llama-cpp-capacitor 0.0.13 → 0.0.22

package/dist/docs.json

@@ -137,6 +137,67 @@
         ],
         "slug": "completion"
       },
+      {
+        "name": "chat",
+        "signature": "(options: { contextId: number; messages: LlamaCppOAICompatibleMessage[]; system?: string; chatTemplate?: string; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>; }) => Promise<NativeCompletionResult>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ contextId: number; messages: LlamaCppOAICompatibleMessage[]; system?: string | undefined; chatTemplate?: string | undefined; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'> | undefined; }"
+          }
+        ],
+        "returns": "Promise<NativeCompletionResult>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [
+          "NativeCompletionResult",
+          "LlamaCppOAICompatibleMessage",
+          "Omit",
+          "NativeCompletionParams"
+        ],
+        "slug": "chat"
+      },
+      {
+        "name": "chatWithSystem",
+        "signature": "(options: { contextId: number; system: string; message: string; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>; }) => Promise<NativeCompletionResult>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ contextId: number; system: string; message: string; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'> | undefined; }"
+          }
+        ],
+        "returns": "Promise<NativeCompletionResult>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [
+          "NativeCompletionResult",
+          "Omit",
+          "NativeCompletionParams"
+        ],
+        "slug": "chatwithsystem"
+      },
+      {
+        "name": "generateText",
+        "signature": "(options: { contextId: number; prompt: string; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>; }) => Promise<NativeCompletionResult>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ contextId: number; prompt: string; params?: Omit<NativeCompletionParams, 'prompt' | 'messages'> | undefined; }"
+          }
+        ],
+        "returns": "Promise<NativeCompletionResult>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [
+          "NativeCompletionResult",
+          "Omit",
+          "NativeCompletionParams"
+        ],
+        "slug": "generatetext"
+      },
       {
         "name": "stopCompletion",
         "signature": "(options: { contextId: number; }) => Promise<void>",
@@ -493,6 +554,82 @@
         "complexTypes": [],
         "slug": "releasevocoder"
       },
+      {
+        "name": "downloadModel",
+        "signature": "(options: { url: string; filename: string; }) => Promise<string>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ url: string; filename: string; }"
+          }
+        ],
+        "returns": "Promise<string>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [],
+        "slug": "downloadmodel"
+      },
+      {
+        "name": "getDownloadProgress",
+        "signature": "(options: { url: string; }) => Promise<{ progress: number; completed: boolean; failed: boolean; errorMessage?: string; localPath?: string; downloadedBytes: number; totalBytes: number; }>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ url: string; }"
+          }
+        ],
+        "returns": "Promise<{ progress: number; completed: boolean; failed: boolean; errorMessage?: string | undefined; localPath?: string | undefined; downloadedBytes: number; totalBytes: number; }>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [],
+        "slug": "getdownloadprogress"
+      },
+      {
+        "name": "cancelDownload",
+        "signature": "(options: { url: string; }) => Promise<boolean>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ url: string; }"
+          }
+        ],
+        "returns": "Promise<boolean>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [],
+        "slug": "canceldownload"
+      },
+      {
+        "name": "getAvailableModels",
+        "signature": "() => Promise<Array<{ name: string; path: string; size: number; }>>",
+        "parameters": [],
+        "returns": "Promise<{ name: string; path: string; size: number; }[]>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [
+          "Array"
+        ],
+        "slug": "getavailablemodels"
+      },
+      {
+        "name": "convertJsonSchemaToGrammar",
+        "signature": "(options: { schema: string; }) => Promise<string>",
+        "parameters": [
+          {
+            "name": "options",
+            "docs": "",
+            "type": "{ schema: string; }"
+          }
+        ],
+        "returns": "Promise<string>",
+        "tags": [],
+        "docs": "",
+        "complexTypes": [],
+        "slug": "convertjsonschematogrammar"
+      },
       {
         "name": "addListener",
         "signature": "(eventName: string, listenerFunc: (data: any) => void) => Promise<void>",
@@ -4031,6 +4168,27 @@
           "complexTypes": [],
           "type": "number | undefined"
         },
+        {
+          "name": "draft_model",
+          "tags": [],
+          "docs": "Path to draft model for speculative decoding (mobile optimization)",
+          "complexTypes": [],
+          "type": "string | undefined"
+        },
+        {
+          "name": "speculative_samples",
+          "tags": [],
+          "docs": "Number of tokens to predict speculatively (default: 3 for mobile)",
+          "complexTypes": [],
+          "type": "number | undefined"
+        },
+        {
+          "name": "mobile_speculative",
+          "tags": [],
+          "docs": "Enable mobile-optimized speculative decoding",
+          "complexTypes": [],
+          "type": "boolean | undefined"
+        },
         {
           "name": "n_gpu_layers",
           "tags": [],
@@ -4263,11 +4421,11 @@
         {
           "name": "tool_calls",
           "tags": [],
-          "docs": "Tool calls",
+          "docs": "Tool calls (parsed from response)",
           "complexTypes": [
             "Array"
           ],
-          "type": "Array<{\r\n    type: 'function';\r\n    function: {\r\n      name: string;\r\n      arguments: string;\r\n    };\r\n    id?: string;\r\n  }>"
+          "type": "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  }>"
         },
         {
           "name": "content",
@@ -4535,7 +4693,7 @@
         {
           "name": "grammar",
           "tags": [],
-          "docs": "Set grammar for grammar-based sampling.  Default: no grammar",
+          "docs": "Set grammar for grammar-based sampling (GBNF format). Default: no grammar\r\nThis will override json_schema if both are provided.",
           "complexTypes": [],
           "type": "string | undefined"
         },
@@ -4807,6 +4965,68 @@
         }
       ]
     },
+    {
+      "name": "LlamaCppOAICompatibleMessage",
+      "slug": "llamacppoaicompatiblemessage",
+      "docs": "",
+      "tags": [],
+      "methods": [],
+      "properties": [
+        {
+          "name": "role",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [],
+          "type": "string"
+        },
+        {
+          "name": "content",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [
+            "LlamaCppMessagePart"
+          ],
+          "type": "string | LlamaCppMessagePart[] | undefined"
+        }
+      ]
+    },
+    {
+      "name": "LlamaCppMessagePart",
+      "slug": "llamacppmessagepart",
+      "docs": "",
+      "tags": [],
+      "methods": [],
+      "properties": [
+        {
+          "name": "type",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [],
+          "type": "string"
+        },
+        {
+          "name": "text",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [],
+          "type": "string | undefined"
+        },
+        {
+          "name": "image_url",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [],
+          "type": "{ url?: string | undefined; } | undefined"
+        },
+        {
+          "name": "input_audio",
+          "tags": [],
+          "docs": "",
+          "complexTypes": [],
+          "type": "{ format: string; data?: string | undefined; url?: string | undefined; } | undefined"
+        }
+      ]
+    },
     {
       "name": "NativeSessionLoadResult",
       "slug": "nativesessionloadresult",
@@ -5507,6 +5727,51 @@
           ]
         }
       ]
+    },
+    {
+      "name": "Omit",
+      "slug": "omit",
+      "docs": "Construct a type with the properties of T except for those in type K.",
+      "types": [
+        {
+          "text": "Pick<T, Exclude<keyof T, K>>",
+          "complexTypes": [
+            "Pick",
+            "T",
+            "Exclude",
+            "K"
+          ]
+        }
+      ]
+    },
+    {
+      "name": "Pick",
+      "slug": "pick",
+      "docs": "From T, pick a set of properties whose keys are in the union K",
+      "types": [
+        {
+          "text": "{\r\n    [P in K]: T[P];\r\n}",
+          "complexTypes": [
+            "K",
+            "T",
+            "P"
+          ]
+        }
+      ]
+    },
+    {
+      "name": "Exclude",
+      "slug": "exclude",
+      "docs": "Exclude from T those types that are assignable to U",
+      "types": [
+        {
+          "text": "T extends U ? never : T",
+          "complexTypes": [
+            "T",
+            "U"
+          ]
+        }
+      ]
     }
   ],
   "pluginConfigs": []

package/dist/esm/definitions.d.ts

@@ -13,6 +13,18 @@ export interface NativeContextParams {
     n_batch?: number;
     n_ubatch?: number;
     n_threads?: number;
+    /**
+     * Path to draft model for speculative decoding (mobile optimization)
+     */
+    draft_model?: string;
+    /**
+     * Number of tokens to predict speculatively (default: 3 for mobile)
+     */
+    speculative_samples?: number;
+    /**
+     * Enable mobile-optimized speculative decoding
+     */
+    mobile_speculative?: boolean;
     /**
      * Number of layers to store in VRAM (Currently only for iOS)
      */
@@ -87,7 +99,8 @@ export interface NativeCompletionParams {
      */
     json_schema?: string;
     /**
-     * Set grammar for grammar-based sampling.  Default: no grammar
+     * Set grammar for grammar-based sampling (GBNF format). Default: no grammar
+     * This will override json_schema if both are provided.
      */
     grammar?: string;
     /**
@@ -268,7 +281,7 @@ export interface NativeCompletionResult {
      */
     reasoning_content: string;
     /**
-     * Tool calls
+     * Tool calls (parsed from response)
      */
     tool_calls: Array<{
         type: 'function';
@@ -483,6 +496,10 @@ export interface CompletionParams extends Omit<NativeCompletionParams, 'emit_par
     chatTemplate?: string;
     chat_template?: string;
     jinja?: boolean;
+    /**
+     * GBNF grammar for structured output. Takes precedence over json_schema.
+     */
+    grammar?: string;
     tools?: object;
     parallel_tool_calls?: object;
     tool_choice?: string;
@@ -545,6 +562,24 @@ export interface LlamaCppPlugin {
         contextId: number;
         params: NativeCompletionParams;
     }): Promise<NativeCompletionResult>;
+    chat(options: {
+        contextId: number;
+        messages: LlamaCppOAICompatibleMessage[];
+        system?: string;
+        chatTemplate?: string;
+        params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;
+    }): Promise<NativeCompletionResult>;
+    chatWithSystem(options: {
+        contextId: number;
+        system: string;
+        message: string;
+        params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;
+    }): Promise<NativeCompletionResult>;
+    generateText(options: {
+        contextId: number;
+        prompt: string;
+        params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;
+    }): Promise<NativeCompletionResult>;
     stopCompletion(options: {
         contextId: number;
     }): Promise<void>;
@@ -648,6 +683,32 @@ export interface LlamaCppPlugin {
     releaseVocoder(options: {
         contextId: number;
     }): Promise<void>;
+    downloadModel(options: {
+        url: string;
+        filename: string;
+    }): Promise<string>;
+    getDownloadProgress(options: {
+        url: string;
+    }): Promise<{
+        progress: number;
+        completed: boolean;
+        failed: boolean;
+        errorMessage?: string;
+        localPath?: string;
+        downloadedBytes: number;
+        totalBytes: number;
+    }>;
+    cancelDownload(options: {
+        url: string;
+    }): Promise<boolean>;
+    getAvailableModels(): Promise<Array<{
+        name: string;
+        path: string;
+        size: number;
+    }>>;
+    convertJsonSchemaToGrammar(options: {
+        schema: string;
+    }): Promise<string>;
     addListener(eventName: string, listenerFunc: (data: any) => void): Promise<void>;
     removeAllListeners(eventName: string): Promise<void>;
 }

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  // Chat-first methods (like llama-cli -sys)\r\n  chat(options: {\r\n    contextId: number;\r\n    messages: LlamaCppOAICompatibleMessage[];\r\n    system?: string; // Simple system prompt like llama-cli -sys\r\n    chatTemplate?: string;\r\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\r\n  }): Promise<NativeCompletionResult>;\r\n\r\n  // Simple chat with system prompt (like llama-cli -sys \"You are a helpful assistant\")\r\n  chatWithSystem(options: {\r\n    contextId: number;\r\n    system: string;\r\n    message: string;\r\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\r\n  }): Promise<NativeCompletionResult>;\r\n\r\n  // Simple text generation (like llama-cli -p \"prompt\")\r\n  generateText(options: {\r\n    contextId: number;\r\n    prompt: string;\r\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\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',