import type { LLM as LLMBase } from '../agents/providers.js';
import { LLM as LLMProviderBase } from '../agents/providers.js';
/** Options for {@link ZRTLLM} (the ZRT-hosted managed model). */
export type ZRTLLMOptions = {
    /** Model identifier. Defaults to `"google.gemma-4-31b"`. */
    model?: string;
    /** Sampling temperature controlling output randomness. Defaults to `0.7`. */
    temperature?: number;
    /** Maximum number of tokens to generate. Defaults to `1024`. */
    maxOutputTokens?: number;
    /** Base URL of the inference service; the default endpoint is resolved automatically. */
    baseUrl?: string | null;
    /** Extra options are ignored for forward compatibility. */
    [key: string]: any;
};
declare class ZRTLLMImpl extends LLMProviderBase {
    _providerName: string;
    model: string;
    temperature: number;
    maxOutputTokens: number;
    _isInference: boolean;
    _inferenceBaseUrl: string | null;
    constructor(opts?: ZRTLLMOptions);
    getRuntimeConfig(): Record<string, any>;
}
/** The ZRT-hosted LLM on Zero Runtime's managed inference. */
export type ZRTLLM = ZRTLLMImpl;
/** Create a {@link ZRTLLM} client (managed model, default `"google.gemma-4-31b"`). */
export declare const ZRTLLM: (opts?: ZRTLLMOptions) => ZRTLLM;
/** Options for {@link LLM.google} (Google / Gemini). */
export type LLMGoogleOptions = {
    /** Provider identifier; normally left unset. */
    provider?: string | null;
    /** API key for the provider. Falls back to the provider's own resolution if unset. */
    apiKey?: string | null;
    /** Model name to use (e.g. a Gemini model id). Defaults to `"gemini-2.5-flash"`. */
    modelId?: string | null;
    /** Provider region/location for the request. */
    location?: string | null;
    /** Sampling temperature; higher values produce more varied output. Defaults to `0.7`. */
    temperature?: number | null;
    /** Maximum number of tokens to generate in the response. */
    maxOutputTokens?: number | null;
    /** Token budget allotted to the model's internal "thinking"/reasoning. */
    thinkingBudget?: number | null;
    /** When true, include the model's thought summaries in the response. */
    includeThoughts?: boolean | null;
    /** Provider safety-filter settings; each entry is a category/threshold pair. */
    safetySettings?: Array<Record<string, any>> | null;
    /** Override the inference endpoint base URL. */
    baseUrl?: string | null;
};
/** Options for {@link LLM.sarvam} (Sarvam AI). */
export type LLMSarvamOptions = {
    /** Provider identifier; normally left unset. */
    provider?: string | null;
    /** API key for the provider. */
    apiKey?: string | null;
    /** Model name to use. */
    modelId?: string | null;
    /** Sampling temperature; higher values produce more varied output. */
    temperature?: number | null;
    /** Maximum number of tokens to generate in the completion. */
    maxCompletionTokens?: number | null;
    /** Tool-selection strategy passed to the provider (e.g. `"auto"`). */
    toolChoice?: string | null;
    /** Reasoning-effort hint passed to the provider. */
    reasoningEffort?: string | null;
    /** Enable wiki grounding for responses. */
    wikiGrounding?: boolean | null;
    /** Override the inference endpoint base URL. */
    baseUrl?: string | null;
};
/** Factory for LLM components used by the direct inference API. */
export declare class LLM {
    /**
     * Build a Google (Gemini) LLM component. Defaults: model `"gemini-2.5-flash"`,
     * temperature `0.7` (the gateway-served defaults). Other options are forwarded
     * only when set; unset options use the provider's own defaults.
     */
    static google(opts?: LLMGoogleOptions): LLMBase;
    /**
     * Build a Sarvam AI LLM component. Only the options you set are forwarded to the
     * provider; unset options use the provider's own defaults.
     */
    static sarvam(opts?: LLMSarvamOptions): LLMBase;
}
export {};