import type { AuraCanonicalAsset } from "./CanonicalAsset.js";
/**
 * Semantic-search upgrade seam over the keyword baseline in `ranking.ts`.
 *
 * The federation layer scores assets through a single seam (`scoreAsset`); this
 * module provides the vector-embedding alternative the comment there promises.
 * Everything here is dependency-free and fully deterministic: the default
 * provider hashes tokens into a fixed-size vector, so the same text always
 * yields the same embedding with no network, keys, randomness, or wall-clock
 * time involved.
 */
/** Pluggable text embedder. Implementations must be deterministic per input. */
export interface EmbeddingProvider {
    /** Embed each input string into a numeric vector of uniform dimension. */
    embed(texts: readonly string[]): Promise<number[][]>;
}
/** Default embedding dimensionality for {@link LocalHashEmbedding}. */
export declare const DEFAULT_EMBEDDING_DIMS = 256;
/**
 * Deterministic, dependency-free bag-of-words embedding.
 *
 * Each token is hashed into one of `dims` buckets and summed (sign derived from
 * a second hash bit so unrelated tokens do not all push in the same direction),
 * then the vector is L2-normalized. Identical text -> identical vector; cosine
 * similarity of identical text is exactly 1.
 */
export declare class LocalHashEmbedding implements EmbeddingProvider {
    readonly dims: number;
    constructor(dims?: number);
    private embedOne;
    embed(texts: readonly string[]): Promise<number[][]>;
}
/**
 * Cosine similarity of two equal-length vectors. Returns 0 when either vector
 * is all zeros or lengths differ. For L2-normalized vectors this equals the dot
 * product, but we recompute norms so it is correct for any input.
 */
export declare function cosineSimilarity(a: readonly number[], b: readonly number[]): number;
/**
 * Canonical text used to embed an asset: title, tags, then description. This is
 * the same field set the ranking baseline scores against, concatenated so a
 * bag-of-words embedding can capture all of them.
 */
export declare function assetEmbeddingText(asset: AuraCanonicalAsset): string;
/**
 * Rank assets against a query by cosine similarity of their embeddings.
 *
 * Embeds the query and every asset's {@link assetEmbeddingText} in one batched
 * call to the provider, then sorts descending by score. Ties break by asset id
 * so the ordering is stable and deterministic.
 */
export declare function embeddingRanker(provider: EmbeddingProvider, query: string, assets: readonly AuraCanonicalAsset[]): Promise<{
    asset: AuraCanonicalAsset;
    score: number;
}[]>;
//# sourceMappingURL=embedding.d.ts.map