llama-cpp-capacitor 0.0.13 → 0.0.21

package/dist/esm/definitions.js.map

@@ -1 +1 @@
-{"version":3,"file":"definitions.js","sourceRoot":"","sources":["../../src/definitions.ts"],"names":[],"mappings":"","sourcesContent":["// Native parameter types that match llama.rn exactly\r\nexport interface NativeEmbeddingParams {\r\n  embd_normalize?: number;\r\n}\r\n\r\nexport interface NativeContextParams {\r\n  model: string;\r\n  /**\r\n   * Chat template to override the default one from the model.\r\n   */\r\n  chat_template?: string;\r\n\r\n  is_model_asset?: boolean;\r\n  use_progress_callback?: boolean;\r\n\r\n  n_ctx?: number;\r\n  n_batch?: number;\r\n  n_ubatch?: number;\r\n\r\n  n_threads?: number;\r\n\r\n  /**\r\n   * Number of layers to store in VRAM (Currently only for iOS)\r\n   */\r\n  n_gpu_layers?: number;\r\n  /**\r\n   * Skip GPU devices (iOS only)\r\n   */\r\n  no_gpu_devices?: boolean;\r\n\r\n  /**\r\n   * Enable flash attention, only recommended in GPU device (Experimental in llama.cpp)\r\n   */\r\n  flash_attn?: boolean;\r\n\r\n  /**\r\n   * KV cache data type for the K (Experimental in llama.cpp)\r\n   */\r\n  cache_type_k?: string;\r\n  /**\r\n   * KV cache data type for the V (Experimental in llama.cpp)\r\n   */\r\n  cache_type_v?: string;\r\n\r\n  use_mlock?: boolean;\r\n  use_mmap?: boolean;\r\n  vocab_only?: boolean;\r\n\r\n  /**\r\n   * Single LoRA adapter path\r\n   */\r\n  lora?: string;\r\n  /**\r\n   * Single LoRA adapter scale\r\n   */\r\n  lora_scaled?: number;\r\n  /**\r\n   * LoRA adapter list\r\n   */\r\n  lora_list?: Array<{ path: string; scaled?: number }>;\r\n\r\n  rope_freq_base?: number;\r\n  rope_freq_scale?: number;\r\n\r\n  pooling_type?: number;\r\n\r\n  /**\r\n   * Enable context shifting to handle prompts larger than context size\r\n   */\r\n  ctx_shift?: boolean;\r\n\r\n  /**\r\n   * Use a unified buffer across the input sequences when computing the attention.\r\n   * Try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix.\r\n   */\r\n  kv_unified?: boolean;\r\n\r\n  /**\r\n   * Use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)\r\n   */\r\n  swa_full?: boolean;\r\n\r\n  /**\r\n   * Number of layers to keep MoE weights on CPU\r\n   */\r\n  n_cpu_moe?: number;\r\n\r\n  // Embedding params\r\n  embedding?: boolean;\r\n  embd_normalize?: number;\r\n}\r\n\r\nexport interface NativeCompletionParams {\r\n  prompt: string;\r\n  n_threads?: number;\r\n  /**\r\n   * Enable Jinja. Default: true if supported by the model\r\n   */\r\n  jinja?: boolean;\r\n  /**\r\n   * JSON schema for convert to grammar for structured JSON output.\r\n   * It will be override by grammar if both are set.\r\n   */\r\n  json_schema?: string;\r\n  /**\r\n   * Set grammar for grammar-based sampling.  Default: no grammar\r\n   */\r\n  grammar?: string;\r\n  /**\r\n   * Lazy grammar sampling, trigger by grammar_triggers. Default: false\r\n   */\r\n  grammar_lazy?: boolean;\r\n  /**\r\n   * Enable thinking if jinja is enabled. Default: true\r\n   */\r\n  enable_thinking?: boolean;\r\n  /**\r\n   * Force thinking to be open. Default: false\r\n   */\r\n  thinking_forced_open?: boolean;\r\n  /**\r\n   * Lazy grammar triggers. Default: []\r\n   */\r\n  grammar_triggers?: Array<{\r\n    type: number;\r\n    value: string;\r\n    token: number;\r\n  }>;\r\n  preserved_tokens?: Array<string>;\r\n  chat_format?: number;\r\n  reasoning_format?: string;\r\n  /**\r\n   * Path to an image file to process before generating text.\r\n   * When provided, the image will be processed and added to the context.\r\n   * Requires multimodal support to be enabled via initMultimodal.\r\n   */\r\n  media_paths?: Array<string>;\r\n  /**\r\n   * Specify a JSON array of stopping strings.\r\n   * These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`\r\n   */\r\n  stop?: Array<string>;\r\n  /**\r\n   * Set the maximum number of tokens to predict when generating text.\r\n   * **Note:** May exceed the set limit slightly if the last token is a partial multibyte character.\r\n   * When 0,no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.\r\n   */\r\n  n_predict?: number;\r\n  /**\r\n   * If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings.\r\n   * Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings.\r\n   * Default: `0`\r\n   */\r\n  n_probs?: number;\r\n  /**\r\n   * Limit the next token selection to the K most probable tokens.  Default: `40`\r\n   */\r\n  top_k?: number;\r\n  /**\r\n   * Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. Default: `0.95`\r\n   */\r\n  top_p?: number;\r\n  /**\r\n   * The minimum probability for a token to be considered, relative to the probability of the most likely token. Default: `0.05`\r\n   */\r\n  min_p?: number;\r\n  /**\r\n   * Set the chance for token removal via XTC sampler. Default: `0.0`, which is disabled.\r\n   */\r\n  xtc_probability?: number;\r\n  /**\r\n   * Set a minimum probability threshold for tokens to be removed via XTC sampler. Default: `0.1` (> `0.5` disables XTC)\r\n   */\r\n  xtc_threshold?: number;\r\n  /**\r\n   * Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.\r\n   */\r\n  typical_p?: number;\r\n  /**\r\n   * Adjust the randomness of the generated text. Default: `0.8`\r\n   */\r\n  temperature?: number;\r\n  /**\r\n   * Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.\r\n   */\r\n  penalty_last_n?: number;\r\n  /**\r\n   * Control the repetition of token sequences in the generated text. Default: `1.0`\r\n   */\r\n  penalty_repeat?: number;\r\n  /**\r\n   * Repeat alpha frequency penalty. Default: `0.0`, which is disabled.\r\n   */\r\n  penalty_freq?: number;\r\n  /**\r\n   * Repeat alpha presence penalty. Default: `0.0`, which is disabled.\r\n   */\r\n  penalty_present?: number;\r\n  /**\r\n   * Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.\r\n   */\r\n  mirostat?: number;\r\n  /**\r\n   * Set the Mirostat target entropy, parameter tau. Default: `5.0`\r\n   */\r\n  mirostat_tau?: number;\r\n  /**\r\n   * Set the Mirostat learning rate, parameter eta. Default: `0.1`\r\n   */\r\n  mirostat_eta?: number;\r\n  /**\r\n   * Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.\r\n   */\r\n  dry_multiplier?: number;\r\n  /**\r\n   * Set the DRY repetition penalty base value. Default: `1.75`\r\n   */\r\n  dry_base?: number;\r\n  /**\r\n   * Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`\r\n   */\r\n  dry_allowed_length?: number;\r\n  /**\r\n   * How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.\r\n   */\r\n  dry_penalty_last_n?: number;\r\n  /**\r\n   * Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\\n', ':', '\"', '*']`\r\n   */\r\n  dry_sequence_breakers?: Array<string>;\r\n  /**\r\n   * Top n sigma sampling as described in academic paper \"Top-nσ: Not All Logits Are You Need\" https://arxiv.org/pdf/2411.07641. Default: `-1.0` (Disabled)\r\n   */\r\n  top_n_sigma?: number;\r\n\r\n  /**\r\n   * Ignore end of stream token and continue generating. Default: `false`\r\n   */\r\n  ignore_eos?: boolean;\r\n  /**\r\n   * Modify the likelihood of a token appearing in the generated text completion.\r\n   * For example, use `\"logit_bias\": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `\"logit_bias\": [[15043,-1.0]]` to decrease its likelihood.\r\n   * Setting the value to false, `\"logit_bias\": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings,\r\n   * e.g.`[[\"Hello, World!\",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does.\r\n   * Default: `[]`\r\n   */\r\n  logit_bias?: Array<Array<number>>;\r\n  /**\r\n   * Set the random number generator (RNG) seed. Default: `-1`, which is a random seed.\r\n   */\r\n  seed?: number;\r\n\r\n  /**\r\n   * Guide tokens for the completion.\r\n   * Help prevent hallucinations by forcing the TTS to use the correct words.\r\n   * Default: `[]`\r\n   */\r\n  guide_tokens?: Array<number>;\r\n\r\n  emit_partial_completion: boolean;\r\n}\r\n\r\nexport interface NativeCompletionTokenProbItem {\r\n  tok_str: string;\r\n  prob: number;\r\n}\r\n\r\nexport interface NativeCompletionTokenProb {\r\n  content: string;\r\n  probs: Array<NativeCompletionTokenProbItem>;\r\n}\r\n\r\nexport interface NativeCompletionResultTimings {\r\n  prompt_n: number;\r\n  prompt_ms: number;\r\n  prompt_per_token_ms: number;\r\n  prompt_per_second: number;\r\n  predicted_n: number;\r\n  predicted_ms: number;\r\n  predicted_per_token_ms: number;\r\n  predicted_per_second: number;\r\n}\r\n\r\nexport interface NativeCompletionResult {\r\n  /**\r\n   * Original text (Ignored reasoning_content / tool_calls)\r\n   */\r\n  text: string;\r\n  /**\r\n   * Reasoning content (parsed for reasoning model)\r\n   */\r\n  reasoning_content: string;\r\n  /**\r\n   * Tool calls\r\n   */\r\n  tool_calls: Array<{\r\n    type: 'function';\r\n    function: {\r\n      name: string;\r\n      arguments: string;\r\n    };\r\n    id?: string;\r\n  }>;\r\n  /**\r\n   * Content text (Filtered text by reasoning_content / tool_calls)\r\n   */\r\n  content: string;\r\n\r\n  chat_format: number;\r\n\r\n  tokens_predicted: number;\r\n  tokens_evaluated: number;\r\n  truncated: boolean;\r\n  stopped_eos: boolean;\r\n  stopped_word: string;\r\n  stopped_limit: number;\r\n  stopping_word: string;\r\n  context_full: boolean;\r\n  interrupted: boolean;\r\n  tokens_cached: number;\r\n  timings: NativeCompletionResultTimings;\r\n\r\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\r\n  audio_tokens?: Array<number>;\r\n}\r\n\r\nexport interface NativeTokenizeResult {\r\n  tokens: Array<number>;\r\n  /**\r\n   * Whether the tokenization contains images\r\n   */\r\n  has_images: boolean;\r\n  /**\r\n   * Bitmap hashes of the images\r\n   */\r\n  bitmap_hashes: Array<number>;\r\n  /**\r\n   * Chunk positions of the text and images\r\n   */\r\n  chunk_pos: Array<number>;\r\n  /**\r\n   * Chunk positions of the images\r\n   */\r\n  chunk_pos_images: Array<number>;\r\n}\r\n\r\nexport interface NativeEmbeddingResult {\r\n  embedding: Array<number>;\r\n}\r\n\r\nexport interface NativeLlamaContext {\r\n  contextId: number;\r\n  model: {\r\n    desc: string;\r\n    size: number;\r\n    nEmbd: number;\r\n    nParams: number;\r\n    chatTemplates: {\r\n      llamaChat: boolean; // Chat template in llama-chat.cpp\r\n      minja: {\r\n        // Chat template supported by minja.hpp\r\n        default: boolean;\r\n        defaultCaps: {\r\n          tools: boolean;\r\n          toolCalls: boolean;\r\n          toolResponses: boolean;\r\n          systemRole: boolean;\r\n          parallelToolCalls: boolean;\r\n          toolCallId: boolean;\r\n        };\r\n        toolUse: boolean;\r\n        toolUseCaps: {\r\n          tools: boolean;\r\n          toolCalls: boolean;\r\n          toolResponses: boolean;\r\n          systemRole: boolean;\r\n          parallelToolCalls: boolean;\r\n          toolCallId: boolean;\r\n        };\r\n      };\r\n    };\r\n    metadata: Object;\r\n    isChatTemplateSupported: boolean; // Deprecated\r\n  };\r\n  /**\r\n   * Loaded library name for Android\r\n   */\r\n  androidLib?: string;\r\n  gpu: boolean;\r\n  reasonNoGPU: string;\r\n}\r\n\r\nexport interface NativeSessionLoadResult {\r\n  tokens_loaded: number;\r\n  prompt: string;\r\n}\r\n\r\nexport interface NativeLlamaMessagePart {\r\n  type: 'text';\r\n  text: string;\r\n}\r\n\r\nexport interface NativeLlamaChatMessage {\r\n  role: string;\r\n  content: string | Array<NativeLlamaMessagePart>;\r\n}\r\n\r\nexport interface FormattedChatResult {\r\n  type: 'jinja' | 'llama-chat';\r\n  prompt: string;\r\n  has_media: boolean;\r\n  media_paths?: Array<string>;\r\n}\r\n\r\nexport interface JinjaFormattedChatResult extends FormattedChatResult {\r\n  chat_format?: number;\r\n  grammar?: string;\r\n  grammar_lazy?: boolean;\r\n  grammar_triggers?: Array<{\r\n    type: number;\r\n    value: string;\r\n    token: number;\r\n  }>;\r\n  thinking_forced_open?: boolean;\r\n  preserved_tokens?: Array<string>;\r\n  additional_stops?: Array<string>;\r\n}\r\n\r\nexport interface NativeImageProcessingResult {\r\n  success: boolean;\r\n  prompt: string;\r\n  error?: string;\r\n}\r\n\r\nexport interface NativeRerankParams {\r\n  normalize?: number;\r\n}\r\n\r\nexport interface NativeRerankResult {\r\n  score: number;\r\n  index: number;\r\n}\r\n\r\n// High-level types for the plugin interface\r\nexport interface LlamaCppMessagePart {\r\n  type: string;\r\n  text?: string;\r\n  image_url?: {\r\n    url?: string;\r\n  };\r\n  input_audio?: {\r\n    format: string;\r\n    data?: string;\r\n    url?: string;\r\n  };\r\n}\r\n\r\nexport interface LlamaCppOAICompatibleMessage {\r\n  role: string;\r\n  content?: string | LlamaCppMessagePart[];\r\n}\r\n\r\nexport interface ToolCall {\r\n  type: 'function';\r\n  id?: string;\r\n  function: {\r\n    name: string;\r\n    arguments: string; // JSON string\r\n  };\r\n}\r\n\r\nexport interface TokenData {\r\n  token: string;\r\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\r\n  // Parsed content from accumulated text\r\n  content?: string;\r\n  reasoning_content?: string;\r\n  tool_calls?: Array<ToolCall>;\r\n  accumulated_text?: string;\r\n}\r\n\r\nexport interface ContextParams extends Omit<\r\n  NativeContextParams,\r\n  'cache_type_k' | 'cache_type_v' | 'pooling_type'\r\n> {\r\n  cache_type_k?:\r\n    | 'f16'\r\n    | 'f32'\r\n    | 'q8_0'\r\n    | 'q4_0'\r\n    | 'q4_1'\r\n    | 'iq4_nl'\r\n    | 'q5_0'\r\n    | 'q5_1';\r\n  cache_type_v?:\r\n    | 'f16'\r\n    | 'f32'\r\n    | 'q8_0'\r\n    | 'q4_0'\r\n    | 'q4_1'\r\n    | 'iq4_nl'\r\n    | 'q5_0'\r\n    | 'q5_1';\r\n  pooling_type?: 'none' | 'mean' | 'cls' | 'last' | 'rank';\r\n}\r\n\r\nexport interface EmbeddingParams extends NativeEmbeddingParams {}\r\n\r\nexport interface RerankParams {\r\n  normalize?: number;\r\n}\r\n\r\nexport interface RerankResult {\r\n  score: number;\r\n  index: number;\r\n  document?: string;\r\n}\r\n\r\nexport interface CompletionResponseFormat {\r\n  type: 'text' | 'json_object' | 'json_schema';\r\n  json_schema?: {\r\n    strict?: boolean;\r\n    schema: object;\r\n  };\r\n  schema?: object; // for json_object type\r\n}\r\n\r\nexport interface CompletionBaseParams {\r\n  prompt?: string;\r\n  messages?: LlamaCppOAICompatibleMessage[];\r\n  chatTemplate?: string; // deprecated\r\n  chat_template?: string;\r\n  jinja?: boolean;\r\n  tools?: object;\r\n  parallel_tool_calls?: object;\r\n  tool_choice?: string;\r\n  response_format?: CompletionResponseFormat;\r\n  media_paths?: string[];\r\n  add_generation_prompt?: boolean;\r\n  /*\r\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\r\n   */\r\n  now?: string | number;\r\n  chat_template_kwargs?: Record<string, string>;\r\n  /**\r\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\r\n   * Used for if last assistant message is for prefill purpose\r\n   */\r\n  prefill_text?: string;\r\n}\r\n\r\nexport interface CompletionParams extends Omit<\r\n  NativeCompletionParams,\r\n  'emit_partial_completion' | 'prompt'\r\n> {\r\n  prompt?: string;\r\n  messages?: LlamaCppOAICompatibleMessage[];\r\n  chatTemplate?: string; // deprecated\r\n  chat_template?: string;\r\n  jinja?: boolean;\r\n  tools?: object;\r\n  parallel_tool_calls?: object;\r\n  tool_choice?: string;\r\n  response_format?: CompletionResponseFormat;\r\n  media_paths?: string[];\r\n  add_generation_prompt?: boolean;\r\n  /*\r\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\r\n   */\r\n  now?: string | number;\r\n  chat_template_kwargs?: Record<string, string>;\r\n  /**\r\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\r\n   * Used for if last assistant message is for prefill purpose\r\n   */\r\n  prefill_text?: string;\r\n}\r\n\r\nexport interface BenchResult {\r\n  modelDesc: string;\r\n  modelSize: number;\r\n  modelNParams: number;\r\n  ppAvg: number;\r\n  ppStd: number;\r\n  tgAvg: number;\r\n  tgStd: number;\r\n}\r\n\r\n// Main plugin interface\r\nexport interface LlamaCppPlugin {\r\n  // Core initialization and management\r\n  toggleNativeLog(options: { enabled: boolean }): Promise<void>;\r\n  setContextLimit(options: { limit: number }): Promise<void>;\r\n  modelInfo(options: { path: string; skip?: string[] }): Promise<Object>;\r\n  initContext(options: { contextId: number; params: NativeContextParams }): Promise<NativeLlamaContext>;\r\n  releaseContext(options: { contextId: number }): Promise<void>;\r\n  releaseAllContexts(): Promise<void>;\r\n\r\n  // Chat and completion\r\n  getFormattedChat(options: {\r\n    contextId: number;\r\n    messages: string;\r\n    chatTemplate?: string;\r\n    params?: {\r\n      jinja?: boolean;\r\n      json_schema?: string;\r\n      tools?: string;\r\n      parallel_tool_calls?: string;\r\n      tool_choice?: string;\r\n      enable_thinking?: boolean;\r\n      add_generation_prompt?: boolean;\r\n      now?: string;\r\n      chat_template_kwargs?: string;\r\n    };\r\n  }): Promise<JinjaFormattedChatResult | string>;\r\n\r\n  completion(options: {\r\n    contextId: number;\r\n    params: NativeCompletionParams;\r\n  }): Promise<NativeCompletionResult>;\r\n\r\n  stopCompletion(options: { contextId: number }): Promise<void>;\r\n\r\n  // Session management\r\n  loadSession(options: {\r\n    contextId: number;\r\n    filepath: string;\r\n  }): Promise<NativeSessionLoadResult>;\r\n\r\n  saveSession(options: {\r\n    contextId: number;\r\n    filepath: string;\r\n    size: number;\r\n  }): Promise<number>;\r\n\r\n  // Tokenization\r\n  tokenize(options: {\r\n    contextId: number;\r\n    text: string;\r\n    imagePaths?: Array<string>;\r\n  }): Promise<NativeTokenizeResult>;\r\n\r\n  detokenize(options: {\r\n    contextId: number;\r\n    tokens: number[];\r\n  }): Promise<string>;\r\n\r\n  // Embeddings and reranking\r\n  embedding(options: {\r\n    contextId: number;\r\n    text: string;\r\n    params: NativeEmbeddingParams;\r\n  }): Promise<NativeEmbeddingResult>;\r\n\r\n  rerank(options: {\r\n    contextId: number;\r\n    query: string;\r\n    documents: Array<string>;\r\n    params?: NativeRerankParams;\r\n  }): Promise<Array<NativeRerankResult>>;\r\n\r\n  // Benchmarking\r\n  bench(options: {\r\n    contextId: number;\r\n    pp: number;\r\n    tg: number;\r\n    pl: number;\r\n    nr: number;\r\n  }): Promise<string>;\r\n\r\n  // LoRA adapters\r\n  applyLoraAdapters(options: {\r\n    contextId: number;\r\n    loraAdapters: Array<{ path: string; scaled?: number }>;\r\n  }): Promise<void>;\r\n\r\n  removeLoraAdapters(options: { contextId: number }): Promise<void>;\r\n\r\n  getLoadedLoraAdapters(options: {\r\n    contextId: number;\r\n  }): Promise<Array<{ path: string; scaled?: number }>>;\r\n\r\n  // Multimodal methods\r\n  initMultimodal(options: {\r\n    contextId: number;\r\n    params: {\r\n      path: string;\r\n      use_gpu: boolean;\r\n    };\r\n  }): Promise<boolean>;\r\n\r\n  isMultimodalEnabled(options: {\r\n    contextId: number;\r\n  }): Promise<boolean>;\r\n\r\n  getMultimodalSupport(options: {\r\n    contextId: number;\r\n  }): Promise<{\r\n    vision: boolean;\r\n    audio: boolean;\r\n  }>;\r\n\r\n  releaseMultimodal(options: {\r\n    contextId: number;\r\n  }): Promise<void>;\r\n\r\n  // TTS methods\r\n  initVocoder(options: {\r\n    contextId: number;\r\n    params: {\r\n      path: string;\r\n      n_batch?: number;\r\n    };\r\n  }): Promise<boolean>;\r\n\r\n  isVocoderEnabled(options: { contextId: number }): Promise<boolean>;\r\n\r\n  getFormattedAudioCompletion(options: {\r\n    contextId: number;\r\n    speakerJsonStr: string;\r\n    textToSpeak: string;\r\n  }): Promise<{\r\n    prompt: string;\r\n    grammar?: string;\r\n  }>;\r\n\r\n  getAudioCompletionGuideTokens(options: {\r\n    contextId: number;\r\n    textToSpeak: string;\r\n  }): Promise<Array<number>>;\r\n\r\n  decodeAudioTokens(options: {\r\n    contextId: number;\r\n    tokens: number[];\r\n  }): Promise<Array<number>>;\r\n\r\n  releaseVocoder(options: { contextId: number }): Promise<void>;\r\n\r\n  // Events\r\n  addListener(eventName: string, listenerFunc: (data: any) => void): Promise<void>;\r\n  removeAllListeners(eventName: string): Promise<void>;\r\n}\r\n"]}
+{"version":3,"file":"definitions.js","sourceRoot":"","sources":["../../src/definitions.ts"],"names":[],"mappings":"","sourcesContent":["// Native parameter types that match llama.rn exactly\r\nexport interface NativeEmbeddingParams {\r\n  embd_normalize?: number;\r\n}\r\n\r\nexport interface NativeContextParams {\r\n  model: string;\r\n  /**\r\n   * Chat template to override the default one from the model.\r\n   */\r\n  chat_template?: string;\r\n\r\n  is_model_asset?: boolean;\r\n  use_progress_callback?: boolean;\r\n\r\n  n_ctx?: number;\r\n  n_batch?: number;\r\n  n_ubatch?: number;\r\n\r\n  n_threads?: number;\r\n\r\n  /**\r\n   * Path to draft model for speculative decoding (mobile optimization)\r\n   */\r\n  draft_model?: string;\r\n  /**\r\n   * Number of tokens to predict speculatively (default: 3 for mobile)\r\n   */\r\n  speculative_samples?: number;\r\n  /**\r\n   * Enable mobile-optimized speculative decoding\r\n   */\r\n  mobile_speculative?: boolean;\r\n\r\n  /**\r\n   * Number of layers to store in VRAM (Currently only for iOS)\r\n   */\r\n  n_gpu_layers?: number;\r\n  /**\r\n   * Skip GPU devices (iOS only)\r\n   */\r\n  no_gpu_devices?: boolean;\r\n\r\n  /**\r\n   * Enable flash attention, only recommended in GPU device (Experimental in llama.cpp)\r\n   */\r\n  flash_attn?: boolean;\r\n\r\n  /**\r\n   * KV cache data type for the K (Experimental in llama.cpp)\r\n   */\r\n  cache_type_k?: string;\r\n  /**\r\n   * KV cache data type for the V (Experimental in llama.cpp)\r\n   */\r\n  cache_type_v?: string;\r\n\r\n  use_mlock?: boolean;\r\n  use_mmap?: boolean;\r\n  vocab_only?: boolean;\r\n\r\n  /**\r\n   * Single LoRA adapter path\r\n   */\r\n  lora?: string;\r\n  /**\r\n   * Single LoRA adapter scale\r\n   */\r\n  lora_scaled?: number;\r\n  /**\r\n   * LoRA adapter list\r\n   */\r\n  lora_list?: Array<{ path: string; scaled?: number }>;\r\n\r\n  rope_freq_base?: number;\r\n  rope_freq_scale?: number;\r\n\r\n  pooling_type?: number;\r\n\r\n  /**\r\n   * Enable context shifting to handle prompts larger than context size\r\n   */\r\n  ctx_shift?: boolean;\r\n\r\n  /**\r\n   * Use a unified buffer across the input sequences when computing the attention.\r\n   * Try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix.\r\n   */\r\n  kv_unified?: boolean;\r\n\r\n  /**\r\n   * Use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)\r\n   */\r\n  swa_full?: boolean;\r\n\r\n  /**\r\n   * Number of layers to keep MoE weights on CPU\r\n   */\r\n  n_cpu_moe?: number;\r\n\r\n  // Embedding params\r\n  embedding?: boolean;\r\n  embd_normalize?: number;\r\n}\r\n\r\nexport interface NativeCompletionParams {\r\n  prompt: string;\r\n  n_threads?: number;\r\n  /**\r\n   * Enable Jinja. Default: true if supported by the model\r\n   */\r\n  jinja?: boolean;\r\n  /**\r\n   * JSON schema for convert to grammar for structured JSON output.\r\n   * It will be override by grammar if both are set.\r\n   */\r\n  json_schema?: string;\r\n  /**\r\n   * Set grammar for grammar-based sampling (GBNF format). Default: no grammar\r\n   * This will override json_schema if both are provided.\r\n   */\r\n  grammar?: string;\r\n  /**\r\n   * Lazy grammar sampling, trigger by grammar_triggers. Default: false\r\n   */\r\n  grammar_lazy?: boolean;\r\n  /**\r\n   * Enable thinking if jinja is enabled. Default: true\r\n   */\r\n  enable_thinking?: boolean;\r\n  /**\r\n   * Force thinking to be open. Default: false\r\n   */\r\n  thinking_forced_open?: boolean;\r\n  /**\r\n   * Lazy grammar triggers. Default: []\r\n   */\r\n  grammar_triggers?: Array<{\r\n    type: number;\r\n    value: string;\r\n    token: number;\r\n  }>;\r\n  preserved_tokens?: Array<string>;\r\n  chat_format?: number;\r\n  reasoning_format?: string;\r\n  /**\r\n   * Path to an image file to process before generating text.\r\n   * When provided, the image will be processed and added to the context.\r\n   * Requires multimodal support to be enabled via initMultimodal.\r\n   */\r\n  media_paths?: Array<string>;\r\n  /**\r\n   * Specify a JSON array of stopping strings.\r\n   * These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`\r\n   */\r\n  stop?: Array<string>;\r\n  /**\r\n   * Set the maximum number of tokens to predict when generating text.\r\n   * **Note:** May exceed the set limit slightly if the last token is a partial multibyte character.\r\n   * When 0,no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.\r\n   */\r\n  n_predict?: number;\r\n  /**\r\n   * If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings.\r\n   * Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings.\r\n   * Default: `0`\r\n   */\r\n  n_probs?: number;\r\n  /**\r\n   * Limit the next token selection to the K most probable tokens.  Default: `40`\r\n   */\r\n  top_k?: number;\r\n  /**\r\n   * Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. Default: `0.95`\r\n   */\r\n  top_p?: number;\r\n  /**\r\n   * The minimum probability for a token to be considered, relative to the probability of the most likely token. Default: `0.05`\r\n   */\r\n  min_p?: number;\r\n  /**\r\n   * Set the chance for token removal via XTC sampler. Default: `0.0`, which is disabled.\r\n   */\r\n  xtc_probability?: number;\r\n  /**\r\n   * Set a minimum probability threshold for tokens to be removed via XTC sampler. Default: `0.1` (> `0.5` disables XTC)\r\n   */\r\n  xtc_threshold?: number;\r\n  /**\r\n   * Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.\r\n   */\r\n  typical_p?: number;\r\n  /**\r\n   * Adjust the randomness of the generated text. Default: `0.8`\r\n   */\r\n  temperature?: number;\r\n  /**\r\n   * Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.\r\n   */\r\n  penalty_last_n?: number;\r\n  /**\r\n   * Control the repetition of token sequences in the generated text. Default: `1.0`\r\n   */\r\n  penalty_repeat?: number;\r\n  /**\r\n   * Repeat alpha frequency penalty. Default: `0.0`, which is disabled.\r\n   */\r\n  penalty_freq?: number;\r\n  /**\r\n   * Repeat alpha presence penalty. Default: `0.0`, which is disabled.\r\n   */\r\n  penalty_present?: number;\r\n  /**\r\n   * Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.\r\n   */\r\n  mirostat?: number;\r\n  /**\r\n   * Set the Mirostat target entropy, parameter tau. Default: `5.0`\r\n   */\r\n  mirostat_tau?: number;\r\n  /**\r\n   * Set the Mirostat learning rate, parameter eta. Default: `0.1`\r\n   */\r\n  mirostat_eta?: number;\r\n  /**\r\n   * Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.\r\n   */\r\n  dry_multiplier?: number;\r\n  /**\r\n   * Set the DRY repetition penalty base value. Default: `1.75`\r\n   */\r\n  dry_base?: number;\r\n  /**\r\n   * Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`\r\n   */\r\n  dry_allowed_length?: number;\r\n  /**\r\n   * How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.\r\n   */\r\n  dry_penalty_last_n?: number;\r\n  /**\r\n   * Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\\n', ':', '\"', '*']`\r\n   */\r\n  dry_sequence_breakers?: Array<string>;\r\n  /**\r\n   * Top n sigma sampling as described in academic paper \"Top-nσ: Not All Logits Are You Need\" https://arxiv.org/pdf/2411.07641. Default: `-1.0` (Disabled)\r\n   */\r\n  top_n_sigma?: number;\r\n\r\n  /**\r\n   * Ignore end of stream token and continue generating. Default: `false`\r\n   */\r\n  ignore_eos?: boolean;\r\n  /**\r\n   * Modify the likelihood of a token appearing in the generated text completion.\r\n   * For example, use `\"logit_bias\": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `\"logit_bias\": [[15043,-1.0]]` to decrease its likelihood.\r\n   * Setting the value to false, `\"logit_bias\": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings,\r\n   * e.g.`[[\"Hello, World!\",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does.\r\n   * Default: `[]`\r\n   */\r\n  logit_bias?: Array<Array<number>>;\r\n  /**\r\n   * Set the random number generator (RNG) seed. Default: `-1`, which is a random seed.\r\n   */\r\n  seed?: number;\r\n\r\n  /**\r\n   * Guide tokens for the completion.\r\n   * Help prevent hallucinations by forcing the TTS to use the correct words.\r\n   * Default: `[]`\r\n   */\r\n  guide_tokens?: Array<number>;\r\n\r\n  emit_partial_completion: boolean;\r\n}\r\n\r\nexport interface NativeCompletionTokenProbItem {\r\n  tok_str: string;\r\n  prob: number;\r\n}\r\n\r\nexport interface NativeCompletionTokenProb {\r\n  content: string;\r\n  probs: Array<NativeCompletionTokenProbItem>;\r\n}\r\n\r\nexport interface NativeCompletionResultTimings {\r\n  prompt_n: number;\r\n  prompt_ms: number;\r\n  prompt_per_token_ms: number;\r\n  prompt_per_second: number;\r\n  predicted_n: number;\r\n  predicted_ms: number;\r\n  predicted_per_token_ms: number;\r\n  predicted_per_second: number;\r\n}\r\n\r\nexport interface NativeCompletionResult {\r\n  /**\r\n   * Original text (Ignored reasoning_content / tool_calls)\r\n   */\r\n  text: string;\r\n  /**\r\n   * Reasoning content (parsed for reasoning model)\r\n   */\r\n  reasoning_content: string;\r\n  /**\r\n   * Tool calls (parsed from response)\r\n   */\r\n  tool_calls: Array<{\r\n    type: 'function';\r\n    function: {\r\n      name: string;\r\n      arguments: string; // JSON string of arguments\r\n    };\r\n    id?: string;\r\n  }>;\r\n  /**\r\n   * Content text (Filtered text by reasoning_content / tool_calls)\r\n   */\r\n  content: string;\r\n\r\n  chat_format: number;\r\n\r\n  tokens_predicted: number;\r\n  tokens_evaluated: number;\r\n  truncated: boolean;\r\n  stopped_eos: boolean;\r\n  stopped_word: string;\r\n  stopped_limit: number;\r\n  stopping_word: string;\r\n  context_full: boolean;\r\n  interrupted: boolean;\r\n  tokens_cached: number;\r\n  timings: NativeCompletionResultTimings;\r\n\r\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\r\n  audio_tokens?: Array<number>;\r\n}\r\n\r\nexport interface NativeTokenizeResult {\r\n  tokens: Array<number>;\r\n  /**\r\n   * Whether the tokenization contains images\r\n   */\r\n  has_images: boolean;\r\n  /**\r\n   * Bitmap hashes of the images\r\n   */\r\n  bitmap_hashes: Array<number>;\r\n  /**\r\n   * Chunk positions of the text and images\r\n   */\r\n  chunk_pos: Array<number>;\r\n  /**\r\n   * Chunk positions of the images\r\n   */\r\n  chunk_pos_images: Array<number>;\r\n}\r\n\r\nexport interface NativeEmbeddingResult {\r\n  embedding: Array<number>;\r\n}\r\n\r\nexport interface NativeLlamaContext {\r\n  contextId: number;\r\n  model: {\r\n    desc: string;\r\n    size: number;\r\n    nEmbd: number;\r\n    nParams: number;\r\n    chatTemplates: {\r\n      llamaChat: boolean; // Chat template in llama-chat.cpp\r\n      minja: {\r\n        // Chat template supported by minja.hpp\r\n        default: boolean;\r\n        defaultCaps: {\r\n          tools: boolean;\r\n          toolCalls: boolean;\r\n          toolResponses: boolean;\r\n          systemRole: boolean;\r\n          parallelToolCalls: boolean;\r\n          toolCallId: boolean;\r\n        };\r\n        toolUse: boolean;\r\n        toolUseCaps: {\r\n          tools: boolean;\r\n          toolCalls: boolean;\r\n          toolResponses: boolean;\r\n          systemRole: boolean;\r\n          parallelToolCalls: boolean;\r\n          toolCallId: boolean;\r\n        };\r\n      };\r\n    };\r\n    metadata: Object;\r\n    isChatTemplateSupported: boolean; // Deprecated\r\n  };\r\n  /**\r\n   * Loaded library name for Android\r\n   */\r\n  androidLib?: string;\r\n  gpu: boolean;\r\n  reasonNoGPU: string;\r\n}\r\n\r\nexport interface NativeSessionLoadResult {\r\n  tokens_loaded: number;\r\n  prompt: string;\r\n}\r\n\r\nexport interface NativeLlamaMessagePart {\r\n  type: 'text';\r\n  text: string;\r\n}\r\n\r\nexport interface NativeLlamaChatMessage {\r\n  role: string;\r\n  content: string | Array<NativeLlamaMessagePart>;\r\n}\r\n\r\nexport interface FormattedChatResult {\r\n  type: 'jinja' | 'llama-chat';\r\n  prompt: string;\r\n  has_media: boolean;\r\n  media_paths?: Array<string>;\r\n}\r\n\r\nexport interface JinjaFormattedChatResult extends FormattedChatResult {\r\n  chat_format?: number;\r\n  grammar?: string;\r\n  grammar_lazy?: boolean;\r\n  grammar_triggers?: Array<{\r\n    type: number;\r\n    value: string;\r\n    token: number;\r\n  }>;\r\n  thinking_forced_open?: boolean;\r\n  preserved_tokens?: Array<string>;\r\n  additional_stops?: Array<string>;\r\n}\r\n\r\nexport interface NativeImageProcessingResult {\r\n  success: boolean;\r\n  prompt: string;\r\n  error?: string;\r\n}\r\n\r\nexport interface NativeRerankParams {\r\n  normalize?: number;\r\n}\r\n\r\nexport interface NativeRerankResult {\r\n  score: number;\r\n  index: number;\r\n}\r\n\r\n// High-level types for the plugin interface\r\nexport interface LlamaCppMessagePart {\r\n  type: string;\r\n  text?: string;\r\n  image_url?: {\r\n    url?: string;\r\n  };\r\n  input_audio?: {\r\n    format: string;\r\n    data?: string;\r\n    url?: string;\r\n  };\r\n}\r\n\r\nexport interface LlamaCppOAICompatibleMessage {\r\n  role: string;\r\n  content?: string | LlamaCppMessagePart[];\r\n}\r\n\r\nexport interface ToolCall {\r\n  type: 'function';\r\n  id?: string;\r\n  function: {\r\n    name: string;\r\n    arguments: string; // JSON string\r\n  };\r\n}\r\n\r\nexport interface TokenData {\r\n  token: string;\r\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\r\n  // Parsed content from accumulated text\r\n  content?: string;\r\n  reasoning_content?: string;\r\n  tool_calls?: Array<ToolCall>;\r\n  accumulated_text?: string;\r\n}\r\n\r\nexport interface ContextParams extends Omit<\r\n  NativeContextParams,\r\n  'cache_type_k' | 'cache_type_v' | 'pooling_type'\r\n> {\r\n  cache_type_k?:\r\n    | 'f16'\r\n    | 'f32'\r\n    | 'q8_0'\r\n    | 'q4_0'\r\n    | 'q4_1'\r\n    | 'iq4_nl'\r\n    | 'q5_0'\r\n    | 'q5_1';\r\n  cache_type_v?:\r\n    | 'f16'\r\n    | 'f32'\r\n    | 'q8_0'\r\n    | 'q4_0'\r\n    | 'q4_1'\r\n    | 'iq4_nl'\r\n    | 'q5_0'\r\n    | 'q5_1';\r\n  pooling_type?: 'none' | 'mean' | 'cls' | 'last' | 'rank';\r\n}\r\n\r\nexport interface EmbeddingParams extends NativeEmbeddingParams {}\r\n\r\nexport interface RerankParams {\r\n  normalize?: number;\r\n}\r\n\r\nexport interface RerankResult {\r\n  score: number;\r\n  index: number;\r\n  document?: string;\r\n}\r\n\r\nexport interface CompletionResponseFormat {\r\n  type: 'text' | 'json_object' | 'json_schema';\r\n  json_schema?: {\r\n    strict?: boolean;\r\n    schema: object;\r\n  };\r\n  schema?: object; // for json_object type\r\n}\r\n\r\nexport interface CompletionBaseParams {\r\n  prompt?: string;\r\n  messages?: LlamaCppOAICompatibleMessage[];\r\n  chatTemplate?: string; // deprecated\r\n  chat_template?: string;\r\n  jinja?: boolean;\r\n  tools?: object;\r\n  parallel_tool_calls?: object;\r\n  tool_choice?: string;\r\n  response_format?: CompletionResponseFormat;\r\n  media_paths?: string[];\r\n  add_generation_prompt?: boolean;\r\n  /*\r\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\r\n   */\r\n  now?: string | number;\r\n  chat_template_kwargs?: Record<string, string>;\r\n  /**\r\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\r\n   * Used for if last assistant message is for prefill purpose\r\n   */\r\n  prefill_text?: string;\r\n}\r\n\r\nexport interface CompletionParams extends Omit<\r\n  NativeCompletionParams,\r\n  'emit_partial_completion' | 'prompt'\r\n> {\r\n  prompt?: string;\r\n  messages?: LlamaCppOAICompatibleMessage[];\r\n  chatTemplate?: string; // deprecated\r\n  chat_template?: string;\r\n  jinja?: boolean;\r\n  /**\r\n   * GBNF grammar for structured output. Takes precedence over json_schema.\r\n   */\r\n  grammar?: string;\r\n  tools?: object;\r\n  parallel_tool_calls?: object;\r\n  tool_choice?: string;\r\n  response_format?: CompletionResponseFormat;\r\n  media_paths?: string[];\r\n  add_generation_prompt?: boolean;\r\n  /*\r\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\r\n   */\r\n  now?: string | number;\r\n  chat_template_kwargs?: Record<string, string>;\r\n  /**\r\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\r\n   * Used for if last assistant message is for prefill purpose\r\n   */\r\n  prefill_text?: string;\r\n}\r\n\r\nexport interface BenchResult {\r\n  modelDesc: string;\r\n  modelSize: number;\r\n  modelNParams: number;\r\n  ppAvg: number;\r\n  ppStd: number;\r\n  tgAvg: number;\r\n  tgStd: number;\r\n}\r\n\r\n// Main plugin interface\r\nexport interface LlamaCppPlugin {\r\n  // Core initialization and management\r\n  toggleNativeLog(options: { enabled: boolean }): Promise<void>;\r\n  setContextLimit(options: { limit: number }): Promise<void>;\r\n  modelInfo(options: { path: string; skip?: string[] }): Promise<Object>;\r\n  initContext(options: { contextId: number; params: NativeContextParams }): Promise<NativeLlamaContext>;\r\n  releaseContext(options: { contextId: number }): Promise<void>;\r\n  releaseAllContexts(): Promise<void>;\r\n\r\n  // Chat and completion\r\n  getFormattedChat(options: {\r\n    contextId: number;\r\n    messages: string;\r\n    chatTemplate?: string;\r\n    params?: {\r\n      jinja?: boolean;\r\n      json_schema?: string;\r\n      tools?: string;\r\n      parallel_tool_calls?: string;\r\n      tool_choice?: string;\r\n      enable_thinking?: boolean;\r\n      add_generation_prompt?: boolean;\r\n      now?: string;\r\n      chat_template_kwargs?: string;\r\n    };\r\n  }): Promise<JinjaFormattedChatResult | string>;\r\n\r\n  completion(options: {\r\n    contextId: number;\r\n    params: NativeCompletionParams;\r\n  }): Promise<NativeCompletionResult>;\r\n\r\n  stopCompletion(options: { contextId: number }): Promise<void>;\r\n\r\n  // Session management\r\n  loadSession(options: {\r\n    contextId: number;\r\n    filepath: string;\r\n  }): Promise<NativeSessionLoadResult>;\r\n\r\n  saveSession(options: {\r\n    contextId: number;\r\n    filepath: string;\r\n    size: number;\r\n  }): Promise<number>;\r\n\r\n  // Tokenization\r\n  tokenize(options: {\r\n    contextId: number;\r\n    text: string;\r\n    imagePaths?: Array<string>;\r\n  }): Promise<NativeTokenizeResult>;\r\n\r\n  detokenize(options: {\r\n    contextId: number;\r\n    tokens: number[];\r\n  }): Promise<string>;\r\n\r\n  // Embeddings and reranking\r\n  embedding(options: {\r\n    contextId: number;\r\n    text: string;\r\n    params: NativeEmbeddingParams;\r\n  }): Promise<NativeEmbeddingResult>;\r\n\r\n  rerank(options: {\r\n    contextId: number;\r\n    query: string;\r\n    documents: Array<string>;\r\n    params?: NativeRerankParams;\r\n  }): Promise<Array<NativeRerankResult>>;\r\n\r\n  // Benchmarking\r\n  bench(options: {\r\n    contextId: number;\r\n    pp: number;\r\n    tg: number;\r\n    pl: number;\r\n    nr: number;\r\n  }): Promise<string>;\r\n\r\n  // LoRA adapters\r\n  applyLoraAdapters(options: {\r\n    contextId: number;\r\n    loraAdapters: Array<{ path: string; scaled?: number }>;\r\n  }): Promise<void>;\r\n\r\n  removeLoraAdapters(options: { contextId: number }): Promise<void>;\r\n\r\n  getLoadedLoraAdapters(options: {\r\n    contextId: number;\r\n  }): Promise<Array<{ path: string; scaled?: number }>>;\r\n\r\n  // Multimodal methods\r\n  initMultimodal(options: {\r\n    contextId: number;\r\n    params: {\r\n      path: string;\r\n      use_gpu: boolean;\r\n    };\r\n  }): Promise<boolean>;\r\n\r\n  isMultimodalEnabled(options: {\r\n    contextId: number;\r\n  }): Promise<boolean>;\r\n\r\n  getMultimodalSupport(options: {\r\n    contextId: number;\r\n  }): Promise<{\r\n    vision: boolean;\r\n    audio: boolean;\r\n  }>;\r\n\r\n  releaseMultimodal(options: {\r\n    contextId: number;\r\n  }): Promise<void>;\r\n\r\n  // TTS methods\r\n  initVocoder(options: {\r\n    contextId: number;\r\n    params: {\r\n      path: string;\r\n      n_batch?: number;\r\n    };\r\n  }): Promise<boolean>;\r\n\r\n  isVocoderEnabled(options: { contextId: number }): Promise<boolean>;\r\n\r\n  getFormattedAudioCompletion(options: {\r\n    contextId: number;\r\n    speakerJsonStr: string;\r\n    textToSpeak: string;\r\n  }): Promise<{\r\n    prompt: string;\r\n    grammar?: string;\r\n  }>;\r\n\r\n  getAudioCompletionGuideTokens(options: {\r\n    contextId: number;\r\n    textToSpeak: string;\r\n  }): Promise<Array<number>>;\r\n\r\n  decodeAudioTokens(options: {\r\n    contextId: number;\r\n    tokens: number[];\r\n  }): Promise<Array<number>>;\r\n\r\n  releaseVocoder(options: { contextId: number }): Promise<void>;\r\n\r\n  // Model download and management\r\n  downloadModel(options: {\r\n    url: string;\r\n    filename: string;\r\n  }): Promise<string>;\r\n\r\n  getDownloadProgress(options: {\r\n    url: string;\r\n  }): Promise<{\r\n    progress: number;\r\n    completed: boolean;\r\n    failed: boolean;\r\n    errorMessage?: string;\r\n    localPath?: string;\r\n    downloadedBytes: number;\r\n    totalBytes: number;\r\n  }>;\r\n\r\n  cancelDownload(options: {\r\n    url: string;\r\n  }): Promise<boolean>;\r\n\r\n  getAvailableModels(): Promise<Array<{\r\n    name: string;\r\n    path: string;\r\n    size: number;\r\n  }>>;\r\n\r\n  // Grammar utilities\r\n  convertJsonSchemaToGrammar(options: {\r\n    schema: string;\r\n  }): Promise<string>;\r\n\r\n  // Events\r\n  addListener(eventName: string, listenerFunc: (data: any) => void): Promise<void>;\r\n  removeAllListeners(eventName: string): Promise<void>;\r\n}\r\n"]}

package/dist/esm/index.d.ts

@@ -173,6 +173,28 @@ export declare function setContextLimit(limit: number): Promise<void>;
 export declare function loadLlamaModelInfo(model: string): Promise<Object>;
 export declare function initLlama({ model, is_model_asset: isModelAsset, pooling_type: poolingType, lora, lora_list: loraList, ...rest }: ContextParams, onProgress?: (progress: number) => void): Promise<LlamaContext>;
 export declare function releaseAllLlama(): Promise<void>;
+export declare function downloadModel(url: string, filename: string): Promise<string>;
+export declare function getDownloadProgress(url: string): Promise<{
+    progress: number;
+    completed: boolean;
+    failed: boolean;
+    errorMessage?: string;
+    localPath?: string;
+    downloadedBytes: number;
+    totalBytes: number;
+}>;
+export declare function cancelDownload(url: string): Promise<boolean>;
+export declare function getAvailableModels(): Promise<Array<{
+    name: string;
+    path: string;
+    size: number;
+}>>;
+/**
+ * Convert a JSON schema to GBNF grammar format
+ * @param schema JSON schema object
+ * @returns Promise resolving to GBNF grammar string
+ */
+export declare function convertJsonSchemaToGrammar(schema: object): Promise<string>;
 export declare const BuildInfo: {
     number: string;
     commit: string;

package/dist/esm/index.js

@@ -39,6 +39,27 @@ const getJsonSchema = (responseFormat) => {
     }
     return null;
 };
+// Utility function to convert JSON schema to GBNF grammar
+const jsonSchemaToGrammar = async (schema) => {
+    // This will call the native method to convert JSON schema to GBNF
+    // For now, we'll return a basic implementation
+    try {
+        const result = await LlamaCpp.convertJsonSchemaToGrammar({ schema: JSON.stringify(schema) });
+        return result;
+    }
+    catch (error) {
+        console.warn('Failed to convert JSON schema to GBNF, using fallback:', error);
+        // Fallback for basic object structure
+        return `root ::= "{" ws object_content ws "}"
+object_content ::= string_field ("," ws string_field)*
+string_field ::= "\\"" [a-zA-Z_][a-zA-Z0-9_]* "\\"" ws ":" ws value
+value ::= string | number | boolean | "null"
+string ::= "\\"" [^"]* "\\""
+number ::= "-"? [0-9]+ ("." [0-9]+)?
+boolean ::= "true" | "false"
+ws ::= [ \\t\\n]*`;
+    }
+};
 export class LlamaContext {
     constructor({ contextId, gpu, reasonNoGPU, model }) {
         this.gpu = false;
@@ -217,10 +238,23 @@ export class LlamaContext {
         if (!nativeParams.media_paths && params.media_paths) {
             nativeParams.media_paths = params.media_paths;
         }
-        if (nativeParams.response_format && !nativeParams.grammar) {
+        // Handle structured output and grammar
+        if (params.grammar) {
+            // Direct GBNF grammar takes precedence
+            nativeParams.grammar = params.grammar;
+        }
+        else if (nativeParams.response_format && !nativeParams.grammar) {
             const jsonSchema = getJsonSchema(params.response_format);
-            if (jsonSchema)
-                nativeParams.json_schema = JSON.stringify(jsonSchema);
+            if (jsonSchema) {
+                // Try to convert JSON schema to GBNF grammar
+                try {
+                    nativeParams.grammar = await jsonSchemaToGrammar(jsonSchema);
+                }
+                catch (error) {
+                    console.warn('Failed to convert JSON schema to grammar, falling back to json_schema parameter:', error);
+                    nativeParams.json_schema = JSON.stringify(jsonSchema);
+                }
+            }
         }
         let tokenListener = callback &&
             LlamaCpp.addListener(EVENT_ON_TOKEN, (evt) => {
@@ -480,6 +514,14 @@ export async function initLlama(_a, onProgress) {
         console.warn(`[LlamaCpp] initLlama: Invalid cache V type: ${rest.cache_type_v}, falling back to f16`);
         delete rest.cache_type_v;
     }
+    // Log speculative decoding configuration if enabled
+    if (rest.draft_model) {
+        console.log(`🚀 Initializing with speculative decoding:
+      - Main model: ${path}
+      - Draft model: ${rest.draft_model}
+      - Speculative samples: ${rest.speculative_samples || 3}
+      - Mobile optimization: ${rest.mobile_speculative !== false ? 'enabled' : 'disabled'}`);
+    }
     const { gpu, reasonNoGPU, model: modelDetails, androidLib, } = await LlamaCpp.initContext({
         contextId,
         params: Object.assign({ model: path, is_model_asset: !!isModelAsset, use_progress_callback: !!onProgress, pooling_type: poolType, lora: loraPath, lora_list: loraAdapters }, rest),
@@ -499,6 +541,27 @@ export async function initLlama(_a, onProgress) {
 export async function releaseAllLlama() {
     return LlamaCpp.releaseAllContexts();
 }
+// Model download and management functions
+export async function downloadModel(url, filename) {
+    return LlamaCpp.downloadModel({ url, filename });
+}
+export async function getDownloadProgress(url) {
+    return LlamaCpp.getDownloadProgress({ url });
+}
+export async function cancelDownload(url) {
+    return LlamaCpp.cancelDownload({ url });
+}
+export async function getAvailableModels() {
+    return LlamaCpp.getAvailableModels();
+}
+/**
+ * Convert a JSON schema to GBNF grammar format
+ * @param schema JSON schema object
+ * @returns Promise resolving to GBNF grammar string
+ */
+export async function convertJsonSchemaToGrammar(schema) {
+    return jsonSchemaToGrammar(schema);
+}
 export const BuildInfo = {
     number: '1.0.0',
     commit: 'capacitor-llama-cpp',

package/dist/esm/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/index.ts"],"names":[],"mappings":";;AAAA,OAAO,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AA+BjD,YAAY;AACZ,MAAM,CAAC,MAAM,kCAAkC,GAAG,aAAa,CAAC;AAEhE,cAAc;AACd,MAAM,8BAA8B,GAAG,iCAAiC,CAAC;AACzE,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAC3C,MAAM,mBAAmB,GAAG,uBAAuB,CAAC;AAEpD,sBAAsB;AACtB,MAAM,QAAQ,GAAG,cAAc,CAAiB,UAAU,CAAC,CAAC;AAE5D,2BAA2B;AAC3B,MAAM,YAAY,GAAiD,EAAE,CAAC;AAEtE,6BAA6B;AAC7B,QAAQ,CAAC,WAAW,CAAC,mBAAmB,EAAE,CAAC,GAAoC,EAAE,EAAE;IACjF,YAAY,CAAC,OAAO,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;AACpE,CAAC,CAAC,CAAC;AAEH,4CAA4C;AAC5C,kBAAA,QAAQ,aAAR,QAAQ,uBAAR,QAAQ,CAAE,eAAe,+CAAzB,QAAQ,EAAoB,EAAE,OAAO,EAAE,KAAK,EAAE,2CAAG,KAAK,mDAAG,GAAG,EAAE,GAAE,CAAC,EAAE;AAiCnE,MAAM,CAAC,MAAM,iCAAiC,GAAG,kCAAkC,CAAC;AA0BpF,MAAM,eAAe,GAAG;IACtB,KAAK;IACL,KAAK;IACL,MAAM;IACN,MAAM;IACN,MAAM;IACN,MAAM;IACN,QAAQ;IACR,MAAM;IACN,MAAM;CACP,CAAC;AAEF,MAAM,aAAa,GAAG,CAAC,cAAyC,EAAE,EAAE;;IAClE,IAAI,CAAA,cAAc,aAAd,cAAc,uBAAd,cAAc,CAAE,IAAI,MAAK,aAAa,EAAE;QAC1C,aAAO,cAAc,CAAC,WAAW,0CAAE,MAAM,CAAC;KAC3C;IACD,IAAI,CAAA,cAAc,aAAd,cAAc,uBAAd,cAAc,CAAE,IAAI,MAAK,aAAa,EAAE;QAC1C,OAAO,cAAc,CAAC,MAAM,IAAI,EAAE,CAAC;KACpC;IACD,OAAO,IAAI,CAAC;AACd,CAAC,CAAC;AAEF,MAAM,OAAO,YAAY;IAMvB,YAAY,EAAE,SAAS,EAAE,GAAG,EAAE,WAAW,EAAE,KAAK,EAAsB;QAJtE,QAAG,GAAY,KAAK,CAAC;QACrB,gBAAW,GAAW,EAAE,CAAC;QAIvB,IAAI,CAAC,EAAE,GAAG,SAAS,CAAC;QACpB,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;IACrB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CAAC,QAAgB;QAChC,IAAI,IAAI,GAAG,QAAQ,CAAC;QACpB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,QAAQ,CAAC,WAAW,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC;IACtE,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CACf,QAAgB,EAChB,OAA+B;QAE/B,OAAO,QAAQ,CAAC,WAAW,CAAC;YAC1B,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,QAAQ;YACR,IAAI,EAAE,CAAA,OAAO,aAAP,OAAO,uBAAP,OAAO,CAAE,SAAS,KAAI,CAAC,CAAC;SAC/B,CAAC,CAAC;IACL,CAAC;IAED,oBAAoB;QAClB,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC,SAAS,CAAC;IAC9C,CAAC;IAED,gBAAgB;QACd,MAAM,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC;QAC3C,OAAO,CAAC,EAAC,KAAK,aAAL,KAAK,uBAAL,KAAK,CAAE,OAAO,CAAA,IAAI,CAAC,EAAC,KAAK,aAAL,KAAK,uBAAL,KAAK,CAAE,OAAO,CAAA,CAAC;IAC9C,CAAC;IAED,KAAK,CAAC,gBAAgB,CACpB,QAAuC,EACvC,QAAwB,EACxB,MAUC;;QAED,MAAM,UAAU,GAAa,EAAE,CAAC;QAChC,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE;YAChC,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE;gBAC9B,MAAM,OAAO,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE;;oBACvC,4BAA4B;oBAC5B,IAAI,IAAI,CAAC,IAAI,KAAK,WAAW,EAAE;wBAC7B,IAAI,IAAI,GAAG,OAAA,IAAI,CAAC,SAAS,0CAAE,GAAG,KAAI,EAAE,CAAC;wBACrC,IAAI,IAAI,aAAJ,IAAI,uBAAJ,IAAI,CAAE,UAAU,CAAC,SAAS;4BAAG,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;wBACtD,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;wBACtB,OAAO;4BACL,IAAI,EAAE,MAAM;4BACZ,IAAI,EAAE,iCAAiC;yBACxC,CAAC;qBACH;yBAAM,IAAI,IAAI,CAAC,IAAI,KAAK,aAAa,EAAE;wBACtC,MAAM,EAAE,WAAW,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC;wBACpC,IAAI,CAAC,KAAK;4BAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;wBAEvD,MAAM,EAAE,MAAM,EAAE,GAAG,KAAK,CAAC;wBACzB,IAAI,MAAM,IAAI,KAAK,IAAI,MAAM,IAAI,KAAK,EAAE;4BACtC,MAAM,IAAI,KAAK,CAAC,6BAA6B,MAAM,EAAE,CAAC,CAAC;yBACxD;wBACD,IAAI,KAAK,CAAC,GAAG,EAAE;4BACb,MAAM,IAAI,GAAG,KAAK,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;4BAChD,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;yBACvB;6BAAM,IAAI,KAAK,CAAC,IAAI,EAAE;4BACrB,UAAU,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;yBAC7B;wBACD,OAAO;4BACL,IAAI,EAAE,MAAM;4BACZ,IAAI,EAAE,iCAAiC;yBACxC,CAAC;qBACH;oBACD,OAAO,IAAI,CAAC;gBACd,CAAC,CAAC,CAAC;gBAEH,uCACK,GAAG,KACN,OAAO,IACP;aACH;YACD,OAAO,GAAG,CAAC;QACb,CAAC,CAA6B,CAAC;QAE/B,MAAM,QAAQ,GAAG,IAAI,CAAC,gBAAgB,EAAE,KAAI,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,KAAK,CAAA,CAAC;QAC1D,IAAI,IAAI,CAAC;QACT,IAAI,QAAQ;YAAE,IAAI,GAAG,QAAQ,CAAC,CAAC,4BAA4B;QAC3D,MAAM,UAAU,GAAG,aAAa,CAAC,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,eAAe,CAAC,CAAC;QAE1D,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,gBAAgB,CAAC;YAC7C,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,QAAQ,EAAE,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC;YAC9B,YAAY,EAAE,IAAI;YAClB,MAAM,EAAE;gBACN,KAAK,EAAE,QAAQ;gBACf,WAAW,EAAE,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,SAAS;gBAChE,KAAK,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,KAAK,EAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS;gBAC/D,mBAAmB,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,mBAAmB,EAC9C,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,mBAAmB,CAAC;oBAC5C,CAAC,CAAC,SAAS;gBACb,WAAW,EAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,WAAW;gBAChC,eAAe,QAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,eAAe,mCAAI,IAAI;gBAChD,qBAAqB,EAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,qBAAqB;gBACpD,GAAG,EAAE,QAAO,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,GAAG,CAAA,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,GAAG;gBAC1E,oBAAoB,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,oBAAoB,EAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CACjE,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,oBAAoB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,EAAE,EAAE;oBACvE,GAAG,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC,0CAA0C;oBAC5E,OAAO,GAAG,CAAC;gBACb,CAAC,EAAE,EAAyB,CAAC,CAC9B,CAAC,CAAC,CAAC,SAAS;aACd;SACF,CAAC,CAAC;QAEH,IAAI,CAAC,QAAQ,EAAE;YACb,OAAO;gBACL,IAAI,EAAE,YAAY;gBAClB,MAAM,EAAE,MAAgB;gBACxB,SAAS,EAAE,UAAU,CAAC,MAAM,GAAG,CAAC;gBAChC,WAAW,EAAE,UAAU;aACxB,CAAC;SACH;QACD,MAAM,WAAW,GAAG,MAAkC,CAAC;QACvD,WAAW,CAAC,IAAI,GAAG,OAAO,CAAC;QAC3B,WAAW,CAAC,SAAS,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;QAC9C,WAAW,CAAC,WAAW,GAAG,UAAU,CAAC;QACrC,OAAO,WAAW,CAAC;IACrB,CAAC;IAED;;;;;;;;;OASG;IACH,KAAK,CAAC,UAAU,CACd,MAAwB,EACxB,QAAoC;QAEpC,MAAM,YAAY,mCACb,MAAM,KACT,MAAM,EAAE,MAAM,CAAC,MAAM,IAAI,EAAE,EAC3B,uBAAuB,EAAE,CAAC,CAAC,QAAQ,GACpC,CAAC;QAEF,IAAI,MAAM,CAAC,QAAQ,EAAE;YACnB,MAAM,eAAe,GAAG,MAAM,IAAI,CAAC,gBAAgB,CACjD,MAAM,CAAC,QAAQ,EACf,MAAM,CAAC,aAAa,IAAI,MAAM,CAAC,YAAY,EAC3C;gBACE,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;gBAC/C,WAAW,EAAE,MAAM,CAAC,WAAW;gBAC/B,eAAe,EAAE,MAAM,CAAC,eAAe;gBACvC,qBAAqB,EAAE,MAAM,CAAC,qBAAqB;gBACnD,GAAG,EAAE,MAAM,CAAC,GAAG;gBACf,oBAAoB,EAAE,MAAM,CAAC,oBAAoB;aAClD,CACF,CAAC;YACF,IAAI,eAAe,CAAC,IAAI,KAAK,OAAO,EAAE;gBACpC,MAAM,WAAW,GAAG,eAA2C,CAAC;gBAEhE,YAAY,CAAC,MAAM,GAAG,WAAW,CAAC,MAAM,IAAI,EAAE,CAAC;gBAC/C,IAAI,OAAO,WAAW,CAAC,WAAW,KAAK,QAAQ;oBAC7C,YAAY,CAAC,WAAW,GAAG,WAAW,CAAC,WAAW,CAAC;gBACrD,IAAI,WAAW,CAAC,OAAO;oBAAE,YAAY,CAAC,OAAO,GAAG,WAAW,CAAC,OAAO,CAAC;gBACpE,IAAI,OAAO,WAAW,CAAC,YAAY,KAAK,SAAS;oBAC/C,YAAY,CAAC,YAAY,GAAG,WAAW,CAAC,YAAY,CAAC;gBACvD,IAAI,WAAW,CAAC,gBAAgB;oBAC9B,YAAY,CAAC,gBAAgB,GAAG,WAAW,CAAC,gBAAgB,CAAC;gBAC/D,IAAI,WAAW,CAAC,gBAAgB;oBAC9B,YAAY,CAAC,gBAAgB,GAAG,WAAW,CAAC,gBAAgB,CAAC;gBAC/D,IAAI,WAAW,CAAC,gBAAgB,EAAE;oBAChC,IAAI,CAAC,YAAY,CAAC,IAAI;wBAAE,YAAY,CAAC,IAAI,GAAG,EAAE,CAAC;oBAC/C,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,WAAW,CAAC,gBAAgB,CAAC,CAAC;iBACzD;gBACD,IAAI,WAAW,CAAC,SAAS,EAAE;oBACzB,YAAY,CAAC,WAAW,GAAG,WAAW,CAAC,WAAW,CAAC;iBACpD;aACF;iBAAM,IAAI,eAAe,CAAC,IAAI,KAAK,YAAY,EAAE;gBAChD,MAAM,eAAe,GAAG,eAAsC,CAAC;gBAC/D,YAAY,CAAC,MAAM,GAAG,eAAe,CAAC,MAAM,IAAI,EAAE,CAAC;gBACnD,IAAI,eAAe,CAAC,SAAS,EAAE;oBAC7B,YAAY,CAAC,WAAW,GAAG,eAAe,CAAC,WAAW,CAAC;iBACxD;aACF;SACF;aAAM;YACL,YAAY,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,EAAE,CAAC;SAC3C;QAED,+EAA+E;QAC/E,IAAI,CAAC,YAAY,CAAC,WAAW,IAAI,MAAM,CAAC,WAAW,EAAE;YACnD,YAAY,CAAC,WAAW,GAAG,MAAM,CAAC,WAAW,CAAC;SAC/C;QAED,IAAI,YAAY,CAAC,eAAe,IAAI,CAAC,YAAY,CAAC,OAAO,EAAE;YACzD,MAAM,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC,eAAe,CAAC,CAAC;YACzD,IAAI,UAAU;gBAAE,YAAY,CAAC,WAAW,GAAG,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC;SACvE;QAED,IAAI,aAAa,GACf,QAAQ;YACR,QAAQ,CAAC,WAAW,CAAC,cAAc,EAAE,CAAC,GAAqB,EAAE,EAAE;gBAC7D,MAAM,EAAE,SAAS,EAAE,WAAW,EAAE,GAAG,GAAG,CAAC;gBACvC,IAAI,SAAS,KAAK,IAAI,CAAC,EAAE;oBAAE,OAAO;gBAClC,QAAQ,CAAC,WAAW,CAAC,CAAC;YACxB,CAAC,CAAC,CAAC;QAEL,IAAI,CAAC,YAAY,CAAC,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,CAAC,CAAC;QAEhE,MAAM,OAAO,GAAG,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC,CAAC;QAClF,OAAO,OAAO;aACX,IAAI,CAAC,CAAC,gBAAgB,EAAE,EAAE;YACzB,aAAa,aAAb,aAAa,uBAAb,aAAa,CAAE,MAAM,GAAG;YACxB,aAAa,GAAG,IAAI,CAAC;YACrB,OAAO,gBAAgB,CAAC;QAC1B,CAAC,CAAC;aACD,KAAK,CAAC,CAAC,GAAQ,EAAE,EAAE;YAClB,aAAa,aAAb,aAAa,uBAAb,aAAa,CAAE,MAAM,GAAG;YACxB,aAAa,GAAG,IAAI,CAAC;YACrB,MAAM,GAAG,CAAC;QACZ,CAAC,CAAC,CAAC;IACP,CAAC;IAED,cAAc;QACZ,OAAO,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACzD,CAAC;IAED;;;;;OAKG;IACH,QAAQ,CACN,IAAY,EACZ,EACE,WAAW,EAAE,UAAU,MAGrB,EAAE;QAEN,OAAO,QAAQ,CAAC,QAAQ,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,UAAU,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,UAAU,CAAC,MAAgB;QACzB,OAAO,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAC7D,CAAC;IAED,SAAS,CACP,IAAY,EACZ,MAAwB;QAExB,OAAO,QAAQ,CAAC,SAAS,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,IAAI,EAAE,EAAE,CAAC,CAAC;IAChF,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,MAAM,CACV,KAAa,EACb,SAAmB,EACnB,MAAqB;QAErB,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAC,MAAM,CAAC;YACpC,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,KAAK;YACL,SAAS;YACT,MAAM,EAAE,MAAM,IAAI,EAAE;SACrB,CAAC,CAAC;QAEH,8DAA8D;QAC9D,OAAO,OAAO;aACX,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,iCACZ,MAAM,KACT,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,IACjC,CAAC;aACF,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;IACvC,CAAC;IAED,KAAK,CAAC,KAAK,CACT,EAAU,EACV,EAAU,EACV,EAAU,EACV,EAAU;QAEV,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QAC5E,MAAM,CAAC,SAAS,EAAE,SAAS,EAAE,YAAY,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,GACpE,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QACrB,OAAO;YACL,SAAS;YACT,SAAS;YACT,YAAY;YACZ,KAAK;YACL,KAAK;YACL,KAAK;YACL,KAAK;SACN,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,iBAAiB,CACrB,QAAkD;QAElD,IAAI,YAAY,GAA6C,EAAE,CAAC;QAChE,IAAI,QAAQ;YACV,YAAY,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAClC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC;gBACrC,MAAM,EAAE,CAAC,CAAC,MAAM;aACjB,CAAC,CAAC,CAAC;QACN,OAAO,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,YAAY,EAAE,CAAC,CAAC;IAC1E,CAAC;IAED,KAAK,CAAC,kBAAkB;QACtB,OAAO,QAAQ,CAAC,kBAAkB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAC7D,CAAC;IAED,KAAK,CAAC,qBAAqB;QAGzB,OAAO,QAAQ,CAAC,qBAAqB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAChE,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,cAAc,CAAC,EACnB,IAAI,EACJ,OAAO,EAAE,MAAM,GAIhB;QACC,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,QAAQ,CAAC,cAAc,CAAC;YAC7B,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,MAAM,EAAE;gBACN,IAAI;gBACJ,OAAO,EAAE,MAAM,aAAN,MAAM,cAAN,MAAM,GAAI,IAAI;aACxB;SACF,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,mBAAmB;QACvB,OAAO,MAAM,QAAQ,CAAC,mBAAmB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACpE,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,oBAAoB;QAIxB,OAAO,MAAM,QAAQ,CAAC,oBAAoB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACrE,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,iBAAiB;QACrB,OAAO,MAAM,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAClE,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,WAAW,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAsC;QAC7E,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,MAAM,QAAQ,CAAC,WAAW,CAAC;YAChC,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,MAAM,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE;SAClC,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,gBAAgB;QACpB,OAAO,MAAM,QAAQ,CAAC,gBAAgB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACjE,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,2BAA2B,CAC/B,OAAsB,EACtB,WAAmB;QAKnB,OAAO,MAAM,QAAQ,CAAC,2BAA2B,CAAC;YAChD,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,cAAc,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE;YACtD,WAAW;SACZ,CAAC,CAAC;IACL,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,6BAA6B,CACjC,WAAmB;QAEnB,OAAO,MAAM,QAAQ,CAAC,6BAA6B,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,WAAW,EAAE,CAAC,CAAC;IAC3F,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,iBAAiB,CAAC,MAAgB;QACtC,OAAO,MAAM,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAC1E,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,cAAc;QAClB,OAAO,MAAM,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAC/D,CAAC;IAED,KAAK,CAAC,OAAO;QACX,OAAO,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACzD,CAAC;CACF;AAED,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,OAAgB;IACpD,OAAO,QAAQ,CAAC,eAAe,CAAC,EAAE,OAAO,EAAE,CAAC,CAAC;AAC/C,CAAC;AAED,MAAM,UAAU,oBAAoB,CAClC,QAA+C;IAE/C,YAAY,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IAC5B,OAAO;QACL,MAAM,EAAE,GAAG,EAAE;YACX,YAAY,CAAC,MAAM,CAAC,YAAY,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,CAAC;QACzD,CAAC;KACF,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,KAAa;IACjD,OAAO,QAAQ,CAAC,eAAe,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;AAC7C,CAAC;AAED,IAAI,gBAAgB,GAAG,CAAC,CAAC;AACzB,MAAM,eAAe,GAAG,GAAG,EAAE,CAC3B,OAAO,CAAC,GAAG,CAAC,QAAQ,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,MAAM,CAAC,CAAC;AAE3E,MAAM,aAAa,GAAG;IACpB,eAAe;IACf,uBAAuB;IACvB,2BAA2B;IAC3B,uBAAuB;IACvB,uBAAuB;CACxB,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAC,KAAa;IACpD,IAAI,IAAI,GAAG,KAAK,CAAC;IACjB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;QAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACrD,OAAO,QAAQ,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,aAAa,EAAE,CAAC,CAAC;AAC3D,CAAC;AAED,MAAM,WAAW,GAAG;IAClB,iCAAiC;IACjC,IAAI,EAAE,CAAC;IACP,IAAI,EAAE,CAAC;IACP,GAAG,EAAE,CAAC;IACN,IAAI,EAAE,CAAC;IACP,IAAI,EAAE,CAAC;CACR,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,EAOgB,EAChB,UAAuC;QARvC,EACE,KAAK,EACL,cAAc,EAAE,YAAY,EAC5B,YAAY,EAAE,WAAW,EACzB,IAAI,EACJ,SAAS,EAAE,QAAQ,OAEL,EADX,IAAI,cANT,gEAOC,CADQ;IAIT,IAAI,IAAI,GAAG,KAAK,CAAC;IACjB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;QAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAErD,IAAI,QAAQ,GAAG,IAAI,CAAC;IACpB,IAAI,QAAQ,aAAR,QAAQ,uBAAR,QAAQ,CAAE,UAAU,CAAC,SAAS;QAAG,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAElE,IAAI,YAAY,GAA6C,EAAE,CAAC;IAChE,IAAI,QAAQ;QACV,YAAY,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAClC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC;YACrC,MAAM,EAAE,CAAC,CAAC,MAAM;SACjB,CAAC,CAAC,CAAC;IAEN,MAAM,SAAS,GAAG,gBAAgB,GAAG,eAAe,EAAE,CAAC;IACvD,gBAAgB,IAAI,CAAC,CAAC;IAEtB,IAAI,sBAAsB,GAAQ,IAAI,CAAC;IACvC,IAAI,UAAU,EAAE;QACd,sBAAsB,GAAG,QAAQ,CAAC,WAAW,CAC3C,8BAA8B,EAC9B,CAAC,GAA4C,EAAE,EAAE;YAC/C,IAAI,GAAG,CAAC,SAAS,KAAK,SAAS;gBAAE,OAAO;YACxC,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC3B,CAAC,CACF,CAAC;KACH;IAED,MAAM,QAAQ,GAAG,WAAW,CAAC,WAAuC,CAAC,CAAC;IAEtE,IAAI,IAAI,CAAC,YAAY,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,YAAY,CAAC,EAAE;QACrE,OAAO,CAAC,IAAI,CAAC,+CAA+C,IAAI,CAAC,YAAY,uBAAuB,CAAC,CAAC;QACtG,OAAO,IAAI,CAAC,YAAY,CAAC;KAC1B;IACD,IAAI,IAAI,CAAC,YAAY,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,YAAY,CAAC,EAAE;QACrE,OAAO,CAAC,IAAI,CAAC,+CAA+C,IAAI,CAAC,YAAY,uBAAuB,CAAC,CAAC;QACtG,OAAO,IAAI,CAAC,YAAY,CAAC;KAC1B;IAED,MAAM,EACJ,GAAG,EACH,WAAW,EACX,KAAK,EAAE,YAAY,EACnB,UAAU,GACX,GAAG,MAAM,QAAQ,CAAC,WAAW,CAAC;QAC7B,SAAS;QACT,MAAM,kBACJ,KAAK,EAAE,IAAI,EACX,cAAc,EAAE,CAAC,CAAC,YAAY,EAC9B,qBAAqB,EAAE,CAAC,CAAC,UAAU,EACnC,YAAY,EAAE,QAAQ,EACtB,IAAI,EAAE,QAAQ,EACd,SAAS,EAAE,YAAY,IACpB,IAAI,CACR;KACF,CAAC,CAAC,KAAK,CAAC,CAAC,GAAQ,EAAE,EAAE;QACpB,sBAAsB,aAAtB,sBAAsB,uBAAtB,sBAAsB,CAAE,MAAM,GAAG;QACjC,MAAM,GAAG,CAAC;IACZ,CAAC,CAAC,CAAC;IACH,sBAAsB,aAAtB,sBAAsB,uBAAtB,sBAAsB,CAAE,MAAM,GAAG;IACjC,OAAO,IAAI,YAAY,CAAC;QACtB,SAAS;QACT,GAAG;QACH,WAAW;QACX,KAAK,EAAE,YAAY;QACnB,UAAU;KACX,CAAC,CAAC;AACL,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,eAAe;IACnC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,CAAC;AACvC,CAAC;AAED,MAAM,CAAC,MAAM,SAAS,GAAG;IACvB,MAAM,EAAE,OAAO;IACf,MAAM,EAAE,qBAAqB;CAC9B,CAAC;AAEF,yCAAyC;AACzC,OAAO,EAAE,QAAQ,EAAE,CAAC","sourcesContent":["import { registerPlugin } from '@capacitor/core';\nimport type {\n  NativeContextParams,\n  NativeLlamaContext,\n  NativeCompletionParams,\n  NativeCompletionTokenProb,\n  NativeCompletionResult,\n  NativeTokenizeResult,\n  NativeEmbeddingResult,\n  NativeSessionLoadResult,\n  NativeEmbeddingParams,\n  NativeRerankParams,\n  NativeRerankResult,\n  NativeCompletionTokenProbItem,\n  NativeCompletionResultTimings,\n  JinjaFormattedChatResult,\n  FormattedChatResult,\n  NativeImageProcessingResult,\n  NativeLlamaChatMessage,\n  LlamaCppMessagePart,\n  LlamaCppOAICompatibleMessage,\n  ContextParams,\n  EmbeddingParams,\n  RerankParams,\n  RerankResult,\n  CompletionResponseFormat,\n  CompletionParams,\n  BenchResult,\n  LlamaCppPlugin,\n} from './definitions';\n\n// Constants\nexport const LLAMACPP_MTMD_DEFAULT_MEDIA_MARKER = '<__media__>';\n\n// Event names\nconst EVENT_ON_INIT_CONTEXT_PROGRESS = '@LlamaCpp_onInitContextProgress';\nconst EVENT_ON_TOKEN = '@LlamaCpp_onToken';\nconst EVENT_ON_NATIVE_LOG = '@LlamaCpp_onNativeLog';\n\n// Register the plugin\nconst LlamaCpp = registerPlugin<LlamaCppPlugin>('LlamaCpp');\n\n// Log listeners management\nconst logListeners: Array<(level: string, text: string) => void> = [];\n\n// Set up native log listener\nLlamaCpp.addListener(EVENT_ON_NATIVE_LOG, (evt: { level: string; text: string }) => {\n  logListeners.forEach((listener) => listener(evt.level, evt.text));\n});\n\n// Trigger unset to use default log callback\nLlamaCpp?.toggleNativeLog?.({ enabled: false })?.catch?.(() => {});\n\n// High-level types for the plugin interface\nexport type RNLlamaMessagePart = LlamaCppMessagePart;\nexport type RNLlamaOAICompatibleMessage = LlamaCppOAICompatibleMessage;\n\n// Re-export all types from definitions\nexport type {\n  NativeContextParams,\n  NativeLlamaContext,\n  NativeCompletionParams,\n  NativeCompletionTokenProb,\n  NativeCompletionResult,\n  NativeTokenizeResult,\n  NativeEmbeddingResult,\n  NativeSessionLoadResult,\n  NativeEmbeddingParams,\n  NativeRerankParams,\n  NativeRerankResult,\n  NativeCompletionTokenProbItem,\n  NativeCompletionResultTimings,\n  FormattedChatResult,\n  JinjaFormattedChatResult,\n  NativeImageProcessingResult,\n  ContextParams,\n  EmbeddingParams,\n  RerankParams,\n  RerankResult,\n  CompletionResponseFormat,\n  CompletionParams,\n  BenchResult,\n};\n\nexport const RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER = LLAMACPP_MTMD_DEFAULT_MEDIA_MARKER;\n\nexport type ToolCall = {\n  type: 'function';\n  id?: string;\n  function: {\n    name: string;\n    arguments: string; // JSON string\n  };\n};\n\nexport type TokenData = {\n  token: string;\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\n  // Parsed content from accumulated text\n  content?: string;\n  reasoning_content?: string;\n  tool_calls?: Array<ToolCall>;\n  accumulated_text?: string;\n};\n\ntype TokenNativeEvent = {\n  contextId: number;\n  tokenResult: TokenData;\n};\n\nconst validCacheTypes = [\n  'f16',\n  'f32',\n  'bf16',\n  'q8_0',\n  'q4_0',\n  'q4_1',\n  'iq4_nl',\n  'q5_0',\n  'q5_1',\n];\n\nconst getJsonSchema = (responseFormat?: CompletionResponseFormat) => {\n  if (responseFormat?.type === 'json_schema') {\n    return responseFormat.json_schema?.schema;\n  }\n  if (responseFormat?.type === 'json_object') {\n    return responseFormat.schema || {};\n  }\n  return null;\n};\n\nexport class LlamaContext {\n  id: number;\n  gpu: boolean = false;\n  reasonNoGPU: string = '';\n  model: NativeLlamaContext['model'];\n\n  constructor({ contextId, gpu, reasonNoGPU, model }: NativeLlamaContext) {\n    this.id = contextId;\n    this.gpu = gpu;\n    this.reasonNoGPU = reasonNoGPU;\n    this.model = model;\n  }\n\n  /**\n   * Load cached prompt & completion state from a file.\n   */\n  async loadSession(filepath: string): Promise<NativeSessionLoadResult> {\n    let path = filepath;\n    if (path.startsWith('file://')) path = path.slice(7);\n    return LlamaCpp.loadSession({ contextId: this.id, filepath: path });\n  }\n\n  /**\n   * Save current cached prompt & completion state to a file.\n   */\n  async saveSession(\n    filepath: string,\n    options?: { tokenSize: number },\n  ): Promise<number> {\n    return LlamaCpp.saveSession({ \n      contextId: this.id, \n      filepath, \n      size: options?.tokenSize || -1 \n    });\n  }\n\n  isLlamaChatSupported(): boolean {\n    return !!this.model.chatTemplates.llamaChat;\n  }\n\n  isJinjaSupported(): boolean {\n    const { minja } = this.model.chatTemplates;\n    return !!minja?.toolUse || !!minja?.default;\n  }\n\n  async getFormattedChat(\n    messages: RNLlamaOAICompatibleMessage[],\n    template?: string | null,\n    params?: {\n      jinja?: boolean;\n      response_format?: CompletionResponseFormat;\n      tools?: object;\n      parallel_tool_calls?: object;\n      tool_choice?: string;\n      enable_thinking?: boolean;\n      add_generation_prompt?: boolean;\n      now?: string | number;\n      chat_template_kwargs?: Record<string, string>;\n    },\n  ): Promise<FormattedChatResult | JinjaFormattedChatResult> {\n    const mediaPaths: string[] = [];\n    const chat = messages.map((msg) => {\n      if (Array.isArray(msg.content)) {\n        const content = msg.content.map((part) => {\n          // Handle multimodal content\n          if (part.type === 'image_url') {\n            let path = part.image_url?.url || '';\n            if (path?.startsWith('file://')) path = path.slice(7);\n            mediaPaths.push(path);\n            return {\n              type: 'text',\n              text: RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER,\n            };\n          } else if (part.type === 'input_audio') {\n            const { input_audio: audio } = part;\n            if (!audio) throw new Error('input_audio is required');\n\n            const { format } = audio;\n            if (format != 'wav' && format != 'mp3') {\n              throw new Error(`Unsupported audio format: ${format}`);\n            }\n            if (audio.url) {\n              const path = audio.url.replace(/file:\\/\\//, '');\n              mediaPaths.push(path);\n            } else if (audio.data) {\n              mediaPaths.push(audio.data);\n            }\n            return {\n              type: 'text',\n              text: RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER,\n            };\n          }\n          return part;\n        });\n\n        return {\n          ...msg,\n          content,\n        };\n      }\n      return msg;\n    }) as NativeLlamaChatMessage[];\n\n    const useJinja = this.isJinjaSupported() && params?.jinja;\n    let tmpl;\n    if (template) tmpl = template; // Force replace if provided\n    const jsonSchema = getJsonSchema(params?.response_format);\n\n    const result = await LlamaCpp.getFormattedChat({\n      contextId: this.id,\n      messages: JSON.stringify(chat),\n      chatTemplate: tmpl,\n      params: {\n        jinja: useJinja,\n        json_schema: jsonSchema ? JSON.stringify(jsonSchema) : undefined,\n        tools: params?.tools ? JSON.stringify(params.tools) : undefined,\n        parallel_tool_calls: params?.parallel_tool_calls\n          ? JSON.stringify(params.parallel_tool_calls)\n          : undefined,\n        tool_choice: params?.tool_choice,\n        enable_thinking: params?.enable_thinking ?? true,\n        add_generation_prompt: params?.add_generation_prompt,\n        now: typeof params?.now === 'number' ? params.now.toString() : params?.now,\n        chat_template_kwargs: params?.chat_template_kwargs ? JSON.stringify(\n          Object.entries(params.chat_template_kwargs).reduce((acc, [key, value]) => {\n            acc[key] = JSON.stringify(value); // Each value is a stringified JSON object\n            return acc;\n          }, {} as Record<string, any>)\n        ) : undefined,\n      },\n    });\n\n    if (!useJinja) {\n      return {\n        type: 'llama-chat',\n        prompt: result as string,\n        has_media: mediaPaths.length > 0,\n        media_paths: mediaPaths,\n      };\n    }\n    const jinjaResult = result as JinjaFormattedChatResult;\n    jinjaResult.type = 'jinja';\n    jinjaResult.has_media = mediaPaths.length > 0;\n    jinjaResult.media_paths = mediaPaths;\n    return jinjaResult;\n  }\n\n  /**\n   * Generate a completion based on the provided parameters\n   * @param params Completion parameters including prompt or messages\n   * @param callback Optional callback for token-by-token streaming\n   * @returns Promise resolving to the completion result\n   *\n   * Note: For multimodal support, you can include an media_paths parameter.\n   * This will process the images and add them to the context before generating text.\n   * Multimodal support must be enabled via initMultimodal() first.\n   */\n  async completion(\n    params: CompletionParams,\n    callback?: (data: TokenData) => void,\n  ): Promise<NativeCompletionResult> {\n    const nativeParams = {\n      ...params,\n      prompt: params.prompt || '',\n      emit_partial_completion: !!callback,\n    };\n\n    if (params.messages) {\n      const formattedResult = await this.getFormattedChat(\n        params.messages,\n        params.chat_template || params.chatTemplate,\n        {\n          jinja: params.jinja,\n          tools: params.tools,\n          parallel_tool_calls: params.parallel_tool_calls,\n          tool_choice: params.tool_choice,\n          enable_thinking: params.enable_thinking,\n          add_generation_prompt: params.add_generation_prompt,\n          now: params.now,\n          chat_template_kwargs: params.chat_template_kwargs,\n        },\n      );\n      if (formattedResult.type === 'jinja') {\n        const jinjaResult = formattedResult as JinjaFormattedChatResult;\n\n        nativeParams.prompt = jinjaResult.prompt || '';\n        if (typeof jinjaResult.chat_format === 'number')\n          nativeParams.chat_format = jinjaResult.chat_format;\n        if (jinjaResult.grammar) nativeParams.grammar = jinjaResult.grammar;\n        if (typeof jinjaResult.grammar_lazy === 'boolean')\n          nativeParams.grammar_lazy = jinjaResult.grammar_lazy;\n        if (jinjaResult.grammar_triggers)\n          nativeParams.grammar_triggers = jinjaResult.grammar_triggers;\n        if (jinjaResult.preserved_tokens)\n          nativeParams.preserved_tokens = jinjaResult.preserved_tokens;\n        if (jinjaResult.additional_stops) {\n          if (!nativeParams.stop) nativeParams.stop = [];\n          nativeParams.stop.push(...jinjaResult.additional_stops);\n        }\n        if (jinjaResult.has_media) {\n          nativeParams.media_paths = jinjaResult.media_paths;\n        }\n      } else if (formattedResult.type === 'llama-chat') {\n        const llamaChatResult = formattedResult as FormattedChatResult;\n        nativeParams.prompt = llamaChatResult.prompt || '';\n        if (llamaChatResult.has_media) {\n          nativeParams.media_paths = llamaChatResult.media_paths;\n        }\n      }\n    } else {\n      nativeParams.prompt = params.prompt || '';\n    }\n\n    // If media_paths were explicitly provided or extracted from messages, use them\n    if (!nativeParams.media_paths && params.media_paths) {\n      nativeParams.media_paths = params.media_paths;\n    }\n\n    if (nativeParams.response_format && !nativeParams.grammar) {\n      const jsonSchema = getJsonSchema(params.response_format);\n      if (jsonSchema) nativeParams.json_schema = JSON.stringify(jsonSchema);\n    }\n\n    let tokenListener: any =\n      callback &&\n      LlamaCpp.addListener(EVENT_ON_TOKEN, (evt: TokenNativeEvent) => {\n        const { contextId, tokenResult } = evt;\n        if (contextId !== this.id) return;\n        callback(tokenResult);\n      });\n\n    if (!nativeParams.prompt) throw new Error('Prompt is required');\n\n    const promise = LlamaCpp.completion({ contextId: this.id, params: nativeParams });\n    return promise\n      .then((completionResult) => {\n        tokenListener?.remove();\n        tokenListener = null;\n        return completionResult;\n      })\n      .catch((err: any) => {\n        tokenListener?.remove();\n        tokenListener = null;\n        throw err;\n      });\n  }\n\n  stopCompletion(): Promise<void> {\n    return LlamaCpp.stopCompletion({ contextId: this.id });\n  }\n\n  /**\n   * Tokenize text or text with images\n   * @param text Text to tokenize\n   * @param params.media_paths Array of image paths to tokenize (if multimodal is enabled)\n   * @returns Promise resolving to the tokenize result\n   */\n  tokenize(\n    text: string,\n    {\n      media_paths: mediaPaths,\n    }: {\n      media_paths?: string[];\n    } = {},\n  ): Promise<NativeTokenizeResult> {\n    return LlamaCpp.tokenize({ contextId: this.id, text, imagePaths: mediaPaths });\n  }\n\n  detokenize(tokens: number[]): Promise<string> {\n    return LlamaCpp.detokenize({ contextId: this.id, tokens });\n  }\n\n  embedding(\n    text: string,\n    params?: EmbeddingParams,\n  ): Promise<NativeEmbeddingResult> {\n    return LlamaCpp.embedding({ contextId: this.id, text, params: params || {} });\n  }\n\n  /**\n   * Rerank documents based on relevance to a query\n   * @param query The query text to rank documents against\n   * @param documents Array of document texts to rank\n   * @param params Optional reranking parameters\n   * @returns Promise resolving to an array of ranking results with scores and indices\n   */\n  async rerank(\n    query: string,\n    documents: string[],\n    params?: RerankParams,\n  ): Promise<RerankResult[]> {\n    const results = await LlamaCpp.rerank({ \n      contextId: this.id, \n      query, \n      documents, \n      params: params || {} \n    });\n\n    // Sort by score descending and add document text if requested\n    return results\n      .map((result) => ({\n        ...result,\n        document: documents[result.index],\n      }))\n      .sort((a, b) => b.score - a.score);\n  }\n\n  async bench(\n    pp: number,\n    tg: number,\n    pl: number,\n    nr: number,\n  ): Promise<BenchResult> {\n    const result = await LlamaCpp.bench({ contextId: this.id, pp, tg, pl, nr });\n    const [modelDesc, modelSize, modelNParams, ppAvg, ppStd, tgAvg, tgStd] =\n      JSON.parse(result);\n    return {\n      modelDesc,\n      modelSize,\n      modelNParams,\n      ppAvg,\n      ppStd,\n      tgAvg,\n      tgStd,\n    };\n  }\n\n  async applyLoraAdapters(\n    loraList: Array<{ path: string; scaled?: number }>,\n  ): Promise<void> {\n    let loraAdapters: Array<{ path: string; scaled?: number }> = [];\n    if (loraList)\n      loraAdapters = loraList.map((l) => ({\n        path: l.path.replace(/file:\\/\\//, ''),\n        scaled: l.scaled,\n      }));\n    return LlamaCpp.applyLoraAdapters({ contextId: this.id, loraAdapters });\n  }\n\n  async removeLoraAdapters(): Promise<void> {\n    return LlamaCpp.removeLoraAdapters({ contextId: this.id });\n  }\n\n  async getLoadedLoraAdapters(): Promise<\n    Array<{ path: string; scaled?: number }>\n  > {\n    return LlamaCpp.getLoadedLoraAdapters({ contextId: this.id });\n  }\n\n  /**\n   * Initialize multimodal support with a mmproj file\n   * @param params Parameters for multimodal support\n   * @param params.path Path to the multimodal projector file\n   * @param params.use_gpu Whether to use GPU\n   * @returns Promise resolving to true if initialization was successful\n   */\n  async initMultimodal({\n    path,\n    use_gpu: useGpu,\n  }: {\n    path: string;\n    use_gpu?: boolean;\n  }): Promise<boolean> {\n    if (path.startsWith('file://')) path = path.slice(7);\n    return LlamaCpp.initMultimodal({\n      contextId: this.id,\n      params: {\n        path,\n        use_gpu: useGpu ?? true,\n      },\n    });\n  }\n\n  /**\n   * Check if multimodal support is enabled\n   * @returns Promise resolving to true if multimodal is enabled\n   */\n  async isMultimodalEnabled(): Promise<boolean> {\n    return await LlamaCpp.isMultimodalEnabled({ contextId: this.id });\n  }\n\n  /**\n   * Check multimodal support\n   * @returns Promise resolving to an object with vision and audio support\n   */\n  async getMultimodalSupport(): Promise<{\n    vision: boolean;\n    audio: boolean;\n  }> {\n    return await LlamaCpp.getMultimodalSupport({ contextId: this.id });\n  }\n\n  /**\n   * Release multimodal support\n   * @returns Promise resolving to void\n   */\n  async releaseMultimodal(): Promise<void> {\n    return await LlamaCpp.releaseMultimodal({ contextId: this.id });\n  }\n\n  /**\n   * Initialize TTS support with a vocoder model\n   * @param params Parameters for TTS support\n   * @param params.path Path to the vocoder model\n   * @param params.n_batch Batch size for the vocoder model\n   * @returns Promise resolving to true if initialization was successful\n   */\n  async initVocoder({ path, n_batch: nBatch }: { path: string; n_batch?: number }): Promise<boolean> {\n    if (path.startsWith('file://')) path = path.slice(7);\n    return await LlamaCpp.initVocoder({ \n      contextId: this.id, \n      params: { path, n_batch: nBatch } \n    });\n  }\n\n  /**\n   * Check if TTS support is enabled\n   * @returns Promise resolving to true if TTS is enabled\n   */\n  async isVocoderEnabled(): Promise<boolean> {\n    return await LlamaCpp.isVocoderEnabled({ contextId: this.id });\n  }\n\n  /**\n   * Get a formatted audio completion prompt\n   * @param speakerJsonStr JSON string representing the speaker\n   * @param textToSpeak Text to speak\n   * @returns Promise resolving to the formatted audio completion result with prompt and grammar\n   */\n  async getFormattedAudioCompletion(\n    speaker: object | null,\n    textToSpeak: string,\n  ): Promise<{\n    prompt: string;\n    grammar?: string;\n  }> {\n    return await LlamaCpp.getFormattedAudioCompletion({\n      contextId: this.id,\n      speakerJsonStr: speaker ? JSON.stringify(speaker) : '',\n      textToSpeak,\n    });\n  }\n\n  /**\n   * Get guide tokens for audio completion\n   * @param textToSpeak Text to speak\n   * @returns Promise resolving to the guide tokens\n   */\n  async getAudioCompletionGuideTokens(\n    textToSpeak: string,\n  ): Promise<Array<number>> {\n    return await LlamaCpp.getAudioCompletionGuideTokens({ contextId: this.id, textToSpeak });\n  }\n\n  /**\n   * Decode audio tokens\n   * @param tokens Array of audio tokens\n   * @returns Promise resolving to the decoded audio tokens\n   */\n  async decodeAudioTokens(tokens: number[]): Promise<Array<number>> {\n    return await LlamaCpp.decodeAudioTokens({ contextId: this.id, tokens });\n  }\n\n  /**\n   * Release TTS support\n   * @returns Promise resolving to void\n   */\n  async releaseVocoder(): Promise<void> {\n    return await LlamaCpp.releaseVocoder({ contextId: this.id });\n  }\n\n  async release(): Promise<void> {\n    return LlamaCpp.releaseContext({ contextId: this.id });\n  }\n}\n\nexport async function toggleNativeLog(enabled: boolean): Promise<void> {\n  return LlamaCpp.toggleNativeLog({ enabled });\n}\n\nexport function addNativeLogListener(\n  listener: (level: string, text: string) => void,\n): { remove: () => void } {\n  logListeners.push(listener);\n  return {\n    remove: () => {\n      logListeners.splice(logListeners.indexOf(listener), 1);\n    },\n  };\n}\n\nexport async function setContextLimit(limit: number): Promise<void> {\n  return LlamaCpp.setContextLimit({ limit });\n}\n\nlet contextIdCounter = 0;\nconst contextIdRandom = () =>\n  process.env.NODE_ENV === 'test' ? 0 : Math.floor(Math.random() * 100000);\n\nconst modelInfoSkip = [\n  // Large fields\n  'tokenizer.ggml.tokens',\n  'tokenizer.ggml.token_type',\n  'tokenizer.ggml.merges',\n  'tokenizer.ggml.scores',\n];\n\nexport async function loadLlamaModelInfo(model: string): Promise<Object> {\n  let path = model;\n  if (path.startsWith('file://')) path = path.slice(7);\n  return LlamaCpp.modelInfo({ path, skip: modelInfoSkip });\n}\n\nconst poolTypeMap = {\n  // -1 is unspecified as undefined\n  none: 0,\n  mean: 1,\n  cls: 2,\n  last: 3,\n  rank: 4,\n};\n\nexport async function initLlama(\n  {\n    model,\n    is_model_asset: isModelAsset,\n    pooling_type: poolingType,\n    lora,\n    lora_list: loraList,\n    ...rest\n  }: ContextParams,\n  onProgress?: (progress: number) => void,\n): Promise<LlamaContext> {\n  let path = model;\n  if (path.startsWith('file://')) path = path.slice(7);\n\n  let loraPath = lora;\n  if (loraPath?.startsWith('file://')) loraPath = loraPath.slice(7);\n\n  let loraAdapters: Array<{ path: string; scaled?: number }> = [];\n  if (loraList)\n    loraAdapters = loraList.map((l) => ({\n      path: l.path.replace(/file:\\/\\//, ''),\n      scaled: l.scaled,\n    }));\n\n  const contextId = contextIdCounter + contextIdRandom();\n  contextIdCounter += 1;\n\n  let removeProgressListener: any = null;\n  if (onProgress) {\n    removeProgressListener = LlamaCpp.addListener(\n      EVENT_ON_INIT_CONTEXT_PROGRESS,\n      (evt: { contextId: number; progress: number }) => {\n        if (evt.contextId !== contextId) return;\n        onProgress(evt.progress);\n      },\n    );\n  }\n\n  const poolType = poolTypeMap[poolingType as keyof typeof poolTypeMap];\n\n  if (rest.cache_type_k && !validCacheTypes.includes(rest.cache_type_k)) {\n    console.warn(`[LlamaCpp] initLlama: Invalid cache K type: ${rest.cache_type_k}, falling back to f16`);\n    delete rest.cache_type_k;\n  }\n  if (rest.cache_type_v && !validCacheTypes.includes(rest.cache_type_v)) {\n    console.warn(`[LlamaCpp] initLlama: Invalid cache V type: ${rest.cache_type_v}, falling back to f16`);\n    delete rest.cache_type_v;\n  }\n\n  const {\n    gpu,\n    reasonNoGPU,\n    model: modelDetails,\n    androidLib,\n  } = await LlamaCpp.initContext({\n    contextId,\n    params: {\n      model: path,\n      is_model_asset: !!isModelAsset,\n      use_progress_callback: !!onProgress,\n      pooling_type: poolType,\n      lora: loraPath,\n      lora_list: loraAdapters,\n      ...rest,\n    },\n  }).catch((err: any) => {\n    removeProgressListener?.remove();\n    throw err;\n  });\n  removeProgressListener?.remove();\n  return new LlamaContext({\n    contextId,\n    gpu,\n    reasonNoGPU,\n    model: modelDetails,\n    androidLib,\n  });\n}\n\nexport async function releaseAllLlama(): Promise<void> {\n  return LlamaCpp.releaseAllContexts();\n}\n\nexport const BuildInfo = {\n  number: '1.0.0',\n  commit: 'capacitor-llama-cpp',\n};\n\n// Re-export the plugin for direct access\nexport { LlamaCpp };\n"]}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/index.ts"],"names":[],"mappings":";;AAAA,OAAO,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AA+BjD,YAAY;AACZ,MAAM,CAAC,MAAM,kCAAkC,GAAG,aAAa,CAAC;AAEhE,cAAc;AACd,MAAM,8BAA8B,GAAG,iCAAiC,CAAC;AACzE,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAC3C,MAAM,mBAAmB,GAAG,uBAAuB,CAAC;AAEpD,sBAAsB;AACtB,MAAM,QAAQ,GAAG,cAAc,CAAiB,UAAU,CAAC,CAAC;AAE5D,2BAA2B;AAC3B,MAAM,YAAY,GAAiD,EAAE,CAAC;AAEtE,6BAA6B;AAC7B,QAAQ,CAAC,WAAW,CAAC,mBAAmB,EAAE,CAAC,GAAoC,EAAE,EAAE;IACjF,YAAY,CAAC,OAAO,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;AACpE,CAAC,CAAC,CAAC;AAEH,4CAA4C;AAC5C,kBAAA,QAAQ,aAAR,QAAQ,uBAAR,QAAQ,CAAE,eAAe,+CAAzB,QAAQ,EAAoB,EAAE,OAAO,EAAE,KAAK,EAAE,2CAAG,KAAK,mDAAG,GAAG,EAAE,GAAE,CAAC,EAAE;AAiCnE,MAAM,CAAC,MAAM,iCAAiC,GAAG,kCAAkC,CAAC;AA0BpF,MAAM,eAAe,GAAG;IACtB,KAAK;IACL,KAAK;IACL,MAAM;IACN,MAAM;IACN,MAAM;IACN,MAAM;IACN,QAAQ;IACR,MAAM;IACN,MAAM;CACP,CAAC;AAEF,MAAM,aAAa,GAAG,CAAC,cAAyC,EAAE,EAAE;;IAClE,IAAI,CAAA,cAAc,aAAd,cAAc,uBAAd,cAAc,CAAE,IAAI,MAAK,aAAa,EAAE;QAC1C,aAAO,cAAc,CAAC,WAAW,0CAAE,MAAM,CAAC;KAC3C;IACD,IAAI,CAAA,cAAc,aAAd,cAAc,uBAAd,cAAc,CAAE,IAAI,MAAK,aAAa,EAAE;QAC1C,OAAO,cAAc,CAAC,MAAM,IAAI,EAAE,CAAC;KACpC;IACD,OAAO,IAAI,CAAC;AACd,CAAC,CAAC;AAEF,0DAA0D;AAC1D,MAAM,mBAAmB,GAAG,KAAK,EAAE,MAAc,EAAmB,EAAE;IACpE,kEAAkE;IAClE,+CAA+C;IAC/C,IAAI;QACF,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,0BAA0B,CAAC,EAAE,MAAM,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAC7F,OAAO,MAAM,CAAC;KACf;IAAC,OAAO,KAAK,EAAE;QACd,OAAO,CAAC,IAAI,CAAC,wDAAwD,EAAE,KAAK,CAAC,CAAC;QAC9E,sCAAsC;QACtC,OAAO;;;;;;;kBAOO,CAAC;KAChB;AACH,CAAC,CAAC;AAEF,MAAM,OAAO,YAAY;IAMvB,YAAY,EAAE,SAAS,EAAE,GAAG,EAAE,WAAW,EAAE,KAAK,EAAsB;QAJtE,QAAG,GAAY,KAAK,CAAC;QACrB,gBAAW,GAAW,EAAE,CAAC;QAIvB,IAAI,CAAC,EAAE,GAAG,SAAS,CAAC;QACpB,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;IACrB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CAAC,QAAgB;QAChC,IAAI,IAAI,GAAG,QAAQ,CAAC;QACpB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,QAAQ,CAAC,WAAW,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC;IACtE,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CACf,QAAgB,EAChB,OAA+B;QAE/B,OAAO,QAAQ,CAAC,WAAW,CAAC;YAC1B,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,QAAQ;YACR,IAAI,EAAE,CAAA,OAAO,aAAP,OAAO,uBAAP,OAAO,CAAE,SAAS,KAAI,CAAC,CAAC;SAC/B,CAAC,CAAC;IACL,CAAC;IAED,oBAAoB;QAClB,OAAO,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC,SAAS,CAAC;IAC9C,CAAC;IAED,gBAAgB;QACd,MAAM,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC;QAC3C,OAAO,CAAC,EAAC,KAAK,aAAL,KAAK,uBAAL,KAAK,CAAE,OAAO,CAAA,IAAI,CAAC,EAAC,KAAK,aAAL,KAAK,uBAAL,KAAK,CAAE,OAAO,CAAA,CAAC;IAC9C,CAAC;IAED,KAAK,CAAC,gBAAgB,CACpB,QAAuC,EACvC,QAAwB,EACxB,MAUC;;QAED,MAAM,UAAU,GAAa,EAAE,CAAC;QAChC,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE;YAChC,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE;gBAC9B,MAAM,OAAO,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE;;oBACvC,4BAA4B;oBAC5B,IAAI,IAAI,CAAC,IAAI,KAAK,WAAW,EAAE;wBAC7B,IAAI,IAAI,GAAG,OAAA,IAAI,CAAC,SAAS,0CAAE,GAAG,KAAI,EAAE,CAAC;wBACrC,IAAI,IAAI,aAAJ,IAAI,uBAAJ,IAAI,CAAE,UAAU,CAAC,SAAS;4BAAG,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;wBACtD,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;wBACtB,OAAO;4BACL,IAAI,EAAE,MAAM;4BACZ,IAAI,EAAE,iCAAiC;yBACxC,CAAC;qBACH;yBAAM,IAAI,IAAI,CAAC,IAAI,KAAK,aAAa,EAAE;wBACtC,MAAM,EAAE,WAAW,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC;wBACpC,IAAI,CAAC,KAAK;4BAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;wBAEvD,MAAM,EAAE,MAAM,EAAE,GAAG,KAAK,CAAC;wBACzB,IAAI,MAAM,IAAI,KAAK,IAAI,MAAM,IAAI,KAAK,EAAE;4BACtC,MAAM,IAAI,KAAK,CAAC,6BAA6B,MAAM,EAAE,CAAC,CAAC;yBACxD;wBACD,IAAI,KAAK,CAAC,GAAG,EAAE;4BACb,MAAM,IAAI,GAAG,KAAK,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;4BAChD,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;yBACvB;6BAAM,IAAI,KAAK,CAAC,IAAI,EAAE;4BACrB,UAAU,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;yBAC7B;wBACD,OAAO;4BACL,IAAI,EAAE,MAAM;4BACZ,IAAI,EAAE,iCAAiC;yBACxC,CAAC;qBACH;oBACD,OAAO,IAAI,CAAC;gBACd,CAAC,CAAC,CAAC;gBAEH,uCACK,GAAG,KACN,OAAO,IACP;aACH;YACD,OAAO,GAAG,CAAC;QACb,CAAC,CAA6B,CAAC;QAE/B,MAAM,QAAQ,GAAG,IAAI,CAAC,gBAAgB,EAAE,KAAI,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,KAAK,CAAA,CAAC;QAC1D,IAAI,IAAI,CAAC;QACT,IAAI,QAAQ;YAAE,IAAI,GAAG,QAAQ,CAAC,CAAC,4BAA4B;QAC3D,MAAM,UAAU,GAAG,aAAa,CAAC,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,eAAe,CAAC,CAAC;QAE1D,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,gBAAgB,CAAC;YAC7C,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,QAAQ,EAAE,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC;YAC9B,YAAY,EAAE,IAAI;YAClB,MAAM,EAAE;gBACN,KAAK,EAAE,QAAQ;gBACf,WAAW,EAAE,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,SAAS;gBAChE,KAAK,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,KAAK,EAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS;gBAC/D,mBAAmB,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,mBAAmB,EAC9C,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,mBAAmB,CAAC;oBAC5C,CAAC,CAAC,SAAS;gBACb,WAAW,EAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,WAAW;gBAChC,eAAe,QAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,eAAe,mCAAI,IAAI;gBAChD,qBAAqB,EAAE,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,qBAAqB;gBACpD,GAAG,EAAE,QAAO,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,GAAG,CAAA,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,GAAG;gBAC1E,oBAAoB,EAAE,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,oBAAoB,EAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CACjE,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,oBAAoB,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,KAAK,CAAC,EAAE,EAAE;oBACvE,GAAG,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC,0CAA0C;oBAC5E,OAAO,GAAG,CAAC;gBACb,CAAC,EAAE,EAAyB,CAAC,CAC9B,CAAC,CAAC,CAAC,SAAS;aACd;SACF,CAAC,CAAC;QAEH,IAAI,CAAC,QAAQ,EAAE;YACb,OAAO;gBACL,IAAI,EAAE,YAAY;gBAClB,MAAM,EAAE,MAAgB;gBACxB,SAAS,EAAE,UAAU,CAAC,MAAM,GAAG,CAAC;gBAChC,WAAW,EAAE,UAAU;aACxB,CAAC;SACH;QACD,MAAM,WAAW,GAAG,MAAkC,CAAC;QACvD,WAAW,CAAC,IAAI,GAAG,OAAO,CAAC;QAC3B,WAAW,CAAC,SAAS,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;QAC9C,WAAW,CAAC,WAAW,GAAG,UAAU,CAAC;QACrC,OAAO,WAAW,CAAC;IACrB,CAAC;IAED;;;;;;;;;OASG;IACH,KAAK,CAAC,UAAU,CACd,MAAwB,EACxB,QAAoC;QAEpC,MAAM,YAAY,mCACb,MAAM,KACT,MAAM,EAAE,MAAM,CAAC,MAAM,IAAI,EAAE,EAC3B,uBAAuB,EAAE,CAAC,CAAC,QAAQ,GACpC,CAAC;QAEF,IAAI,MAAM,CAAC,QAAQ,EAAE;YACnB,MAAM,eAAe,GAAG,MAAM,IAAI,CAAC,gBAAgB,CACjD,MAAM,CAAC,QAAQ,EACf,MAAM,CAAC,aAAa,IAAI,MAAM,CAAC,YAAY,EAC3C;gBACE,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;gBAC/C,WAAW,EAAE,MAAM,CAAC,WAAW;gBAC/B,eAAe,EAAE,MAAM,CAAC,eAAe;gBACvC,qBAAqB,EAAE,MAAM,CAAC,qBAAqB;gBACnD,GAAG,EAAE,MAAM,CAAC,GAAG;gBACf,oBAAoB,EAAE,MAAM,CAAC,oBAAoB;aAClD,CACF,CAAC;YACF,IAAI,eAAe,CAAC,IAAI,KAAK,OAAO,EAAE;gBACpC,MAAM,WAAW,GAAG,eAA2C,CAAC;gBAEhE,YAAY,CAAC,MAAM,GAAG,WAAW,CAAC,MAAM,IAAI,EAAE,CAAC;gBAC/C,IAAI,OAAO,WAAW,CAAC,WAAW,KAAK,QAAQ;oBAC7C,YAAY,CAAC,WAAW,GAAG,WAAW,CAAC,WAAW,CAAC;gBACrD,IAAI,WAAW,CAAC,OAAO;oBAAE,YAAY,CAAC,OAAO,GAAG,WAAW,CAAC,OAAO,CAAC;gBACpE,IAAI,OAAO,WAAW,CAAC,YAAY,KAAK,SAAS;oBAC/C,YAAY,CAAC,YAAY,GAAG,WAAW,CAAC,YAAY,CAAC;gBACvD,IAAI,WAAW,CAAC,gBAAgB;oBAC9B,YAAY,CAAC,gBAAgB,GAAG,WAAW,CAAC,gBAAgB,CAAC;gBAC/D,IAAI,WAAW,CAAC,gBAAgB;oBAC9B,YAAY,CAAC,gBAAgB,GAAG,WAAW,CAAC,gBAAgB,CAAC;gBAC/D,IAAI,WAAW,CAAC,gBAAgB,EAAE;oBAChC,IAAI,CAAC,YAAY,CAAC,IAAI;wBAAE,YAAY,CAAC,IAAI,GAAG,EAAE,CAAC;oBAC/C,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,WAAW,CAAC,gBAAgB,CAAC,CAAC;iBACzD;gBACD,IAAI,WAAW,CAAC,SAAS,EAAE;oBACzB,YAAY,CAAC,WAAW,GAAG,WAAW,CAAC,WAAW,CAAC;iBACpD;aACF;iBAAM,IAAI,eAAe,CAAC,IAAI,KAAK,YAAY,EAAE;gBAChD,MAAM,eAAe,GAAG,eAAsC,CAAC;gBAC/D,YAAY,CAAC,MAAM,GAAG,eAAe,CAAC,MAAM,IAAI,EAAE,CAAC;gBACnD,IAAI,eAAe,CAAC,SAAS,EAAE;oBAC7B,YAAY,CAAC,WAAW,GAAG,eAAe,CAAC,WAAW,CAAC;iBACxD;aACF;SACF;aAAM;YACL,YAAY,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,EAAE,CAAC;SAC3C;QAED,+EAA+E;QAC/E,IAAI,CAAC,YAAY,CAAC,WAAW,IAAI,MAAM,CAAC,WAAW,EAAE;YACnD,YAAY,CAAC,WAAW,GAAG,MAAM,CAAC,WAAW,CAAC;SAC/C;QAED,uCAAuC;QACvC,IAAI,MAAM,CAAC,OAAO,EAAE;YAClB,uCAAuC;YACvC,YAAY,CAAC,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;SACvC;aAAM,IAAI,YAAY,CAAC,eAAe,IAAI,CAAC,YAAY,CAAC,OAAO,EAAE;YAChE,MAAM,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC,eAAe,CAAC,CAAC;YACzD,IAAI,UAAU,EAAE;gBACd,6CAA6C;gBAC7C,IAAI;oBACF,YAAY,CAAC,OAAO,GAAG,MAAM,mBAAmB,CAAC,UAAU,CAAC,CAAC;iBAC9D;gBAAC,OAAO,KAAK,EAAE;oBACd,OAAO,CAAC,IAAI,CAAC,kFAAkF,EAAE,KAAK,CAAC,CAAC;oBACxG,YAAY,CAAC,WAAW,GAAG,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC;iBACvD;aACF;SACF;QAED,IAAI,aAAa,GACf,QAAQ;YACR,QAAQ,CAAC,WAAW,CAAC,cAAc,EAAE,CAAC,GAAqB,EAAE,EAAE;gBAC7D,MAAM,EAAE,SAAS,EAAE,WAAW,EAAE,GAAG,GAAG,CAAC;gBACvC,IAAI,SAAS,KAAK,IAAI,CAAC,EAAE;oBAAE,OAAO;gBAClC,QAAQ,CAAC,WAAW,CAAC,CAAC;YACxB,CAAC,CAAC,CAAC;QAEL,IAAI,CAAC,YAAY,CAAC,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,CAAC,CAAC;QAEhE,MAAM,OAAO,GAAG,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC,CAAC;QAClF,OAAO,OAAO;aACX,IAAI,CAAC,CAAC,gBAAgB,EAAE,EAAE;YACzB,aAAa,aAAb,aAAa,uBAAb,aAAa,CAAE,MAAM,GAAG;YACxB,aAAa,GAAG,IAAI,CAAC;YACrB,OAAO,gBAAgB,CAAC;QAC1B,CAAC,CAAC;aACD,KAAK,CAAC,CAAC,GAAQ,EAAE,EAAE;YAClB,aAAa,aAAb,aAAa,uBAAb,aAAa,CAAE,MAAM,GAAG;YACxB,aAAa,GAAG,IAAI,CAAC;YACrB,MAAM,GAAG,CAAC;QACZ,CAAC,CAAC,CAAC;IACP,CAAC;IAED,cAAc;QACZ,OAAO,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACzD,CAAC;IAED;;;;;OAKG;IACH,QAAQ,CACN,IAAY,EACZ,EACE,WAAW,EAAE,UAAU,MAGrB,EAAE;QAEN,OAAO,QAAQ,CAAC,QAAQ,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,UAAU,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,UAAU,CAAC,MAAgB;QACzB,OAAO,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAC7D,CAAC;IAED,SAAS,CACP,IAAY,EACZ,MAAwB;QAExB,OAAO,QAAQ,CAAC,SAAS,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,IAAI,EAAE,EAAE,CAAC,CAAC;IAChF,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,MAAM,CACV,KAAa,EACb,SAAmB,EACnB,MAAqB;QAErB,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAC,MAAM,CAAC;YACpC,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,KAAK;YACL,SAAS;YACT,MAAM,EAAE,MAAM,IAAI,EAAE;SACrB,CAAC,CAAC;QAEH,8DAA8D;QAC9D,OAAO,OAAO;aACX,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,iCACZ,MAAM,KACT,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,IACjC,CAAC;aACF,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;IACvC,CAAC;IAED,KAAK,CAAC,KAAK,CACT,EAAU,EACV,EAAU,EACV,EAAU,EACV,EAAU;QAEV,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QAC5E,MAAM,CAAC,SAAS,EAAE,SAAS,EAAE,YAAY,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,GACpE,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QACrB,OAAO;YACL,SAAS;YACT,SAAS;YACT,YAAY;YACZ,KAAK;YACL,KAAK;YACL,KAAK;YACL,KAAK;SACN,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,iBAAiB,CACrB,QAAkD;QAElD,IAAI,YAAY,GAA6C,EAAE,CAAC;QAChE,IAAI,QAAQ;YACV,YAAY,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAClC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC;gBACrC,MAAM,EAAE,CAAC,CAAC,MAAM;aACjB,CAAC,CAAC,CAAC;QACN,OAAO,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,YAAY,EAAE,CAAC,CAAC;IAC1E,CAAC;IAED,KAAK,CAAC,kBAAkB;QACtB,OAAO,QAAQ,CAAC,kBAAkB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAC7D,CAAC;IAED,KAAK,CAAC,qBAAqB;QAGzB,OAAO,QAAQ,CAAC,qBAAqB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAChE,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,cAAc,CAAC,EACnB,IAAI,EACJ,OAAO,EAAE,MAAM,GAIhB;QACC,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,QAAQ,CAAC,cAAc,CAAC;YAC7B,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,MAAM,EAAE;gBACN,IAAI;gBACJ,OAAO,EAAE,MAAM,aAAN,MAAM,cAAN,MAAM,GAAI,IAAI;aACxB;SACF,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,mBAAmB;QACvB,OAAO,MAAM,QAAQ,CAAC,mBAAmB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACpE,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,oBAAoB;QAIxB,OAAO,MAAM,QAAQ,CAAC,oBAAoB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACrE,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,iBAAiB;QACrB,OAAO,MAAM,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAClE,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,WAAW,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAsC;QAC7E,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACrD,OAAO,MAAM,QAAQ,CAAC,WAAW,CAAC;YAChC,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,MAAM,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE;SAClC,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,gBAAgB;QACpB,OAAO,MAAM,QAAQ,CAAC,gBAAgB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACjE,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,2BAA2B,CAC/B,OAAsB,EACtB,WAAmB;QAKnB,OAAO,MAAM,QAAQ,CAAC,2BAA2B,CAAC;YAChD,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,cAAc,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE;YACtD,WAAW;SACZ,CAAC,CAAC;IACL,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,6BAA6B,CACjC,WAAmB;QAEnB,OAAO,MAAM,QAAQ,CAAC,6BAA6B,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,WAAW,EAAE,CAAC,CAAC;IAC3F,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,iBAAiB,CAAC,MAAgB;QACtC,OAAO,MAAM,QAAQ,CAAC,iBAAiB,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAC1E,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,cAAc;QAClB,OAAO,MAAM,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAC/D,CAAC;IAED,KAAK,CAAC,OAAO;QACX,OAAO,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IACzD,CAAC;CACF;AAED,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,OAAgB;IACpD,OAAO,QAAQ,CAAC,eAAe,CAAC,EAAE,OAAO,EAAE,CAAC,CAAC;AAC/C,CAAC;AAED,MAAM,UAAU,oBAAoB,CAClC,QAA+C;IAE/C,YAAY,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IAC5B,OAAO;QACL,MAAM,EAAE,GAAG,EAAE;YACX,YAAY,CAAC,MAAM,CAAC,YAAY,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,CAAC;QACzD,CAAC;KACF,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,KAAa;IACjD,OAAO,QAAQ,CAAC,eAAe,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;AAC7C,CAAC;AAED,IAAI,gBAAgB,GAAG,CAAC,CAAC;AACzB,MAAM,eAAe,GAAG,GAAG,EAAE,CAC3B,OAAO,CAAC,GAAG,CAAC,QAAQ,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,MAAM,CAAC,CAAC;AAE3E,MAAM,aAAa,GAAG;IACpB,eAAe;IACf,uBAAuB;IACvB,2BAA2B;IAC3B,uBAAuB;IACvB,uBAAuB;CACxB,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAC,KAAa;IACpD,IAAI,IAAI,GAAG,KAAK,CAAC;IACjB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;QAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACrD,OAAO,QAAQ,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,aAAa,EAAE,CAAC,CAAC;AAC3D,CAAC;AAED,MAAM,WAAW,GAAG;IAClB,iCAAiC;IACjC,IAAI,EAAE,CAAC;IACP,IAAI,EAAE,CAAC;IACP,GAAG,EAAE,CAAC;IACN,IAAI,EAAE,CAAC;IACP,IAAI,EAAE,CAAC;CACR,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,EAOgB,EAChB,UAAuC;QARvC,EACE,KAAK,EACL,cAAc,EAAE,YAAY,EAC5B,YAAY,EAAE,WAAW,EACzB,IAAI,EACJ,SAAS,EAAE,QAAQ,OAEL,EADX,IAAI,cANT,gEAOC,CADQ;IAIT,IAAI,IAAI,GAAG,KAAK,CAAC;IACjB,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC;QAAE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAErD,IAAI,QAAQ,GAAG,IAAI,CAAC;IACpB,IAAI,QAAQ,aAAR,QAAQ,uBAAR,QAAQ,CAAE,UAAU,CAAC,SAAS;QAAG,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAElE,IAAI,YAAY,GAA6C,EAAE,CAAC;IAChE,IAAI,QAAQ;QACV,YAAY,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAClC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC;YACrC,MAAM,EAAE,CAAC,CAAC,MAAM;SACjB,CAAC,CAAC,CAAC;IAEN,MAAM,SAAS,GAAG,gBAAgB,GAAG,eAAe,EAAE,CAAC;IACvD,gBAAgB,IAAI,CAAC,CAAC;IAEtB,IAAI,sBAAsB,GAAQ,IAAI,CAAC;IACvC,IAAI,UAAU,EAAE;QACd,sBAAsB,GAAG,QAAQ,CAAC,WAAW,CAC3C,8BAA8B,EAC9B,CAAC,GAA4C,EAAE,EAAE;YAC/C,IAAI,GAAG,CAAC,SAAS,KAAK,SAAS;gBAAE,OAAO;YACxC,UAAU,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC3B,CAAC,CACF,CAAC;KACH;IAED,MAAM,QAAQ,GAAG,WAAW,CAAC,WAAuC,CAAC,CAAC;IAEtE,IAAI,IAAI,CAAC,YAAY,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,YAAY,CAAC,EAAE;QACrE,OAAO,CAAC,IAAI,CAAC,+CAA+C,IAAI,CAAC,YAAY,uBAAuB,CAAC,CAAC;QACtG,OAAO,IAAI,CAAC,YAAY,CAAC;KAC1B;IACD,IAAI,IAAI,CAAC,YAAY,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,YAAY,CAAC,EAAE;QACrE,OAAO,CAAC,IAAI,CAAC,+CAA+C,IAAI,CAAC,YAAY,uBAAuB,CAAC,CAAC;QACtG,OAAO,IAAI,CAAC,YAAY,CAAC;KAC1B;IAED,oDAAoD;IACpD,IAAI,IAAI,CAAC,WAAW,EAAE;QACpB,OAAO,CAAC,GAAG,CAAC;sBACM,IAAI;uBACH,IAAI,CAAC,WAAW;+BACR,IAAI,CAAC,mBAAmB,IAAI,CAAC;+BAC7B,IAAI,CAAC,kBAAkB,KAAK,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,EAAE,CAAC,CAAC;KAC1F;IAED,MAAM,EACJ,GAAG,EACH,WAAW,EACX,KAAK,EAAE,YAAY,EACnB,UAAU,GACX,GAAG,MAAM,QAAQ,CAAC,WAAW,CAAC;QAC7B,SAAS;QACT,MAAM,kBACJ,KAAK,EAAE,IAAI,EACX,cAAc,EAAE,CAAC,CAAC,YAAY,EAC9B,qBAAqB,EAAE,CAAC,CAAC,UAAU,EACnC,YAAY,EAAE,QAAQ,EACtB,IAAI,EAAE,QAAQ,EACd,SAAS,EAAE,YAAY,IACpB,IAAI,CACR;KACF,CAAC,CAAC,KAAK,CAAC,CAAC,GAAQ,EAAE,EAAE;QACpB,sBAAsB,aAAtB,sBAAsB,uBAAtB,sBAAsB,CAAE,MAAM,GAAG;QACjC,MAAM,GAAG,CAAC;IACZ,CAAC,CAAC,CAAC;IACH,sBAAsB,aAAtB,sBAAsB,uBAAtB,sBAAsB,CAAE,MAAM,GAAG;IACjC,OAAO,IAAI,YAAY,CAAC;QACtB,SAAS;QACT,GAAG;QACH,WAAW;QACX,KAAK,EAAE,YAAY;QACnB,UAAU;KACX,CAAC,CAAC;AACL,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,eAAe;IACnC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,CAAC;AACvC,CAAC;AAED,0CAA0C;AAC1C,MAAM,CAAC,KAAK,UAAU,aAAa,CAAC,GAAW,EAAE,QAAgB;IAC/D,OAAO,QAAQ,CAAC,aAAa,CAAC,EAAE,GAAG,EAAE,QAAQ,EAAE,CAAC,CAAC;AACnD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,mBAAmB,CAAC,GAAW;IASnD,OAAO,QAAQ,CAAC,mBAAmB,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;AAC/C,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,GAAW;IAC9C,OAAO,QAAQ,CAAC,cAAc,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;AAC1C,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,kBAAkB;IAKtC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,CAAC;AACvC,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,0BAA0B,CAAC,MAAc;IAC7D,OAAO,mBAAmB,CAAC,MAAM,CAAC,CAAC;AACrC,CAAC;AAED,MAAM,CAAC,MAAM,SAAS,GAAG;IACvB,MAAM,EAAE,OAAO;IACf,MAAM,EAAE,qBAAqB;CAC9B,CAAC;AAEF,yCAAyC;AACzC,OAAO,EAAE,QAAQ,EAAE,CAAC","sourcesContent":["import { registerPlugin } from '@capacitor/core';\r\nimport type {\r\n  NativeContextParams,\r\n  NativeLlamaContext,\r\n  NativeCompletionParams,\r\n  NativeCompletionTokenProb,\r\n  NativeCompletionResult,\r\n  NativeTokenizeResult,\r\n  NativeEmbeddingResult,\r\n  NativeSessionLoadResult,\r\n  NativeEmbeddingParams,\r\n  NativeRerankParams,\r\n  NativeRerankResult,\r\n  NativeCompletionTokenProbItem,\r\n  NativeCompletionResultTimings,\r\n  JinjaFormattedChatResult,\r\n  FormattedChatResult,\r\n  NativeImageProcessingResult,\r\n  NativeLlamaChatMessage,\r\n  LlamaCppMessagePart,\r\n  LlamaCppOAICompatibleMessage,\r\n  ContextParams,\r\n  EmbeddingParams,\r\n  RerankParams,\r\n  RerankResult,\r\n  CompletionResponseFormat,\r\n  CompletionParams,\r\n  BenchResult,\r\n  LlamaCppPlugin,\r\n} from './definitions';\r\n\r\n// Constants\r\nexport const LLAMACPP_MTMD_DEFAULT_MEDIA_MARKER = '<__media__>';\r\n\r\n// Event names\r\nconst EVENT_ON_INIT_CONTEXT_PROGRESS = '@LlamaCpp_onInitContextProgress';\r\nconst EVENT_ON_TOKEN = '@LlamaCpp_onToken';\r\nconst EVENT_ON_NATIVE_LOG = '@LlamaCpp_onNativeLog';\r\n\r\n// Register the plugin\r\nconst LlamaCpp = registerPlugin<LlamaCppPlugin>('LlamaCpp');\r\n\r\n// Log listeners management\r\nconst logListeners: Array<(level: string, text: string) => void> = [];\r\n\r\n// Set up native log listener\r\nLlamaCpp.addListener(EVENT_ON_NATIVE_LOG, (evt: { level: string; text: string }) => {\r\n  logListeners.forEach((listener) => listener(evt.level, evt.text));\r\n});\r\n\r\n// Trigger unset to use default log callback\r\nLlamaCpp?.toggleNativeLog?.({ enabled: false })?.catch?.(() => {});\r\n\r\n// High-level types for the plugin interface\r\nexport type RNLlamaMessagePart = LlamaCppMessagePart;\r\nexport type RNLlamaOAICompatibleMessage = LlamaCppOAICompatibleMessage;\r\n\r\n// Re-export all types from definitions\r\nexport type {\r\n  NativeContextParams,\r\n  NativeLlamaContext,\r\n  NativeCompletionParams,\r\n  NativeCompletionTokenProb,\r\n  NativeCompletionResult,\r\n  NativeTokenizeResult,\r\n  NativeEmbeddingResult,\r\n  NativeSessionLoadResult,\r\n  NativeEmbeddingParams,\r\n  NativeRerankParams,\r\n  NativeRerankResult,\r\n  NativeCompletionTokenProbItem,\r\n  NativeCompletionResultTimings,\r\n  FormattedChatResult,\r\n  JinjaFormattedChatResult,\r\n  NativeImageProcessingResult,\r\n  ContextParams,\r\n  EmbeddingParams,\r\n  RerankParams,\r\n  RerankResult,\r\n  CompletionResponseFormat,\r\n  CompletionParams,\r\n  BenchResult,\r\n};\r\n\r\nexport const RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER = LLAMACPP_MTMD_DEFAULT_MEDIA_MARKER;\r\n\r\nexport type ToolCall = {\r\n  type: 'function';\r\n  id?: string;\r\n  function: {\r\n    name: string;\r\n    arguments: string; // JSON string\r\n  };\r\n};\r\n\r\nexport type TokenData = {\r\n  token: string;\r\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\r\n  // Parsed content from accumulated text\r\n  content?: string;\r\n  reasoning_content?: string;\r\n  tool_calls?: Array<ToolCall>;\r\n  accumulated_text?: string;\r\n};\r\n\r\ntype TokenNativeEvent = {\r\n  contextId: number;\r\n  tokenResult: TokenData;\r\n};\r\n\r\nconst validCacheTypes = [\r\n  'f16',\r\n  'f32',\r\n  'bf16',\r\n  'q8_0',\r\n  'q4_0',\r\n  'q4_1',\r\n  'iq4_nl',\r\n  'q5_0',\r\n  'q5_1',\r\n];\r\n\r\nconst getJsonSchema = (responseFormat?: CompletionResponseFormat) => {\r\n  if (responseFormat?.type === 'json_schema') {\r\n    return responseFormat.json_schema?.schema;\r\n  }\r\n  if (responseFormat?.type === 'json_object') {\r\n    return responseFormat.schema || {};\r\n  }\r\n  return null;\r\n};\r\n\r\n// Utility function to convert JSON schema to GBNF grammar\r\nconst jsonSchemaToGrammar = async (schema: object): Promise<string> => {\r\n  // This will call the native method to convert JSON schema to GBNF\r\n  // For now, we'll return a basic implementation\r\n  try {\r\n    const result = await LlamaCpp.convertJsonSchemaToGrammar({ schema: JSON.stringify(schema) });\r\n    return result;\r\n  } catch (error) {\r\n    console.warn('Failed to convert JSON schema to GBNF, using fallback:', error);\r\n    // Fallback for basic object structure\r\n    return `root ::= \"{\" ws object_content ws \"}\"\r\nobject_content ::= string_field (\",\" ws string_field)*\r\nstring_field ::= \"\\\\\"\" [a-zA-Z_][a-zA-Z0-9_]* \"\\\\\"\" ws \":\" ws value\r\nvalue ::= string | number | boolean | \"null\"\r\nstring ::= \"\\\\\"\" [^\"]* \"\\\\\"\"\r\nnumber ::= \"-\"? [0-9]+ (\".\" [0-9]+)?\r\nboolean ::= \"true\" | \"false\"\r\nws ::= [ \\\\t\\\\n]*`;\r\n  }\r\n};\r\n\r\nexport class LlamaContext {\r\n  id: number;\r\n  gpu: boolean = false;\r\n  reasonNoGPU: string = '';\r\n  model: NativeLlamaContext['model'];\r\n\r\n  constructor({ contextId, gpu, reasonNoGPU, model }: NativeLlamaContext) {\r\n    this.id = contextId;\r\n    this.gpu = gpu;\r\n    this.reasonNoGPU = reasonNoGPU;\r\n    this.model = model;\r\n  }\r\n\r\n  /**\r\n   * Load cached prompt & completion state from a file.\r\n   */\r\n  async loadSession(filepath: string): Promise<NativeSessionLoadResult> {\r\n    let path = filepath;\r\n    if (path.startsWith('file://')) path = path.slice(7);\r\n    return LlamaCpp.loadSession({ contextId: this.id, filepath: path });\r\n  }\r\n\r\n  /**\r\n   * Save current cached prompt & completion state to a file.\r\n   */\r\n  async saveSession(\r\n    filepath: string,\r\n    options?: { tokenSize: number },\r\n  ): Promise<number> {\r\n    return LlamaCpp.saveSession({ \r\n      contextId: this.id, \r\n      filepath, \r\n      size: options?.tokenSize || -1 \r\n    });\r\n  }\r\n\r\n  isLlamaChatSupported(): boolean {\r\n    return !!this.model.chatTemplates.llamaChat;\r\n  }\r\n\r\n  isJinjaSupported(): boolean {\r\n    const { minja } = this.model.chatTemplates;\r\n    return !!minja?.toolUse || !!minja?.default;\r\n  }\r\n\r\n  async getFormattedChat(\r\n    messages: RNLlamaOAICompatibleMessage[],\r\n    template?: string | null,\r\n    params?: {\r\n      jinja?: boolean;\r\n      response_format?: CompletionResponseFormat;\r\n      tools?: object;\r\n      parallel_tool_calls?: object;\r\n      tool_choice?: string;\r\n      enable_thinking?: boolean;\r\n      add_generation_prompt?: boolean;\r\n      now?: string | number;\r\n      chat_template_kwargs?: Record<string, string>;\r\n    },\r\n  ): Promise<FormattedChatResult | JinjaFormattedChatResult> {\r\n    const mediaPaths: string[] = [];\r\n    const chat = messages.map((msg) => {\r\n      if (Array.isArray(msg.content)) {\r\n        const content = msg.content.map((part) => {\r\n          // Handle multimodal content\r\n          if (part.type === 'image_url') {\r\n            let path = part.image_url?.url || '';\r\n            if (path?.startsWith('file://')) path = path.slice(7);\r\n            mediaPaths.push(path);\r\n            return {\r\n              type: 'text',\r\n              text: RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER,\r\n            };\r\n          } else if (part.type === 'input_audio') {\r\n            const { input_audio: audio } = part;\r\n            if (!audio) throw new Error('input_audio is required');\r\n\r\n            const { format } = audio;\r\n            if (format != 'wav' && format != 'mp3') {\r\n              throw new Error(`Unsupported audio format: ${format}`);\r\n            }\r\n            if (audio.url) {\r\n              const path = audio.url.replace(/file:\\/\\//, '');\r\n              mediaPaths.push(path);\r\n            } else if (audio.data) {\r\n              mediaPaths.push(audio.data);\r\n            }\r\n            return {\r\n              type: 'text',\r\n              text: RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER,\r\n            };\r\n          }\r\n          return part;\r\n        });\r\n\r\n        return {\r\n          ...msg,\r\n          content,\r\n        };\r\n      }\r\n      return msg;\r\n    }) as NativeLlamaChatMessage[];\r\n\r\n    const useJinja = this.isJinjaSupported() && params?.jinja;\r\n    let tmpl;\r\n    if (template) tmpl = template; // Force replace if provided\r\n    const jsonSchema = getJsonSchema(params?.response_format);\r\n\r\n    const result = await LlamaCpp.getFormattedChat({\r\n      contextId: this.id,\r\n      messages: JSON.stringify(chat),\r\n      chatTemplate: tmpl,\r\n      params: {\r\n        jinja: useJinja,\r\n        json_schema: jsonSchema ? JSON.stringify(jsonSchema) : undefined,\r\n        tools: params?.tools ? JSON.stringify(params.tools) : undefined,\r\n        parallel_tool_calls: params?.parallel_tool_calls\r\n          ? JSON.stringify(params.parallel_tool_calls)\r\n          : undefined,\r\n        tool_choice: params?.tool_choice,\r\n        enable_thinking: params?.enable_thinking ?? true,\r\n        add_generation_prompt: params?.add_generation_prompt,\r\n        now: typeof params?.now === 'number' ? params.now.toString() : params?.now,\r\n        chat_template_kwargs: params?.chat_template_kwargs ? JSON.stringify(\r\n          Object.entries(params.chat_template_kwargs).reduce((acc, [key, value]) => {\r\n            acc[key] = JSON.stringify(value); // Each value is a stringified JSON object\r\n            return acc;\r\n          }, {} as Record<string, any>)\r\n        ) : undefined,\r\n      },\r\n    });\r\n\r\n    if (!useJinja) {\r\n      return {\r\n        type: 'llama-chat',\r\n        prompt: result as string,\r\n        has_media: mediaPaths.length > 0,\r\n        media_paths: mediaPaths,\r\n      };\r\n    }\r\n    const jinjaResult = result as JinjaFormattedChatResult;\r\n    jinjaResult.type = 'jinja';\r\n    jinjaResult.has_media = mediaPaths.length > 0;\r\n    jinjaResult.media_paths = mediaPaths;\r\n    return jinjaResult;\r\n  }\r\n\r\n  /**\r\n   * Generate a completion based on the provided parameters\r\n   * @param params Completion parameters including prompt or messages\r\n   * @param callback Optional callback for token-by-token streaming\r\n   * @returns Promise resolving to the completion result\r\n   *\r\n   * Note: For multimodal support, you can include an media_paths parameter.\r\n   * This will process the images and add them to the context before generating text.\r\n   * Multimodal support must be enabled via initMultimodal() first.\r\n   */\r\n  async completion(\r\n    params: CompletionParams,\r\n    callback?: (data: TokenData) => void,\r\n  ): Promise<NativeCompletionResult> {\r\n    const nativeParams = {\r\n      ...params,\r\n      prompt: params.prompt || '',\r\n      emit_partial_completion: !!callback,\r\n    };\r\n\r\n    if (params.messages) {\r\n      const formattedResult = await this.getFormattedChat(\r\n        params.messages,\r\n        params.chat_template || params.chatTemplate,\r\n        {\r\n          jinja: params.jinja,\r\n          tools: params.tools,\r\n          parallel_tool_calls: params.parallel_tool_calls,\r\n          tool_choice: params.tool_choice,\r\n          enable_thinking: params.enable_thinking,\r\n          add_generation_prompt: params.add_generation_prompt,\r\n          now: params.now,\r\n          chat_template_kwargs: params.chat_template_kwargs,\r\n        },\r\n      );\r\n      if (formattedResult.type === 'jinja') {\r\n        const jinjaResult = formattedResult as JinjaFormattedChatResult;\r\n\r\n        nativeParams.prompt = jinjaResult.prompt || '';\r\n        if (typeof jinjaResult.chat_format === 'number')\r\n          nativeParams.chat_format = jinjaResult.chat_format;\r\n        if (jinjaResult.grammar) nativeParams.grammar = jinjaResult.grammar;\r\n        if (typeof jinjaResult.grammar_lazy === 'boolean')\r\n          nativeParams.grammar_lazy = jinjaResult.grammar_lazy;\r\n        if (jinjaResult.grammar_triggers)\r\n          nativeParams.grammar_triggers = jinjaResult.grammar_triggers;\r\n        if (jinjaResult.preserved_tokens)\r\n          nativeParams.preserved_tokens = jinjaResult.preserved_tokens;\r\n        if (jinjaResult.additional_stops) {\r\n          if (!nativeParams.stop) nativeParams.stop = [];\r\n          nativeParams.stop.push(...jinjaResult.additional_stops);\r\n        }\r\n        if (jinjaResult.has_media) {\r\n          nativeParams.media_paths = jinjaResult.media_paths;\r\n        }\r\n      } else if (formattedResult.type === 'llama-chat') {\r\n        const llamaChatResult = formattedResult as FormattedChatResult;\r\n        nativeParams.prompt = llamaChatResult.prompt || '';\r\n        if (llamaChatResult.has_media) {\r\n          nativeParams.media_paths = llamaChatResult.media_paths;\r\n        }\r\n      }\r\n    } else {\r\n      nativeParams.prompt = params.prompt || '';\r\n    }\r\n\r\n    // If media_paths were explicitly provided or extracted from messages, use them\r\n    if (!nativeParams.media_paths && params.media_paths) {\r\n      nativeParams.media_paths = params.media_paths;\r\n    }\r\n\r\n    // Handle structured output and grammar\r\n    if (params.grammar) {\r\n      // Direct GBNF grammar takes precedence\r\n      nativeParams.grammar = params.grammar;\r\n    } else if (nativeParams.response_format && !nativeParams.grammar) {\r\n      const jsonSchema = getJsonSchema(params.response_format);\r\n      if (jsonSchema) {\r\n        // Try to convert JSON schema to GBNF grammar\r\n        try {\r\n          nativeParams.grammar = await jsonSchemaToGrammar(jsonSchema);\r\n        } catch (error) {\r\n          console.warn('Failed to convert JSON schema to grammar, falling back to json_schema parameter:', error);\r\n          nativeParams.json_schema = JSON.stringify(jsonSchema);\r\n        }\r\n      }\r\n    }\r\n\r\n    let tokenListener: any =\r\n      callback &&\r\n      LlamaCpp.addListener(EVENT_ON_TOKEN, (evt: TokenNativeEvent) => {\r\n        const { contextId, tokenResult } = evt;\r\n        if (contextId !== this.id) return;\r\n        callback(tokenResult);\r\n      });\r\n\r\n    if (!nativeParams.prompt) throw new Error('Prompt is required');\r\n\r\n    const promise = LlamaCpp.completion({ contextId: this.id, params: nativeParams });\r\n    return promise\r\n      .then((completionResult) => {\r\n        tokenListener?.remove();\r\n        tokenListener = null;\r\n        return completionResult;\r\n      })\r\n      .catch((err: any) => {\r\n        tokenListener?.remove();\r\n        tokenListener = null;\r\n        throw err;\r\n      });\r\n  }\r\n\r\n  stopCompletion(): Promise<void> {\r\n    return LlamaCpp.stopCompletion({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Tokenize text or text with images\r\n   * @param text Text to tokenize\r\n   * @param params.media_paths Array of image paths to tokenize (if multimodal is enabled)\r\n   * @returns Promise resolving to the tokenize result\r\n   */\r\n  tokenize(\r\n    text: string,\r\n    {\r\n      media_paths: mediaPaths,\r\n    }: {\r\n      media_paths?: string[];\r\n    } = {},\r\n  ): Promise<NativeTokenizeResult> {\r\n    return LlamaCpp.tokenize({ contextId: this.id, text, imagePaths: mediaPaths });\r\n  }\r\n\r\n  detokenize(tokens: number[]): Promise<string> {\r\n    return LlamaCpp.detokenize({ contextId: this.id, tokens });\r\n  }\r\n\r\n  embedding(\r\n    text: string,\r\n    params?: EmbeddingParams,\r\n  ): Promise<NativeEmbeddingResult> {\r\n    return LlamaCpp.embedding({ contextId: this.id, text, params: params || {} });\r\n  }\r\n\r\n  /**\r\n   * Rerank documents based on relevance to a query\r\n   * @param query The query text to rank documents against\r\n   * @param documents Array of document texts to rank\r\n   * @param params Optional reranking parameters\r\n   * @returns Promise resolving to an array of ranking results with scores and indices\r\n   */\r\n  async rerank(\r\n    query: string,\r\n    documents: string[],\r\n    params?: RerankParams,\r\n  ): Promise<RerankResult[]> {\r\n    const results = await LlamaCpp.rerank({ \r\n      contextId: this.id, \r\n      query, \r\n      documents, \r\n      params: params || {} \r\n    });\r\n\r\n    // Sort by score descending and add document text if requested\r\n    return results\r\n      .map((result) => ({\r\n        ...result,\r\n        document: documents[result.index],\r\n      }))\r\n      .sort((a, b) => b.score - a.score);\r\n  }\r\n\r\n  async bench(\r\n    pp: number,\r\n    tg: number,\r\n    pl: number,\r\n    nr: number,\r\n  ): Promise<BenchResult> {\r\n    const result = await LlamaCpp.bench({ contextId: this.id, pp, tg, pl, nr });\r\n    const [modelDesc, modelSize, modelNParams, ppAvg, ppStd, tgAvg, tgStd] =\r\n      JSON.parse(result);\r\n    return {\r\n      modelDesc,\r\n      modelSize,\r\n      modelNParams,\r\n      ppAvg,\r\n      ppStd,\r\n      tgAvg,\r\n      tgStd,\r\n    };\r\n  }\r\n\r\n  async applyLoraAdapters(\r\n    loraList: Array<{ path: string; scaled?: number }>,\r\n  ): Promise<void> {\r\n    let loraAdapters: Array<{ path: string; scaled?: number }> = [];\r\n    if (loraList)\r\n      loraAdapters = loraList.map((l) => ({\r\n        path: l.path.replace(/file:\\/\\//, ''),\r\n        scaled: l.scaled,\r\n      }));\r\n    return LlamaCpp.applyLoraAdapters({ contextId: this.id, loraAdapters });\r\n  }\r\n\r\n  async removeLoraAdapters(): Promise<void> {\r\n    return LlamaCpp.removeLoraAdapters({ contextId: this.id });\r\n  }\r\n\r\n  async getLoadedLoraAdapters(): Promise<\r\n    Array<{ path: string; scaled?: number }>\r\n  > {\r\n    return LlamaCpp.getLoadedLoraAdapters({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Initialize multimodal support with a mmproj file\r\n   * @param params Parameters for multimodal support\r\n   * @param params.path Path to the multimodal projector file\r\n   * @param params.use_gpu Whether to use GPU\r\n   * @returns Promise resolving to true if initialization was successful\r\n   */\r\n  async initMultimodal({\r\n    path,\r\n    use_gpu: useGpu,\r\n  }: {\r\n    path: string;\r\n    use_gpu?: boolean;\r\n  }): Promise<boolean> {\r\n    if (path.startsWith('file://')) path = path.slice(7);\r\n    return LlamaCpp.initMultimodal({\r\n      contextId: this.id,\r\n      params: {\r\n        path,\r\n        use_gpu: useGpu ?? true,\r\n      },\r\n    });\r\n  }\r\n\r\n  /**\r\n   * Check if multimodal support is enabled\r\n   * @returns Promise resolving to true if multimodal is enabled\r\n   */\r\n  async isMultimodalEnabled(): Promise<boolean> {\r\n    return await LlamaCpp.isMultimodalEnabled({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Check multimodal support\r\n   * @returns Promise resolving to an object with vision and audio support\r\n   */\r\n  async getMultimodalSupport(): Promise<{\r\n    vision: boolean;\r\n    audio: boolean;\r\n  }> {\r\n    return await LlamaCpp.getMultimodalSupport({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Release multimodal support\r\n   * @returns Promise resolving to void\r\n   */\r\n  async releaseMultimodal(): Promise<void> {\r\n    return await LlamaCpp.releaseMultimodal({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Initialize TTS support with a vocoder model\r\n   * @param params Parameters for TTS support\r\n   * @param params.path Path to the vocoder model\r\n   * @param params.n_batch Batch size for the vocoder model\r\n   * @returns Promise resolving to true if initialization was successful\r\n   */\r\n  async initVocoder({ path, n_batch: nBatch }: { path: string; n_batch?: number }): Promise<boolean> {\r\n    if (path.startsWith('file://')) path = path.slice(7);\r\n    return await LlamaCpp.initVocoder({ \r\n      contextId: this.id, \r\n      params: { path, n_batch: nBatch } \r\n    });\r\n  }\r\n\r\n  /**\r\n   * Check if TTS support is enabled\r\n   * @returns Promise resolving to true if TTS is enabled\r\n   */\r\n  async isVocoderEnabled(): Promise<boolean> {\r\n    return await LlamaCpp.isVocoderEnabled({ contextId: this.id });\r\n  }\r\n\r\n  /**\r\n   * Get a formatted audio completion prompt\r\n   * @param speakerJsonStr JSON string representing the speaker\r\n   * @param textToSpeak Text to speak\r\n   * @returns Promise resolving to the formatted audio completion result with prompt and grammar\r\n   */\r\n  async getFormattedAudioCompletion(\r\n    speaker: object | null,\r\n    textToSpeak: string,\r\n  ): Promise<{\r\n    prompt: string;\r\n    grammar?: string;\r\n  }> {\r\n    return await LlamaCpp.getFormattedAudioCompletion({\r\n      contextId: this.id,\r\n      speakerJsonStr: speaker ? JSON.stringify(speaker) : '',\r\n      textToSpeak,\r\n    });\r\n  }\r\n\r\n  /**\r\n   * Get guide tokens for audio completion\r\n   * @param textToSpeak Text to speak\r\n   * @returns Promise resolving to the guide tokens\r\n   */\r\n  async getAudioCompletionGuideTokens(\r\n    textToSpeak: string,\r\n  ): Promise<Array<number>> {\r\n    return await LlamaCpp.getAudioCompletionGuideTokens({ contextId: this.id, textToSpeak });\r\n  }\r\n\r\n  /**\r\n   * Decode audio tokens\r\n   * @param tokens Array of audio tokens\r\n   * @returns Promise resolving to the decoded audio tokens\r\n   */\r\n  async decodeAudioTokens(tokens: number[]): Promise<Array<number>> {\r\n    return await LlamaCpp.decodeAudioTokens({ contextId: this.id, tokens });\r\n  }\r\n\r\n  /**\r\n   * Release TTS support\r\n   * @returns Promise resolving to void\r\n   */\r\n  async releaseVocoder(): Promise<void> {\r\n    return await LlamaCpp.releaseVocoder({ contextId: this.id });\r\n  }\r\n\r\n  async release(): Promise<void> {\r\n    return LlamaCpp.releaseContext({ contextId: this.id });\r\n  }\r\n}\r\n\r\nexport async function toggleNativeLog(enabled: boolean): Promise<void> {\r\n  return LlamaCpp.toggleNativeLog({ enabled });\r\n}\r\n\r\nexport function addNativeLogListener(\r\n  listener: (level: string, text: string) => void,\r\n): { remove: () => void } {\r\n  logListeners.push(listener);\r\n  return {\r\n    remove: () => {\r\n      logListeners.splice(logListeners.indexOf(listener), 1);\r\n    },\r\n  };\r\n}\r\n\r\nexport async function setContextLimit(limit: number): Promise<void> {\r\n  return LlamaCpp.setContextLimit({ limit });\r\n}\r\n\r\nlet contextIdCounter = 0;\r\nconst contextIdRandom = () =>\r\n  process.env.NODE_ENV === 'test' ? 0 : Math.floor(Math.random() * 100000);\r\n\r\nconst modelInfoSkip = [\r\n  // Large fields\r\n  'tokenizer.ggml.tokens',\r\n  'tokenizer.ggml.token_type',\r\n  'tokenizer.ggml.merges',\r\n  'tokenizer.ggml.scores',\r\n];\r\n\r\nexport async function loadLlamaModelInfo(model: string): Promise<Object> {\r\n  let path = model;\r\n  if (path.startsWith('file://')) path = path.slice(7);\r\n  return LlamaCpp.modelInfo({ path, skip: modelInfoSkip });\r\n}\r\n\r\nconst poolTypeMap = {\r\n  // -1 is unspecified as undefined\r\n  none: 0,\r\n  mean: 1,\r\n  cls: 2,\r\n  last: 3,\r\n  rank: 4,\r\n};\r\n\r\nexport async function initLlama(\r\n  {\r\n    model,\r\n    is_model_asset: isModelAsset,\r\n    pooling_type: poolingType,\r\n    lora,\r\n    lora_list: loraList,\r\n    ...rest\r\n  }: ContextParams,\r\n  onProgress?: (progress: number) => void,\r\n): Promise<LlamaContext> {\r\n  let path = model;\r\n  if (path.startsWith('file://')) path = path.slice(7);\r\n\r\n  let loraPath = lora;\r\n  if (loraPath?.startsWith('file://')) loraPath = loraPath.slice(7);\r\n\r\n  let loraAdapters: Array<{ path: string; scaled?: number }> = [];\r\n  if (loraList)\r\n    loraAdapters = loraList.map((l) => ({\r\n      path: l.path.replace(/file:\\/\\//, ''),\r\n      scaled: l.scaled,\r\n    }));\r\n\r\n  const contextId = contextIdCounter + contextIdRandom();\r\n  contextIdCounter += 1;\r\n\r\n  let removeProgressListener: any = null;\r\n  if (onProgress) {\r\n    removeProgressListener = LlamaCpp.addListener(\r\n      EVENT_ON_INIT_CONTEXT_PROGRESS,\r\n      (evt: { contextId: number; progress: number }) => {\r\n        if (evt.contextId !== contextId) return;\r\n        onProgress(evt.progress);\r\n      },\r\n    );\r\n  }\r\n\r\n  const poolType = poolTypeMap[poolingType as keyof typeof poolTypeMap];\r\n\r\n  if (rest.cache_type_k && !validCacheTypes.includes(rest.cache_type_k)) {\r\n    console.warn(`[LlamaCpp] initLlama: Invalid cache K type: ${rest.cache_type_k}, falling back to f16`);\r\n    delete rest.cache_type_k;\r\n  }\r\n  if (rest.cache_type_v && !validCacheTypes.includes(rest.cache_type_v)) {\r\n    console.warn(`[LlamaCpp] initLlama: Invalid cache V type: ${rest.cache_type_v}, falling back to f16`);\r\n    delete rest.cache_type_v;\r\n  }\r\n\r\n  // Log speculative decoding configuration if enabled\r\n  if (rest.draft_model) {\r\n    console.log(`🚀 Initializing with speculative decoding:\r\n      - Main model: ${path}\r\n      - Draft model: ${rest.draft_model}\r\n      - Speculative samples: ${rest.speculative_samples || 3}\r\n      - Mobile optimization: ${rest.mobile_speculative !== false ? 'enabled' : 'disabled'}`);\r\n  }\r\n\r\n  const {\r\n    gpu,\r\n    reasonNoGPU,\r\n    model: modelDetails,\r\n    androidLib,\r\n  } = await LlamaCpp.initContext({\r\n    contextId,\r\n    params: {\r\n      model: path,\r\n      is_model_asset: !!isModelAsset,\r\n      use_progress_callback: !!onProgress,\r\n      pooling_type: poolType,\r\n      lora: loraPath,\r\n      lora_list: loraAdapters,\r\n      ...rest,\r\n    },\r\n  }).catch((err: any) => {\r\n    removeProgressListener?.remove();\r\n    throw err;\r\n  });\r\n  removeProgressListener?.remove();\r\n  return new LlamaContext({\r\n    contextId,\r\n    gpu,\r\n    reasonNoGPU,\r\n    model: modelDetails,\r\n    androidLib,\r\n  });\r\n}\r\n\r\nexport async function releaseAllLlama(): Promise<void> {\r\n  return LlamaCpp.releaseAllContexts();\r\n}\r\n\r\n// Model download and management functions\r\nexport async function downloadModel(url: string, filename: string): Promise<string> {\r\n  return LlamaCpp.downloadModel({ url, filename });\r\n}\r\n\r\nexport async function getDownloadProgress(url: string): Promise<{\r\n  progress: number;\r\n  completed: boolean;\r\n  failed: boolean;\r\n  errorMessage?: string;\r\n  localPath?: string;\r\n  downloadedBytes: number;\r\n  totalBytes: number;\r\n}> {\r\n  return LlamaCpp.getDownloadProgress({ url });\r\n}\r\n\r\nexport async function cancelDownload(url: string): Promise<boolean> {\r\n  return LlamaCpp.cancelDownload({ url });\r\n}\r\n\r\nexport async function getAvailableModels(): Promise<Array<{\r\n  name: string;\r\n  path: string;\r\n  size: number;\r\n}>> {\r\n  return LlamaCpp.getAvailableModels();\r\n}\r\n\r\n/**\r\n * Convert a JSON schema to GBNF grammar format\r\n * @param schema JSON schema object\r\n * @returns Promise resolving to GBNF grammar string\r\n */\r\nexport async function convertJsonSchemaToGrammar(schema: object): Promise<string> {\r\n  return jsonSchemaToGrammar(schema);\r\n}\r\n\r\nexport const BuildInfo = {\r\n  number: '1.0.0',\r\n  commit: 'capacitor-llama-cpp',\r\n};\r\n\r\n// Re-export the plugin for direct access\r\nexport { LlamaCpp };\r\n"]}

package/dist/plugin.cjs.js

@@ -42,6 +42,27 @@ const getJsonSchema = (responseFormat) => {
     }
     return null;
 };
+// Utility function to convert JSON schema to GBNF grammar
+const jsonSchemaToGrammar = async (schema) => {
+    // This will call the native method to convert JSON schema to GBNF
+    // For now, we'll return a basic implementation
+    try {
+        const result = await LlamaCpp.convertJsonSchemaToGrammar({ schema: JSON.stringify(schema) });
+        return result;
+    }
+    catch (error) {
+        console.warn('Failed to convert JSON schema to GBNF, using fallback:', error);
+        // Fallback for basic object structure
+        return `root ::= "{" ws object_content ws "}"
+object_content ::= string_field ("," ws string_field)*
+string_field ::= "\\"" [a-zA-Z_][a-zA-Z0-9_]* "\\"" ws ":" ws value
+value ::= string | number | boolean | "null"
+string ::= "\\"" [^"]* "\\""
+number ::= "-"? [0-9]+ ("." [0-9]+)?
+boolean ::= "true" | "false"
+ws ::= [ \\t\\n]*`;
+    }
+};
 class LlamaContext {
     constructor({ contextId, gpu, reasonNoGPU, model }) {
         this.gpu = false;
@@ -220,10 +241,23 @@ class LlamaContext {
         if (!nativeParams.media_paths && params.media_paths) {
             nativeParams.media_paths = params.media_paths;
         }
-        if (nativeParams.response_format && !nativeParams.grammar) {
+        // Handle structured output and grammar
+        if (params.grammar) {
+            // Direct GBNF grammar takes precedence
+            nativeParams.grammar = params.grammar;
+        }
+        else if (nativeParams.response_format && !nativeParams.grammar) {
             const jsonSchema = getJsonSchema(params.response_format);
-            if (jsonSchema)
-                nativeParams.json_schema = JSON.stringify(jsonSchema);
+            if (jsonSchema) {
+                // Try to convert JSON schema to GBNF grammar
+                try {
+                    nativeParams.grammar = await jsonSchemaToGrammar(jsonSchema);
+                }
+                catch (error) {
+                    console.warn('Failed to convert JSON schema to grammar, falling back to json_schema parameter:', error);
+                    nativeParams.json_schema = JSON.stringify(jsonSchema);
+                }
+            }
         }
         let tokenListener = callback &&
             LlamaCpp.addListener(EVENT_ON_TOKEN, (evt) => {
@@ -483,6 +517,14 @@ async function initLlama(_a, onProgress) {
         console.warn(`[LlamaCpp] initLlama: Invalid cache V type: ${rest.cache_type_v}, falling back to f16`);
         delete rest.cache_type_v;
     }
+    // Log speculative decoding configuration if enabled
+    if (rest.draft_model) {
+        console.log(`🚀 Initializing with speculative decoding:
+      - Main model: ${path}
+      - Draft model: ${rest.draft_model}
+      - Speculative samples: ${rest.speculative_samples || 3}
+      - Mobile optimization: ${rest.mobile_speculative !== false ? 'enabled' : 'disabled'}`);
+    }
     const { gpu, reasonNoGPU, model: modelDetails, androidLib, } = await LlamaCpp.initContext({
         contextId,
         params: Object.assign({ model: path, is_model_asset: !!isModelAsset, use_progress_callback: !!onProgress, pooling_type: poolType, lora: loraPath, lora_list: loraAdapters }, rest),
@@ -502,6 +544,27 @@ async function initLlama(_a, onProgress) {
 async function releaseAllLlama() {
     return LlamaCpp.releaseAllContexts();
 }
+// Model download and management functions
+async function downloadModel(url, filename) {
+    return LlamaCpp.downloadModel({ url, filename });
+}
+async function getDownloadProgress(url) {
+    return LlamaCpp.getDownloadProgress({ url });
+}
+async function cancelDownload(url) {
+    return LlamaCpp.cancelDownload({ url });
+}
+async function getAvailableModels() {
+    return LlamaCpp.getAvailableModels();
+}
+/**
+ * Convert a JSON schema to GBNF grammar format
+ * @param schema JSON schema object
+ * @returns Promise resolving to GBNF grammar string
+ */
+async function convertJsonSchemaToGrammar(schema) {
+    return jsonSchemaToGrammar(schema);
+}
 const BuildInfo = {
     number: '1.0.0',
     commit: 'capacitor-llama-cpp',
@@ -513,6 +576,11 @@ exports.LlamaContext = LlamaContext;
 exports.LlamaCpp = LlamaCpp;
 exports.RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER = RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER;
 exports.addNativeLogListener = addNativeLogListener;
+exports.cancelDownload = cancelDownload;
+exports.convertJsonSchemaToGrammar = convertJsonSchemaToGrammar;
+exports.downloadModel = downloadModel;
+exports.getAvailableModels = getAvailableModels;
+exports.getDownloadProgress = getDownloadProgress;
 exports.initLlama = initLlama;
 exports.loadLlamaModelInfo = loadLlamaModelInfo;
 exports.releaseAllLlama = releaseAllLlama;