/**
 * Semantic search with cosine similarity and keyword fallback.
 * Uses embedder for vector search, falls back to term-frequency
 * keyword search when ONNX is unavailable.
 */
import type { Embedder } from "./embedder.js";
import type { IIndexManager } from "../storage/interfaces.js";
import type { BlackboardEntry, Decision } from "../utils/types.js";
export interface BlackboardSearchResult {
    entry: BlackboardEntry;
    relevance: number;
}
export interface DecisionSearchResult {
    decision: Decision;
    relevance: number;
}
export interface SearchResults<T> {
    results: T[];
    /**
     * Pre-slice match count — never capped by limit (field D9). Membership is
     * tested on RAW scores, before any status de-boost: semantic mode counts
     * raw cosine >= SEARCH_NOISE_FLOOR (every embedded item gets a score, so
     * an unfloored count would equal the corpus size; noise sits ~0.26-0.28);
     * keyword mode counts every literal term hit (score > 0 — TF scores are a
     * different scale and a literal hit is never noise). The results page
     * itself is NOT floored — it stays a ranked page.
     */
    total_matched: number;
    fallback_mode: boolean;
}
/**
 * Relevance below this is indistinguishable from noise (measured ~0.26-0.28
 * cosine for nonsense queries on MiniLM). Used for total_matched counting
 * here and for assemble's semantic-admission floor.
 */
export declare const SEARCH_NOISE_FLOOR = 0.3;
/**
 * Version stamp for count fields on search responses (2026-08-15 field
 * audit, ask 2): total_matched's meaning has changed across builds
 * (post-slice page length before 2.9.0; floored pre-page count since) and
 * field docs pinned to different generations coexist. Responses carry this
 * literal so a caller can tell which semantics it is reading. Bump ONLY
 * when count semantics actually change.
 */
export declare const COUNT_SEMANTICS = "pre_page_floored_v2";
export declare class SearchEngine {
    private readonly embedder;
    private readonly indexManager;
    constructor(embedder: Embedder, indexManager: IIndexManager);
    /** Search blackboard entries by semantic similarity or keyword fallback. */
    searchBlackboard(query: string, entries: BlackboardEntry[], options?: {
        entry_types?: string[];
        limit?: number;
    }): Promise<SearchResults<BlackboardSearchResult>>;
    /** Search decisions by semantic similarity or keyword fallback. */
    searchDecisions(query: string, decisions: Decision[], options?: {
        limit?: number;
    }): Promise<SearchResults<DecisionSearchResult>>;
}
/**
 * Cosine similarity for pre-normalized vectors (dot product).
 * Since all-MiniLM-L6-v2 outputs normalized vectors, cosine similarity
 * simplifies to the dot product.
 */
export declare function cosineSimilarity(a: number[], b: number[]): number;
/**
 * Term-frequency based keyword search for fallback mode.
 * Scores each item by how many query terms appear and how often.
 */
export declare function keywordSearch(query: string, items: {
    id: string;
    text: string;
}[], limit: number): {
    id: string;
    score: number;
}[];
//# sourceMappingURL=search.d.ts.map