/**
 * Embedding pipeline using @huggingface/transformers.
 *
 * Generates normalized embeddings for semantic search. The embedding
 * dimension is derived at runtime from the model's actual output.
 * First use downloads the model (~23MB) to the configured cache directory.
 */
import type { IndexProgressCallback } from "./types.js";
export interface EmbedderOptions {
    /** Absolute path to the model cache directory (e.g. ~/.dreb/agent/models/). */
    modelCacheDir: string;
    /** HuggingFace model name. Default: 'Xenova/all-MiniLM-L6-v2'. */
    modelName?: string;
    /** Number of texts to embed per batch. Default: 32. */
    batchSize?: number;
}
export declare class Embedder {
    private readonly modelCacheDir;
    private readonly modelName;
    private readonly batchSize;
    private extractor;
    private initPromise;
    private resolvedDimension;
    constructor(options: EmbedderOptions);
    /**
     * Initialize the model pipeline. Must be called before embedding.
     *
     * On first use this downloads the ONNX model to `modelCacheDir`.
     * Subsequent calls reuse the cached model.
     */
    initialize(): Promise<void>;
    /**
     * Embed documents for indexing.
     *
     * Applies model-specific prefixes if required, then processes texts
     * in batches of `batchSize` for memory efficiency.
     */
    embedDocuments(texts: string[], onProgress?: IndexProgressCallback): Promise<Float32Array[]>;
    /**
     * Embed a query for search.
     *
     * Applies model-specific query prefix if required.
     */
    embedQuery(query: string): Promise<Float32Array>;
    /** Get the embedding dimension. Returns the model's actual dimension once known, or 384 as default. */
    get dimension(): number;
    /** Dispose the pipeline to free memory. */
    dispose(): void;
    /** Throw if initialize() hasn't been called yet. */
    private ensureInitialized;
}
//# sourceMappingURL=embedder.d.ts.map