/**
 * Local embedding generation for iranti's vector search layer.
 *
 * Produces dense float vectors from free-form text without requiring an
 * external model API. The algorithm is intentionally simple and fast:
 *  - Tokenise + lowercase the input
 *  - Expand tokens with a `TOKEN_SYNONYMS` table (e.g. "hq" → ["headquarters",
 *    "location", "city"]) so semantically equivalent queries surface matches
 *  - Hash each token with two FNV1a seeds into a fixed-dimension vector using
 *    the random-projection / feature-hashing trick (dimension defaults to 256,
 *    capped at 1024 via `IRANTI_EMBEDDING_DIM`)
 *  - Apply length-based token weighting (`1 + log(len)`) so rare long tokens
 *    outweigh stopwords
 *  - L2-normalise the final vector for cosine-distance comparisons
 *
 * This is intentionally NOT a neural embedding. It trades recall accuracy for
 * zero-latency, zero-cost, fully deterministic generation. The synonym table
 * is the primary quality lever — extend it as new domain vocabulary emerges.
 *
 * Key exports:
 *  - `generateEmbedding(text, dims?)`  — produce a vector
 *  - `cosineSimilarity(left, right)`   — compute similarity
 *  - `buildEmbeddingText(input)`       — canonical text from an EntryInput
 *  - `toPgVectorLiteral(vector)`       — format for pgvector `::vector` cast
 *  - `EMBEDDING_DIMENSIONS`            — active dimension count
 */
export declare const EMBEDDING_DIMENSIONS: number;
export declare function cosineSimilarity(left: number[], right: number[]): number;
type EmbeddingTextInput = {
    entityType?: string;
    entityId?: string;
    key: string;
    summary: string;
    valueRaw: unknown;
};
export declare function buildEmbeddingText(input: EmbeddingTextInput): string;
export declare function generateEmbedding(text: string, dimensions?: number): number[];
export declare function toPgVectorLiteral(vector: number[]): string;
export {};
//# sourceMappingURL=embeddings.d.ts.map