import type { EmbeddingConfig, EmbeddingRequest, EmbeddingResponse, EmbeddingsProvider } from '../types/embeddings.js';
import type { EmbeddingManagerRuntime } from './manager.js';
/**
 * What `EmbeddingManager` loads on the first embedding: routing, caching, batching, the budget,
 * retries, and the provider calls. Kept apart so a client that never embeds never imports it.
 */
export declare class EmbeddingEngine {
    private readonly config;
    private readonly runtime;
    /** The manager's registry, shared so adapters registered here are visible to it and back. */
    private readonly providers;
    private readonly tokenizer;
    private readonly lifecycle;
    private readonly localCache?;
    private autoRegistered;
    constructor(config: EmbeddingConfig, runtime: EmbeddingManagerRuntime, 
    /** The manager's registry, shared so adapters registered here are visible to it and back. */
    providers: Map<string, EmbeddingsProvider>);
    /**
     * Embeds one text or a batch.
     *
     * The single-string form returns one vector at `vectors[0]`; a batch returns them in input order
     * regardless of how many provider calls the split required.
     */
    embed(request: EmbeddingRequest): Promise<EmbeddingResponse>;
    /** Answers an admitted request: the cache first, then the provider for whatever is left. */
    private runEmbed;
    /** Embeds one text and returns the vector alone, for callers that never need the metadata. */
    embedOne(text: string, options?: Omit<EmbeddingRequest, 'input'>): Promise<number[]>;
    private resolveRoute;
    /**
     * Uses the adapter's own default model when the request named none.
     *
     * Without this a caller with a single non-OpenAI adapter would be routed to the `auto` alias
     * target, which that adapter does not serve.
     */
    private withProviderDefaultModel;
    /** Builds the adapters the client's provider credentials imply, unless one of that name exists. */
    private ensureAutoRegistered;
    private runBatches;
    private callWithRetry;
    private callProvider;
    private postProcess;
    /** Applies the per-request cost limit, when one is set, and returns the estimate it priced. */
    private enforceCostBudget;
    /** What the inputs would cost on a model, in US dollars, for the shared budget. */
    private estimateCost;
    private estimate;
    private cacheContext;
    /**
     * Every input's cached vector, or `undefined`. A shared store answers the whole request in one
     * call, keyed by hash; an adapter is asked for each input at once, under the keys it always had,
     * so a fleet mid-rollout still shares entries.
     */
    private readCache;
    private writeCache;
    /** Keys for a shared store. Loaded with the store, so an engine without one never hashes. */
    private storeKeys;
}