/**
 * Multilingual sentence encoder — the one optional, local model (no model API).
 *
 * Pure-TS SGNS hit a wall on the Korean long tail (프리다/후킹 are too rare to
 * learn from this corpus alone). A pretrained multilingual encoder brings
 * EXTERNAL knowledge — it already knows 후킹 ≈ hooking — so cross-lingual
 * retrieval works at the sentence level without any per-corpus training.
 *
 * It runs entirely on-device via transformers.js (onnxruntime under the hood):
 * zero model API calls, zero cloud account. The dependency is an OPTIONAL peer — it is lazy-imported
 * here, so the core framework still installs and runs without it; only the dense
 * encoder paths require it. Model weights download once to ~/.nexus/models/hf.
 */
/** Small, strong multilingual sentence-embedding model (384-dim, good Korean). */
export declare const ENCODER_MODEL = "Xenova/paraphrase-multilingual-MiniLM-L12-v2";
export declare const ENCODER_DIM = 384;
export declare function isEncoderInstalled(): boolean;
/** Lazily load the feature-extraction pipeline (downloads the model on first use). */
export declare function getEncoder(dataDir: string): Promise<any>;
/**
 * Embed texts into L2-normalized sentence vectors (mean-pooled). Returns a flat
 * row-major N×dim Float32Array. Batched to bound memory.
 */
export declare function embedTexts(dataDir: string, texts: string[], batchSize?: number, onProgress?: (done: number, total: number) => void): Promise<{
    vectors: Float32Array;
    dim: number;
}>;
/** Embed a single text → its sentence vector. */
export declare function embedOne(dataDir: string, text: string): Promise<Float32Array>;