import { Metrics } from './metrics.ts';
import { type Device, type RuntimeOptions } from './runtime.ts';
import { type ModelSource } from './source.ts';
export interface EmbedOptions extends RuntimeOptions {
    /** Quantization variant. Default q8 — a good size/quality point for embedders. */
    dtype?: string;
    /** 'auto' (default) uses WebGPU when available, else WASM/CPU. */
    device?: Device;
    onProgress?: (p: unknown) => void;
}
/** Embedding model wrapper (feature-extraction) with batching + similarity. */
export declare class NexusEmbedder {
    private extractor;
    readonly device: string;
    readonly modelId: string;
    readonly metrics: Metrics;
    private constructor();
    /**
     * Load an embedding model from an explicit source:
     *
     *   NexusEmbedder.load({ hub: 'Xenova/bge-small-en-v1.5' })
     *   NexusEmbedder.load({ base: '/models/', id: 'BAAI/bge-small-en-v1.5' })
     *   NexusEmbedder.load({ archive: fileFromInput })
     */
    static load(source: ModelSource, opts?: EmbedOptions): Promise<NexusEmbedder>;
    /** Embed one text into a normalized vector. */
    embed(text: string): Promise<Float32Array>;
    /** Embed many texts; returns one normalized vector per text. */
    embedBatch(texts: string[]): Promise<Float32Array[]>;
    dispose(): Promise<void>;
}
/** Cosine similarity of two normalized vectors (= dot product). */
export declare function similarity(a: Float32Array, b: Float32Array): number;