/**
 * Optional semantic re-ranking for The Brain.
 *
 * TheBrainV2 retrieves lexically: BM25 over an inverted index, Jaccard re-rank, plus a
 * hand-written table of bilingual synonym clusters. That table is the tell — it exists
 * because a memory stored as "race condition al flushear el índice invertido" is invisible
 * to a search for "why do memories disappear", and every new vocabulary gap needs another
 * hand-written cluster. Embeddings close that class of gap without a maintained word list.
 *
 * What it deliberately does NOT do:
 *  - become a dependency. Off unless LEMMA_BRAIN_EMBEDDINGS is set, and every failure path
 *    (no Ollama, no model, timeout, malformed response) silently degrades to pure BM25.
 *    The module's zero-dependency, sub-10ms-per-query design stays the default.
 *  - replace retrieval. BM25 still selects the candidate set; embeddings only reorder it.
 *    A memory the inverted index never surfaces can't be rescued here — which is the
 *    conservative direction, since a wrong reordering costs a place in a list while a wrong
 *    retrieval costs a wrong answer.
 *  - block. One probe decides availability for the whole session, each request has a hard
 *    timeout, and the number of vectors computed per call is capped.
 *
 * Vectors live in their own sidecar file, never in entries.ndjson: 768 floats per entry
 * would multiply the corpus every session re-reads at startup by an order of magnitude, for
 * data only this optional path ever reads.
 */
/** Enabled only on an explicit opt-in — see the module comment. */
export declare function embeddingsEnabled(): boolean;
export interface EmbeddedCandidate {
    id: string;
    /** Lexical similarity from TheBrainV2.search(), already normalized 0-1. */
    similarity: number;
    /** Text to embed when this candidate has no vector yet. */
    text: string;
}
export declare function resetEmbeddingAvailability(): void;
export declare function cosine(a: number[], b: number[]): number;
/** Sidecar vector store, keyed by entry id and tagged with the model that produced it. */
export declare class EmbeddingSidecar {
    private file;
    private vectors;
    private model;
    private loaded;
    constructor(file: string);
    private load;
    get(id: string): number[] | undefined;
    set(id: string, vec: number[]): void;
    /** Drops vectors for ids no longer in the corpus, then rewrites the sidecar atomically. */
    flush(liveIds: Set<string>): void;
}
/**
 * Blend semantic similarity into an already-ranked candidate list.
 *
 * Returns the ids in their new order along with the blended scores, or null when embeddings
 * are unavailable — the caller then keeps the lexical ordering untouched. Candidates with no
 * vector (budget exhausted this call) keep their lexical score rather than being scored
 * against a zero vector, which would drop them below every scored candidate for no reason
 * other than arriving late in the queue.
 */
export declare function semanticRerank(query: string, candidates: EmbeddedCandidate[], sidecar: EmbeddingSidecar): Promise<Map<string, number> | null>;
//# sourceMappingURL=BrainEmbeddings.d.ts.map