import type { LanguageModelV4, LanguageModelV4CallOptions, LanguageModelV4FinishReason, LanguageModelV4FunctionTool, LanguageModelV4GenerateResult, LanguageModelV4Message, LanguageModelV4StreamResult, LanguageModelV4ToolChoice, LanguageModelV4Usage } from "@ai-sdk/provider";
import { type ChatMessage } from "./native-binding.js";
import { type LlamaCppCacheConfig, type LlamaCppMemorySafetyConfig, type LlamaCppModelMemoryInfo, type LlamaCppReasoningEffort, type LlamaCppReasoningConfig } from "./llama-cpp-provider-config.js";
export interface LlamaCppModelConfig {
    modelPath: string;
    /**
     * Path to the multimodal projector GGUF file.
     *
     * Required when image/file parts are included in user messages.
     */
    mmprojPath?: string;
    /**
     * Maximum context size.
     *
     * This setting is highly dependent on the model and the memory available on
     * the machine. Higher values can consume significant memory and may freeze
     * the machine if set too high. Monitor system memory when increasing this
     * value.
     */
    contextSize?: number;
    memorySafety?: LlamaCppMemorySafetyConfig;
    memory?: LlamaCppModelMemoryInfo;
    gpuLayers?: number;
    threads?: number;
    /**
     * Enable verbose debug output from llama.cpp.
     * Default: false
     */
    debug?: boolean;
    /**
     * Print the final chat-template-rendered prompt sent to llama.cpp to stderr.
     *
     * This may include private user data. Intended for local debugging only.
     * Default: false
     */
    logPrompts?: boolean;
    /**
     * Chat template to use for formatting messages.
     * - "auto" (default): Use the template embedded in the GGUF model file
     * - Template name: Use a specific built-in template (e.g., "llama3", "chatml", "gemma")
     *
     * Available templates: chatml, llama2, llama2-sys, llama3, llama4, mistral-v1,
     * mistral-v3, mistral-v7, phi3, phi4, gemma, falcon3, zephyr, deepseek, deepseek2,
     * deepseek3, command-r, and more.
     */
    chatTemplate?: string;
    /**
     * Extract model thinking into AI SDK reasoning parts.
     */
    reasoning?: LlamaCppReasoningConfig;
    cache?: LlamaCppCacheConfig;
}
export interface LlamaCppGenerationConfig {
    maxTokens?: number;
    temperature?: number;
    topP?: number;
    topK?: number;
    stopSequences?: string[];
}
interface ResolvedReasoningConfig {
    opening: string;
    closing: string;
    promptPrefix?: string;
    budgetTokens?: number;
}
export interface ParsedReasoningPart {
    type: "text" | "reasoning";
    text: string;
}
export declare function resolveReasoningConfig(reasoning?: LlamaCppReasoningConfig, effort?: LlamaCppReasoningEffort): ResolvedReasoningConfig | undefined;
export declare function splitReasoningContent(text: string, reasoning: Pick<ResolvedReasoningConfig, "opening" | "closing">): ParsedReasoningPart[];
export declare function convertFinishReason(reason: string): LanguageModelV4FinishReason;
export declare function convertUsage(promptTokens: number, completionTokens: number, cache?: {
    read?: number;
    write?: number;
}): LanguageModelV4Usage;
/**
 * Represents a parsed tool call from the model output.
 */
export interface ParsedToolCall {
    id: string;
    name: string;
    arguments: Record<string, unknown>;
}
/**
 * Generate a GBNF grammar for tool calls based on the provided tool definitions.
 * This grammar constrains the model to produce valid JSON tool calls.
 */
export declare function generateToolCallGrammar(tools: LanguageModelV4FunctionTool[]): string;
/**
 * Parse the model output to extract tool calls.
 * Supports multiple formats:
 * - Single tool call: {"name": "...", "arguments": {...}}
 * - Array of tool calls: [{"name": "...", "arguments": {...}}, ...]
 * - Legacy format with tool_calls wrapper: {"tool_calls": [...]}
 * Returns null if the output is not a valid tool call JSON.
 */
export declare function parseToolCalls(text: string): ParsedToolCall[] | null;
/**
 * Build a system prompt that instructs the model to use tools.
 */
export declare function buildToolSystemPrompt(tools: LanguageModelV4FunctionTool[], toolChoice?: LanguageModelV4ToolChoice): string;
/**
 * Convert AI SDK messages to simple role/content format for the native layer.
 * The native layer will apply the appropriate chat template.
 */
export declare function convertMessages(messages: LanguageModelV4Message[], reasoning?: Pick<ResolvedReasoningConfig, "promptPrefix">): ChatMessage[];
export declare class LlamaCppLanguageModel implements LanguageModelV4 {
    readonly specificationVersion: "v4";
    readonly provider = "llama.cpp";
    readonly modelId: string;
    readonly contextSize: number;
    /**
     * Supported URL patterns - empty since we only support local files
     */
    readonly supportedUrls: Record<string, RegExp[]>;
    private modelHandle;
    private readonly config;
    private initPromise;
    private loadedContextSize;
    private readonly activeRequests;
    constructor(config: LlamaCppModelConfig);
    private ensureModelLoaded;
    dispose(): Promise<void>;
    private prepareGenerationRequest;
    doGenerate(options: LanguageModelV4CallOptions): Promise<LanguageModelV4GenerateResult>;
    doStream(options: LanguageModelV4CallOptions): Promise<LanguageModelV4StreamResult>;
    private runWithAbortSignal;
}
export {};
//# sourceMappingURL=llama-cpp-language-model.d.ts.map