import type { AIProviderName } from "../constants/enums.js";
import type { NeurolinkCredentials, StreamOptions, TextGenerationOptions } from "../types/index.js";
import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
/**
 * Ollama Provider — direct HTTP, no AI SDK.
 *
 * Wraps a local (or remote/Cloud) Ollama server via its OpenAI-compatible
 * `/v1` API. All request / stream / multi-step tool-loop orchestration lives
 * in `OpenAIChatCompletionsProvider`; this class declares configuration plus
 * the Ollama-specific behaviour:
 *
 *   1. `/v1` base-URL normalization (accepts a bare `OLLAMA_BASE_URL` host).
 *   2. No-auth-by-default with an optional `OLLAMA_API_KEY` for Ollama Cloud.
 *   3. Configurable per-model tool gating via `OLLAMA_TOOL_CAPABLE_MODELS` /
 *      `modelConfig` (`supportsTools`).
 *   4. Elevated request timeout for slow large local models (5-minute base
 *      default, overridable via `OLLAMA_TIMEOUT`).
 *   5. Native `/v1/embeddings` (`embed` / `embedMany`).
 *   6. Rich, actionable error mapping (`ollama serve` / `ollama pull` hints).
 *
 * Model discovery uses the base's `/v1/models` probe (Ollama supports it).
 *
 * @see https://docs.ollama.com/api/openai-compatibility
 */
export declare class OllamaProvider extends OpenAIChatCompletionsProvider {
    constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["ollama"]);
    protected getProviderName(): AIProviderName;
    protected getDefaultModel(): string;
    protected getFallbackModelName(): string;
    protected formatProviderError(error: unknown): Error;
    /**
     * Ollama proxies many local models with varying tool support. When
     * `OLLAMA_TOOL_CAPABLE_MODELS` (or `modelConfig`'s
     * `modelBehavior.toolCapableModels`) is configured, gate tools on a
     * substring match against the current model; with no list configured,
     * assume tools are supported (don't disable on absent evidence).
     */
    supportsTools(): boolean;
    /**
     * Local models are slow; the base already defaults Ollama to a 5-minute
     * timeout and honors a per-call `options.timeout`. Preserve the legacy
     * `OLLAMA_TIMEOUT` env override for callers who relied on it, applied only
     * when no explicit per-call timeout is set. Parsed with the shared
     * `parseTimeout` so both millisecond numbers ("240000") and duration strings
     * ("4m", "30s") work; a malformed value is ignored in favour of the default.
     */
    getTimeout(options: TextGenerationOptions | StreamOptions): number;
    /**
     * Health check: probe `/v1/models` and require at least one installed model.
     * A reachable server with zero models would let `resolveModelName()` fall
     * back to a model that the first real request can't serve, so report unusable.
     */
    validateConfiguration(): Promise<boolean>;
    getConfiguration(): {
        provider: AIProviderName;
        model: string;
        defaultModel: string;
        baseURL: string;
    };
    /**
     * Generate an embedding for a single text input via native /v1/embeddings.
     * Uses `OLLAMA_EMBEDDING_MODEL` (default `nomic-embed-text`); the embedding
     * model must be pulled locally (`ollama pull nomic-embed-text`).
     */
    embed(text: string, modelName?: string): Promise<number[]>;
    /**
     * Generate embeddings for multiple text inputs via native /v1/embeddings.
     */
    embedMany(texts: string[], modelName?: string): Promise<number[][]>;
    private callEmbeddings;
}