/**
 * Pluggable embedder. Same shape as the LLM client: OpenAI-compatible
 * endpoint by default, with an optional Anthropic bypass if anyone ever
 * asks (Anthropic doesn't publish an embeddings API today, so for now the
 * interface is strictly OpenAI-compatible).
 *
 * Defaults assume the user's local Ollama has nomic-embed-text, which
 * produces 768-dim vectors matching the vec0 table declared in migration 1.
 */
export interface Embedder {
    /** Human-readable model ID for logging / trace. */
    readonly modelName: string;
    /** Embedding dimension the model produces. */
    readonly dimension: number;
    /** Embed one string; returns a Float32Array of length `dimension`. */
    embed(text: string): Promise<Float32Array>;
    /** Embed a batch; should be at least as efficient as N serial calls. */
    embedBatch(texts: string[]): Promise<Float32Array[]>;
}
export interface EmbedderConfig {
    apiUrl?: string;
    apiKey?: string;
    model?: string;
    dimension?: number;
    timeout?: number;
}
export declare class OpenAICompatibleEmbedder implements Embedder {
    readonly modelName: string;
    readonly dimension: number;
    private apiUrl;
    private apiKey;
    private timeout;
    constructor(config?: EmbedderConfig);
    embed(text: string): Promise<Float32Array>;
    embedBatch(texts: string[]): Promise<Float32Array[]>;
}
export declare class EmbedderError extends Error {
    statusCode: number;
    details?: string | undefined;
    constructor(message: string, statusCode: number, details?: string | undefined);
}
/** Serialize a Float32Array for sqlite-vec's vec_f32() SQL function. */
export declare function vecToJson(v: Float32Array): string;
export declare function getEmbedder(): Embedder;
/** Test-only: reset the cached embedder (forces env re-read on next getEmbedder()). */
export declare function resetEmbedder(): void;