/**
 * Lexical full-text ranking over artifact bodies (BM25).
 *
 * The artifact index stores only metadata + a 500-char summary per node
 * (see `MetadataEntry`), so `aiwg index query` is metadata-scoped. This
 * module adds a lexical full-text vector: `aiwg index query "..." --fulltext`
 * reads the candidate nodes' source bodies and ranks them with Okapi BM25.
 *
 * This is deliberately distinct from the semantic/embedding vector
 * (`embedding-index.ts`, `--semantic`): full-text is lexical recall (exact
 * term matches in the body), semantic is conceptual. Both are composable.
 *
 * No persisted inverted index is built — ranking reads candidate bodies at
 * query time. This keeps the on-disk index free of regenerable full-text
 * bloat (the anti-bloat posture of #1488) and stays a pure query-side
 * feature with no index-build change.
 *
 * @implements #1494
 * @tests @test/unit/artifacts/fulltext.test.ts
 */
/** A document to rank: an opaque id plus its searchable body text. */
export interface FullTextDoc {
    id: string;
    text: string;
}
/** A ranked hit: the doc id, its normalized score (top = 1.0), and which query terms matched. */
export interface FullTextHit {
    id: string;
    /** Normalized BM25 score in (0, 1]; the top hit is 1.0. */
    score: number;
    /** Raw (un-normalized) BM25 score — useful for thresholding/debugging. */
    rawScore: number;
    /** Distinct query terms that occurred in this doc's body. */
    matchedTerms: string[];
}
export interface Bm25Options {
    /** Term-frequency saturation. Higher → repeated terms keep mattering. Default 1.5. */
    k1?: number;
    /** Length normalization. 0 = none, 1 = full. Default 0.75. */
    b?: number;
}
/**
 * Tokenize body/query text into lowercased terms. Splits on non-alphanumeric
 * (keeping intra-word hyphens), drops stop-words and single-char tokens.
 */
export declare function tokenizeText(text: string): string[];
/**
 * Rank documents against a query using Okapi BM25.
 *
 * Returns only docs that matched at least one query term, sorted by score
 * descending. Scores are normalized so the top hit is 1.0 (BM25 is otherwise
 * unbounded), matching the 0–1 score convention the rest of the query surface
 * uses; `rawScore` carries the un-normalized value.
 */
export declare function bm25Rank(docs: FullTextDoc[], query: string, options?: Bm25Options): FullTextHit[];
//# sourceMappingURL=fulltext.d.ts.map