
 *
 * Defaults: factor 0.5, regex covers `chore(release)`, `Merge `, `bump `,
 * `[Dream Cycle]`, `publish 3.x.y`. Tune via env or call site.
 */
export declare const META_COMMIT_REGEX: RegExp;
export declare function typePenalty(name: string | undefined, factor?: number, regex?: RegExp): number;
/**
 * BM25 score of one document against a query.
 * Standard Okapi formula with k1=1.5 and b=0.75.
 */
export declare function bm25Score(queryTokens: string[], docTokens: string[], stats: CorpusStats, k1?: number, b?: number): number;
/**
 * Min-max normalise a score vector to [0, 1].
 * Returns the original vector if it is constant (avoid divide-by-zero).
 */
export declare function normalise(scores: number[]): number[];
/**
 * Combine cosine and BM25 scores. Both vectors must be aligned by docIndex.
 * Returns hybridScores[] aligned to the same index.
 * alpha controls the cosine weight; (1-alpha) is the BM25 weight.
 */
export declare function hybridScores(cosine: number[], bm25: number[], alpha?: number): number[];
/**
 * Multi-field BM25 score — treats subject and body as separate fields with
 * independent token frequencies, then combines `subjectWeight * subjectBM25
 * + bodyWeight * bodyBM25`. Subject (commit title / pattern name) carries
 * the high-signal tokens (file names, action verbs, ADR refs); body is
 * often boilerplate. Default 3:1 weight reflects that asymmetry.
 *
 * Caller must build separate CorpusStats for subjects and bodies (their
 * IDF distributions differ — subjects are short, bodies long).
 */
export declare function multiFieldBM25(queryTokens: string[], subjectTokens: string[], bodyTokens: string[], subjectStats: CorpusStats, bodyStats: CorpusStats, subjectWeight?: number, bodyWeight?: number): number;
/**
 * Cosine similarity between two equal-length numeric vectors.
 * Returns 0 if either has zero norm.
 */
export declare function cosineSim(a: number[], b: number[]): number;
/**
 * Maximal Marginal Relevance rerank.
 *
 * Greedy selection: at each step pick the candidate that maximises
 *   lambda * relevance(c) - (1-lambda) * max(similarity(c, picked))
 *
 * lambda=1.0 → pure relevance (no diversity adjustment)
 * lambda=0.0 → pure diversity
 *
 * Default lambda=0.5 balances both.
 */
export declare function mmrRerank<T extends {
    embedding: number[];
}>(candidates: Array<T & {
    relevance: number;
}>, k: number, lambda?: number): Array<T & {
    relevance: number;
    mmrScore: number;
}>;
//# sourceMappingURL=hybrid-retrieval.d.ts.map