import type { AIProviderName } from "../constants/enums.js";
import type { NeurolinkCredentials } from "../types/index.js";
import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
/**
 * Cohere Provider — direct HTTP, no AI SDK.
 *
 * Routes Command R / Command R+ chat completions through Cohere's
 * OpenAI-compatible endpoint at /compatibility/v1. All request/stream/
 * tool-loop orchestration lives in `OpenAIChatCompletionsProvider`; this
 * class only declares configuration and provider-specific error mapping.
 *
 * Embed v3 is exposed via `embed()` / `embedMany()` backed by a native
 * POST to /v2/embed (the compatibility path is chat-only).
 *
 * @see https://docs.cohere.com/docs/compatibility-api
 * @see https://docs.cohere.com/reference/embed
 */
export declare class CohereProvider extends OpenAIChatCompletionsProvider {
    constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["cohere"]);
    protected getProviderName(): AIProviderName;
    protected getDefaultModel(): string;
    protected getFallbackModelName(): string;
    protected getFallbackModels(): string[];
    protected formatProviderError(error: unknown): Error;
    validateConfiguration(): Promise<boolean>;
    getConfiguration(): {
        provider: AIProviderName;
        model: string;
        defaultModel: string;
        baseURL: string;
    };
    /**
     * Default embedding model for Cohere.
     */
    protected getDefaultEmbeddingModel(): string;
    /**
     * Generate an embedding for a single text via Cohere's native /v2/embed
     * endpoint. Returns the float[] embedding vector.
     *
     * The shared OpenAI-compatible /compatibility/v1 path is chat-only; embed
     * lives on the native API (POST /v2/embed). Documented at
     * https://docs.cohere.com/reference/embed.
     */
    embed(text: string, modelName?: string): Promise<number[]>;
    /**
     * Batch embedding via Cohere's native /v2/embed endpoint. Cohere caps at
     * 96 inputs per request; larger batches are chunked.
     *
     * Partial failures are not surfaced: if any chunk request fails the whole
     * call rejects and already-embedded chunks are discarded — callers should
     * retry the full input.
     */
    embedMany(texts: string[], modelName?: string): Promise<number[][]>;
}