export declare const DEFAULT_OLLAMA_HOST = "http://127.0.0.1:11434";
export interface InstalledModel {
    name: string;
    sizeBytes: number;
    modifiedAt: string;
}
export interface ChatOptions {
    model: string;
    system?: string;
    user: string;
    temperature?: number;
    /**
     * Whether to let the model emit a <think>…</think> chain before the answer.
     * Defaults to false — the bridge's whole point is fast, cheap delegation;
     * thinking models burn minutes on trivial tasks. Opt in per-call if needed.
     */
    think?: boolean | 'low' | 'medium' | 'high';
    /**
     * Ollama keep_alive parameter (seconds, duration string, or -1 for
     * forever). Controls how long the model stays loaded after the call.
     */
    keepAlive?: string | number;
    /**
     * JSON schema passed to Ollama `format:` field for grammar-constrained
     * structured output. Used by classify (enum schema) and extract (full
     * object schema). Callers must pre-sanitize with sanitizeSchemaForOllama()
     * to avoid GBNF crashes on pattern/format:email/format:uri constraints.
     */
    format?: object;
    /**
     * Maximum number of tokens the model may generate. Passed as
     * `options.num_predict` to Ollama. Default: unlimited.
     */
    numPredict?: number;
    /**
     * Context window size in tokens. Passed as `options.num_ctx` to Ollama.
     * Ollama's runtime default is 4096 regardless of the model's maximum;
     * inputs exceeding the window are silently left-truncated. Set per tier
     * to avoid data loss on longer documents.
     */
    numCtx?: number;
    signal?: AbortSignal;
}
/**
 * Result returned by OllamaClient.chat(). Includes the text content plus
 * token usage counters from Ollama (0 when the daemon doesn't report them).
 */
export interface ChatResult {
    text: string;
    /** Tokens used to encode the prompt (Ollama: prompt_eval_count). */
    promptTokens: number;
    /** Tokens generated in the completion (Ollama: eval_count). */
    completionTokens: number;
}
export declare class OllamaDaemonError extends Error {
    readonly cause?: unknown | undefined;
    constructor(message: string, cause?: unknown | undefined);
}
export declare class OllamaClient {
    readonly host: string;
    private readonly ollama;
    constructor(host?: string);
    ping(): Promise<{
        version: string;
    }>;
    listInstalled(): Promise<InstalledModel[]>;
    chat(opts: ChatOptions): Promise<ChatResult>;
}
//# sourceMappingURL=client.d.ts.map