/**
 * Vector knowledge — embedding-backed semantic retrieval.
 *
 * Phase-1 sibling of the reference engines' vector store. `VectorKnowledge` embeds
 * documents and queries and retrieves by cosine similarity, satisfying the same
 * {@link Knowledge} interface as the lexical retriever (so the agent accepts
 * either). The `Embedder` is pluggable: the default `HashEmbedder` is deterministic
 * and offline (feature-hashed bag-of-words) — good for tests and a zero-dependency
 * default — while a gateway embedder drops in behind the same interface for true
 * semantics.
 */
import type { Knowledge, KnowledgeHit } from './knowledge.js';
/** A small deterministic non-cryptographic hash (FNV-1a, 32-bit). */
export declare function hashToken(token: string): number;
export interface Embedder {
    embed(text: string): number[];
}
/**
 * Deterministic, offline feature-hashing embedder. Hashes each token into one of
 * `dim` buckets (signed) and L2-normalizes. No learned semantics, but a real
 * vector with cosine geometry — docs sharing tokens land near each other.
 */
export declare class HashEmbedder implements Embedder {
    private readonly dim;
    constructor(dim?: number);
    embed(text: string): number[];
}
/** An embedding-backed knowledge store with cosine-similarity retrieval. */
export declare class VectorKnowledge implements Knowledge {
    private readonly embedder;
    private readonly docs;
    constructor(embedder?: Embedder);
    ingest(content: string, source: string): void;
    query(query: string, topK?: number): KnowledgeHit[];
}
//# sourceMappingURL=vector.d.ts.map