/**
 * Zero-Dependency Semantic Similarity Engine
 *
 * Captures meaning without embeddings or external APIs.
 *
 * Three mechanisms:
 *
 * 1. SYNONYM GRAPH
 *    Hard-coded domain knowledge mapping related concepts.
 *    "deploy" ↔ "release" ↔ "publish" ↔ "배포" ↔ "ship"
 *    Bilingual: English ↔ Korean for dev terminology.
 *
 * 2. CO-OCCURRENCE PMI (Pointwise Mutual Information)
 *    Learns from the corpus itself: words that appear together
 *    more than expected by chance are semantically related.
 *    "authentication" frequently near "password" → related.
 *    No training, no model — pure statistics.
 *
 * 3. QUERY EXPANSION
 *    Expands search query with synonyms + co-occurrence neighbors.
 *    "deploy" → ["deploy", "release", "publish", "배포", "ship"]
 *    Then runs BM25 on expanded query → semantic matches found.
 */
/** Each group = mutually synonymous concepts. */
export declare const SYNONYM_GROUPS: string[][];
/** Get synonyms for a word. */
export declare function getSynonyms(word: string): string[];
export type CoOccurrenceModel = {
    /** Get top-N semantically related words for a given word. */
    getRelated: (word: string, topN?: number) => {
        word: string;
        pmi: number;
    }[];
    /** Rebuild from a new corpus. */
    rebuild: (documents: string[]) => void;
    /** Number of unique terms. */
    vocabSize: () => number;
};
export declare function stem(word: string): string;
/** Transliterate Korean loanword to English equivalent. */
export declare function transliterate(word: string): string | undefined;
/** Trigram Jaccard similarity between two words. Returns 0-1. */
export declare function trigramSimilarity(a: string, b: string): number;
export declare function tokenize(text: string): string[];
export declare function createCoOccurrenceModel(): CoOccurrenceModel;
export type ExpandedQuery = {
    /** Original query tokens. */
    original: string[];
    /** Expanded tokens (includes original + synonyms + co-occurrence). */
    expanded: string[];
    /** Which tokens were added and why. */
    expansions: {
        token: string;
        source: "synonym" | "cooccurrence" | "stem" | "transliteration" | "embedding";
        relatedTo: string;
    }[];
};
/** Anything that can return related words — both CoOccurrenceModel and the
 *  learned EmbeddingModel satisfy this, so either can feed query expansion. */
export type RelatedSource = {
    getRelated: (word: string, topN?: number) => {
        word: string;
        pmi: number;
    }[];
};
/**
 * Expand a query with semantic neighbors.
 *
 * "deploy error" → ["deploy", "release", "publish", "배포",
 *                    "error", "bug", "issue", "에러",
 *                    + co-occurrence neighbors]
 */
export declare function expandQuery(query: string, coModel?: CoOccurrenceModel, maxCoOccurrencePerToken?: number, embModel?: RelatedSource, maxEmbeddingPerToken?: number, minEmbeddingScore?: number): ExpandedQuery;
/**
 * Compute semantic similarity between two texts.
 *
 * Five signals combined:
 * 1. Direct token overlap (Jaccard)
 * 2. Synonym + transliteration expanded overlap
 * 3. Stemmed overlap
 * 4. Trigram fuzzy match (catches morphological variants)
 * 5. Co-occurrence similarity
 *
 * Returns 0-1.
 */
export declare function semanticSimilarity(textA: string, textB: string, coModel?: CoOccurrenceModel): number;