import 'dart:io';

class AiConstants {
  AiConstants._();

  // Server ports
  static const int chatServerPort = 8080;
  static const int embeddingServerPort = 8081;

  // API endpoints
  static String get chatApiUrl =>
      'http://localhost:$chatServerPort/v1/chat/completions';
  static String get embeddingApiUrl =>
      'http://localhost:$embeddingServerPort/v1/embeddings';
  static String chatHealthUrl = 'http://localhost:$chatServerPort/health';
  static String embeddingHealthUrl =
      'http://localhost:$embeddingServerPort/health';

  // Model files
  static const String qwenModelFile = 'qwen3-4b-instruct-2507-q4_k_m.gguf';
  static const String nomicModelFile = 'nomic-embed-text-v1.5.Q4_K_M.gguf';
  static const String nomicModelName = 'nomic-embed-text-v1.5.Q4_K_M';

  // llama.cpp binary
  static const String llamaBuild = 'b8130';
  static String get serverBinaryName =>
      Platform.isWindows ? 'llama-server.exe' : 'llama-server';

  // Model download URLs
  static const String nomicModelUrl =
      'https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF/resolve/main/nomic-embed-text-v1.5.Q4_K_M.gguf';
  static const String qwenModelUrl =
      'https://huggingface.co/bartowski/Qwen_Qwen3-4B-Instruct-2507-GGUF/resolve/main/Qwen_Qwen3-4B-Instruct-2507-Q4_K_M.gguf';

  // Human-readable model labels (single source of truth for UI copy)
  static const String chatModelLabel = 'Qwen3 4B Instruct Q4_K_M';
  static const String embeddingModelLabel = 'Nomic Embed Text v1.5 Q4_K_M';

  // Minimum plausible file sizes — a partially downloaded file is smaller.
  static const int qwenModelMinBytes = 2_000_000_000; // full file ≈ 2.4 GB
  static const int nomicModelMinBytes = 50_000_000; // full file ≈ 84 MB
  static const int serverBinaryMinBytes = 1_000_000;

  // Server configuration
  static const int qwenContextSize = 8192;
  static const int nomicContextSize = 2048;
  static const int gpuLayerOffload = 99;
  static const double chatTemperature = 0.7;
  static const int chatMaxTokens = 1536;

  /// How many most-recent messages are sent back to the model as history.
  static const int chatHistoryLimit = 12;

  /// Nomic v1.5 requires task prefixes for good retrieval quality.
  static const String nomicDocumentPrefix = 'search_document: ';
  static const String nomicQueryPrefix = 'search_query: ';

  // Timeouts
  static const Duration serverConnectTimeout = Duration(seconds: 30);
  static const Duration serverReceiveTimeout = Duration(minutes: 5);
  static const Duration embeddingConnectTimeout = Duration(seconds: 10);
  static const Duration embeddingReceiveTimeout = Duration(seconds: 60);

  /// Cold-loading a multi-GB model can take minutes: on the first Vulkan
  /// run llama.cpp additionally compiles GPU pipelines for the device.
  static const Duration serverStartupTimeout = Duration(minutes: 5);

  /// CPU fallback loads via mmap and needs no pipeline compilation.
  static const Duration serverStartupTimeoutCpu = Duration(minutes: 3);
  static const Duration healthPollInterval = Duration(milliseconds: 750);

  // System prompt
  static const String baseSystemPrompt =
      'You are the AI coach inside TrainHub, a training hub for ITF Taekwon-Do. '
      'You help with: designing drills and training plans (patterns/tul, '
      'sparring/matsogi, kicks, conditioning), analysing technique, belt '
      'grading preparation, and sports science questions. '
      'Use correct ITF terminology (e.g. dollyo chagi, yop chagi, tul names). '
      'Always answer in the same language the user writes in. '
      'Be concise and practical — coaches read this between rounds.';
}