expo-ai-kit 0.6.0 → 0.8.0

package/build/models.js

@@ -19,6 +19,7 @@ export const MODEL_REGISTRY = [
         contextWindow: 8_000,
         minRamBytes: 2_000_000_000, // 2GB — LiteRT-LM memory-maps weights, actual RSS ~1.5GB
         supportedPlatforms: ['ios', 'android'],
+        license: 'Gemma',
     },
     {
         id: 'gemma-e4b',
@@ -31,6 +32,65 @@ export const MODEL_REGISTRY = [
         contextWindow: 16_000,
         minRamBytes: 3_000_000_000, // 3GB — LiteRT-LM memory-maps weights
         supportedPlatforms: ['ios', 'android'],
+        license: 'Gemma',
+    },
+    // --- Qwen3 (Apache-2.0) — official litert-community builds. A size ladder
+    // from a sub-GB model that runs anywhere up to a 4B that rivals Gemma E4B. ---
+    {
+        id: 'qwen3-0.6b',
+        name: 'Qwen3 0.6B',
+        parameterCount: '0.6B',
+        quantization: 'mixed-int4',
+        downloadUrl: 'https://huggingface.co/litert-community/Qwen3-0.6B/resolve/main/qwen3_0_6b_mixed_int4.litertlm',
+        sha256: 'b1baab462f6be49d70eada79d715c2c52cd9ece0cad00bddf6a2c097d23498e9',
+        sizeBytes: 497_664_000, // 475MB (exact, HF LFS)
+        // The int4 build ships a 2048-token KV. TODO: benchmark on device.
+        contextWindow: 2_048,
+        minRamBytes: 1_000_000_000, // 1GB — tiny; runs on virtually any modern device
+        supportedPlatforms: ['ios', 'android'],
+        license: 'Apache-2.0',
+    },
+    {
+        id: 'qwen3-1.7b',
+        name: 'Qwen3 1.7B',
+        parameterCount: '1.7B',
+        quantization: 'dynamic-int8',
+        downloadUrl: 'https://huggingface.co/litert-community/Qwen3-1.7B/resolve/main/Qwen3_1.7B.litertlm',
+        sha256: '66064a4e9269cb693e124c4e3040bcb8a446b10bca42663896329495add3861c',
+        sizeBytes: 2_056_729_520, // 2.06GB (exact, HF LFS)
+        // Conservative default (not marked in the filename). TODO: benchmark on device.
+        contextWindow: 4_096,
+        minRamBytes: 2_000_000_000, // 2GB
+        supportedPlatforms: ['ios', 'android'],
+        license: 'Apache-2.0',
+    },
+    {
+        id: 'qwen3-4b',
+        name: 'Qwen3 4B',
+        parameterCount: '4B',
+        quantization: 'mixed-int4',
+        downloadUrl: 'https://huggingface.co/litert-community/Qwen3-4B/resolve/main/qwen3_4b_mixed_int4.litertlm',
+        sha256: 'f0794bc77efeaaf4f7af815f04c483b19b8f2ae4a102cef1b7b760a25848a18e',
+        sizeBytes: 2_659_057_664, // 2.66GB (exact, HF LFS)
+        // Conservative default (not marked in the filename). TODO: benchmark on device.
+        contextWindow: 4_096,
+        minRamBytes: 3_000_000_000, // 3GB — 4B params need more headroom than the similarly-sized E2B
+        supportedPlatforms: ['ios', 'android'],
+        license: 'Apache-2.0',
+    },
+    // --- Phi-4 Mini (MIT) — strong reasoning; q8 build, the heaviest downloadable. ---
+    {
+        id: 'phi-4-mini',
+        name: 'Phi-4 Mini',
+        parameterCount: '3.8B',
+        quantization: 'int8',
+        downloadUrl: 'https://huggingface.co/litert-community/Phi-4-mini-instruct/resolve/main/Phi-4-mini-instruct_multi-prefill-seq_q8_ekv4096.litertlm',
+        sha256: '7764d4deb53800578307be33039476b38a6c370fff71bedb3c0552563e23ab02',
+        sizeBytes: 3_910_090_752, // 3.91GB (exact, HF LFS)
+        contextWindow: 4_096, // ekv4096 build
+        minRamBytes: 4_000_000_000, // 4GB — q8 weights, heaviest downloadable
+        supportedPlatforms: ['ios', 'android'],
+        license: 'MIT',
     },
 ];
 /**

package/build/models.js.map

@@ -1 +1 @@
-{"version":3,"file":"models.js","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AA+BH,MAAM,CAAC,MAAM,cAAc,GAAyB;IAClD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,iBAAiB;QAC/B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,0CAA0C;QAC1C,0CAA0C;QAC1C,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,yDAAyD;QACrF,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;KACvC;IACD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,eAAe;QAC7B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,aAAa,EAAE,MAAM;QACrB,WAAW,EAAE,aAAa,EAAE,sCAAsC;QAClE,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;KACvC;CACF,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * Model Registry\n *\n * Defines all downloadable models known to expo-ai-kit.\n * getDownloadableModels() reads from this registry and enriches\n * each entry with on-device status from the native layer.\n */\n\nexport type ModelRegistryEntry = {\n  /** Unique model identifier used in setModel/downloadModel */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label */\n  parameterCount: string;\n  /** Quantization variant */\n  quantization: string;\n  /** URL to download the LiteRT-LM model file */\n  downloadUrl: string;\n  /** SHA256 hash for integrity verification after download */\n  sha256: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /**\n   * Practical context window (max tokens) for this model on constrained devices.\n   *\n   * These are conservative defaults, NOT the base model's theoretical max.\n   * These values should be benchmarked and adjusted during testing with\n   * real devices.\n   */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run this model */\n  minRamBytes: number;\n  /** Platforms this model can run on */\n  supportedPlatforms: ('ios' | 'android')[];\n};\n\nexport const MODEL_REGISTRY: ModelRegistryEntry[] = [\n  {\n    id: 'gemma-e2b',\n    name: 'Gemma 4 E2B',\n    parameterCount: '2.3B',\n    quantization: 'mixed-2/4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',\n    sha256: '181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c',\n    sizeBytes: 2_588_147_712, // 2.59GB (exact, HF LFS)\n    // Conservative limit for 4GB RAM devices.\n    // TODO: Benchmark during Phase 2 testing.\n    contextWindow: 8_000,\n    minRamBytes: 2_000_000_000, // 2GB — LiteRT-LM memory-maps weights, actual RSS ~1.5GB\n    supportedPlatforms: ['ios', 'android'],\n  },\n  {\n    id: 'gemma-e4b',\n    name: 'Gemma 4 E4B',\n    parameterCount: '4.5B',\n    quantization: 'mixed-4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',\n    sha256: '0b2a8980ce155fd97673d8e820b4d29d9c7d99b8fa6806f425d969b145bd52e0',\n    sizeBytes: 3_659_530_240, // 3.66GB (exact, HF LFS)\n    contextWindow: 16_000,\n    minRamBytes: 3_000_000_000, // 3GB — LiteRT-LM memory-maps weights\n    supportedPlatforms: ['ios', 'android'],\n  },\n];\n\n/**\n * Look up a model registry entry by ID.\n * Returns undefined if not found.\n */\nexport function getRegistryEntry(modelId: string): ModelRegistryEntry | undefined {\n  return MODEL_REGISTRY.find((m) => m.id === modelId);\n}\n"]}
+{"version":3,"file":"models.js","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAsCH,MAAM,CAAC,MAAM,cAAc,GAAyB;IAClD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,iBAAiB;QAC/B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,0CAA0C;QAC1C,0CAA0C;QAC1C,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,yDAAyD;QACrF,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,OAAO;KACjB;IACD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,eAAe;QAC7B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,aAAa,EAAE,MAAM;QACrB,WAAW,EAAE,aAAa,EAAE,sCAAsC;QAClE,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,OAAO;KACjB;IACD,2EAA2E;IAC3E,+EAA+E;IAC/E;QACE,EAAE,EAAE,YAAY;QAChB,IAAI,EAAE,YAAY;QAClB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,YAAY;QAC1B,WAAW,EACT,gGAAgG;QAClG,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,WAAW,EAAE,wBAAwB;QAChD,mEAAmE;QACnE,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,kDAAkD;QAC9E,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,YAAY;KACtB;IACD;QACE,EAAE,EAAE,YAAY;QAChB,IAAI,EAAE,YAAY;QAClB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,cAAc;QAC5B,WAAW,EACT,qFAAqF;QACvF,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,gFAAgF;QAChF,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,MAAM;QAClC,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,YAAY;KACtB;IACD;QACE,EAAE,EAAE,UAAU;QACd,IAAI,EAAE,UAAU;QAChB,cAAc,EAAE,IAAI;QACpB,YAAY,EAAE,YAAY;QAC1B,WAAW,EACT,4FAA4F;QAC9F,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,gFAAgF;QAChF,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,kEAAkE;QAC9F,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,YAAY;KACtB;IACD,oFAAoF;IACpF;QACE,EAAE,EAAE,YAAY;QAChB,IAAI,EAAE,YAAY;QAClB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,MAAM;QACpB,WAAW,EACT,oIAAoI;QACtI,MAAM,EAAE,kEAAkE;QAC1E,SAAS,EAAE,aAAa,EAAE,yBAAyB;QACnD,aAAa,EAAE,KAAK,EAAE,gBAAgB;QACtC,WAAW,EAAE,aAAa,EAAE,0CAA0C;QACtE,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;QACtC,OAAO,EAAE,KAAK;KACf;CACF,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * Model Registry\n *\n * Defines all downloadable models known to expo-ai-kit.\n * getDownloadableModels() reads from this registry and enriches\n * each entry with on-device status from the native layer.\n */\n\nexport type ModelRegistryEntry = {\n  /** Unique model identifier used in setModel/downloadModel */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label */\n  parameterCount: string;\n  /** Quantization variant */\n  quantization: string;\n  /** URL to download the LiteRT-LM model file */\n  downloadUrl: string;\n  /** SHA256 hash for integrity verification after download */\n  sha256: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /**\n   * Practical context window (max tokens) for this model on constrained devices.\n   *\n   * These are conservative defaults, NOT the base model's theoretical max.\n   * These values should be benchmarked and adjusted during testing with\n   * real devices.\n   */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run this model */\n  minRamBytes: number;\n  /** Platforms this model can run on */\n  supportedPlatforms: ('ios' | 'android')[];\n  /**\n   * License the model weights are distributed under — an SPDX identifier\n   * (e.g. 'Apache-2.0', 'MIT') or a family name for non-OSI terms (e.g. 'Gemma',\n   * 'Llama-3.2'). Surfaced on {@link DownloadableModel} so app developers can\n   * check their obligations before shipping a model to users.\n   */\n  license: string;\n};\n\nexport const MODEL_REGISTRY: ModelRegistryEntry[] = [\n  {\n    id: 'gemma-e2b',\n    name: 'Gemma 4 E2B',\n    parameterCount: '2.3B',\n    quantization: 'mixed-2/4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',\n    sha256: '181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c',\n    sizeBytes: 2_588_147_712, // 2.59GB (exact, HF LFS)\n    // Conservative limit for 4GB RAM devices.\n    // TODO: Benchmark during Phase 2 testing.\n    contextWindow: 8_000,\n    minRamBytes: 2_000_000_000, // 2GB — LiteRT-LM memory-maps weights, actual RSS ~1.5GB\n    supportedPlatforms: ['ios', 'android'],\n    license: 'Gemma',\n  },\n  {\n    id: 'gemma-e4b',\n    name: 'Gemma 4 E4B',\n    parameterCount: '4.5B',\n    quantization: 'mixed-4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',\n    sha256: '0b2a8980ce155fd97673d8e820b4d29d9c7d99b8fa6806f425d969b145bd52e0',\n    sizeBytes: 3_659_530_240, // 3.66GB (exact, HF LFS)\n    contextWindow: 16_000,\n    minRamBytes: 3_000_000_000, // 3GB — LiteRT-LM memory-maps weights\n    supportedPlatforms: ['ios', 'android'],\n    license: 'Gemma',\n  },\n  // --- Qwen3 (Apache-2.0) — official litert-community builds. A size ladder\n  // from a sub-GB model that runs anywhere up to a 4B that rivals Gemma E4B. ---\n  {\n    id: 'qwen3-0.6b',\n    name: 'Qwen3 0.6B',\n    parameterCount: '0.6B',\n    quantization: 'mixed-int4',\n    downloadUrl:\n      'https://huggingface.co/litert-community/Qwen3-0.6B/resolve/main/qwen3_0_6b_mixed_int4.litertlm',\n    sha256: 'b1baab462f6be49d70eada79d715c2c52cd9ece0cad00bddf6a2c097d23498e9',\n    sizeBytes: 497_664_000, // 475MB (exact, HF LFS)\n    // The int4 build ships a 2048-token KV. TODO: benchmark on device.\n    contextWindow: 2_048,\n    minRamBytes: 1_000_000_000, // 1GB — tiny; runs on virtually any modern device\n    supportedPlatforms: ['ios', 'android'],\n    license: 'Apache-2.0',\n  },\n  {\n    id: 'qwen3-1.7b',\n    name: 'Qwen3 1.7B',\n    parameterCount: '1.7B',\n    quantization: 'dynamic-int8',\n    downloadUrl:\n      'https://huggingface.co/litert-community/Qwen3-1.7B/resolve/main/Qwen3_1.7B.litertlm',\n    sha256: '66064a4e9269cb693e124c4e3040bcb8a446b10bca42663896329495add3861c',\n    sizeBytes: 2_056_729_520, // 2.06GB (exact, HF LFS)\n    // Conservative default (not marked in the filename). TODO: benchmark on device.\n    contextWindow: 4_096,\n    minRamBytes: 2_000_000_000, // 2GB\n    supportedPlatforms: ['ios', 'android'],\n    license: 'Apache-2.0',\n  },\n  {\n    id: 'qwen3-4b',\n    name: 'Qwen3 4B',\n    parameterCount: '4B',\n    quantization: 'mixed-int4',\n    downloadUrl:\n      'https://huggingface.co/litert-community/Qwen3-4B/resolve/main/qwen3_4b_mixed_int4.litertlm',\n    sha256: 'f0794bc77efeaaf4f7af815f04c483b19b8f2ae4a102cef1b7b760a25848a18e',\n    sizeBytes: 2_659_057_664, // 2.66GB (exact, HF LFS)\n    // Conservative default (not marked in the filename). TODO: benchmark on device.\n    contextWindow: 4_096,\n    minRamBytes: 3_000_000_000, // 3GB — 4B params need more headroom than the similarly-sized E2B\n    supportedPlatforms: ['ios', 'android'],\n    license: 'Apache-2.0',\n  },\n  // --- Phi-4 Mini (MIT) — strong reasoning; q8 build, the heaviest downloadable. ---\n  {\n    id: 'phi-4-mini',\n    name: 'Phi-4 Mini',\n    parameterCount: '3.8B',\n    quantization: 'int8',\n    downloadUrl:\n      'https://huggingface.co/litert-community/Phi-4-mini-instruct/resolve/main/Phi-4-mini-instruct_multi-prefill-seq_q8_ekv4096.litertlm',\n    sha256: '7764d4deb53800578307be33039476b38a6c370fff71bedb3c0552563e23ab02',\n    sizeBytes: 3_910_090_752, // 3.91GB (exact, HF LFS)\n    contextWindow: 4_096, // ekv4096 build\n    minRamBytes: 4_000_000_000, // 4GB — q8 weights, heaviest downloadable\n    supportedPlatforms: ['ios', 'android'],\n    license: 'MIT',\n  },\n];\n\n/**\n * Look up a model registry entry by ID.\n * Returns undefined if not found.\n */\nexport function getRegistryEntry(modelId: string): ModelRegistryEntry | undefined {\n  return MODEL_REGISTRY.find((m) => m.id === modelId);\n}\n"]}

package/build/tools.d.ts

@@ -0,0 +1,41 @@
+import type { JSONSchema, ToolSet } from './types';
+/** Build the instruction appended to the system prompt to enable tool calling. */
+export declare function buildToolInstruction(tools: ToolSet): string;
+/**
+ * What {@link parseToolCall} found in a model response.
+ *
+ * - `tool`: a well-formed call to a known tool (args still need schema validation).
+ * - `unknown-tool`: looked like a tool call but the name isn't in the tool set.
+ * - `text`: no tool call — treat the response as the final answer.
+ */
+export type ParsedToolCall = {
+    kind: 'tool';
+    toolName: string;
+    args: unknown;
+} | {
+    kind: 'unknown-tool';
+    toolName: string;
+} | {
+    kind: 'text';
+};
+/**
+ * Detect a tool call in model output.
+ *
+ * A response is a tool call when it contains a JSON object (possibly wrapped in
+ * prose or a ```json fence) with a string `tool` field. We tolerate `arguments`
+ * or `args` for the payload. If `tool` names something not in `toolNames` it's
+ * reported as `unknown-tool` so the loop can re-prompt instead of leaking the
+ * raw JSON as an answer. Anything else is plain `text`.
+ */
+export declare function parseToolCall(text: string, toolNames: string[]): ParsedToolCall;
+/** Follow-up prompt when the model named a tool that doesn't exist. */
+export declare function buildUnknownToolRepair(toolName: string, toolNames: string[]): string;
+/** Follow-up prompt when a tool's proposed arguments failed schema validation. */
+export declare function buildToolArgsRepair(toolName: string, errors: string[]): string;
+/**
+ * Render a tool result as the user-turn text fed back to the model.
+ * Non-string results are JSON-encoded; strings pass through as-is.
+ */
+export declare function formatToolResult(toolName: string, result: unknown): string;
+export type { JSONSchema };
+//# sourceMappingURL=tools.d.ts.map

package/build/tools.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tools.d.ts","sourceRoot":"","sources":["../src/tools.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,UAAU,EAAE,OAAO,EAAE,MAAM,SAAS,CAAC;AAgBnD,kFAAkF;AAClF,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAwB3D;AAED;;;;;;GAMG;AACH,MAAM,MAAM,cAAc,GACtB;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,OAAO,CAAA;CAAE,GACjD;IAAE,IAAI,EAAE,cAAc,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,GAC1C;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAErB;;;;;;;;GAQG;AACH,wBAAgB,aAAa,CAAC,IAAI,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,GAAG,cAAc,CAgB/E;AAED,uEAAuE;AACvE,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,GAAG,MAAM,CAMpF;AAED,kFAAkF;AAClF,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,CAQ9E;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,GAAG,MAAM,CAI1E;AAeD,YAAY,EAAE,UAAU,EAAE,CAAC"}

package/build/tools.js

@@ -0,0 +1,86 @@
+import { extractJson } from './structured';
+// ---------------------------------------------------------------------------
+// Pure helpers for generateText() tool calling.
+//
+// Like structured.ts, this module imports no native module so its logic can be
+// unit-tested in plain Node. generateText() (in index.ts) drives the inference
+// + tool-execution loop on top of these.
+//
+// On-device backends have no native tool-call channel, so we define a tiny text
+// protocol: the model emits a JSON object `{"tool": "<name>", "arguments": {…}}`
+// to request a call, or plain text to answer. We parse that back out tolerantly
+// (reusing extractJson), validate the name + args, and run the tool in JS.
+// ---------------------------------------------------------------------------
+/** Build the instruction appended to the system prompt to enable tool calling. */
+export function buildToolInstruction(tools) {
+    const names = Object.keys(tools);
+    const lines = [
+        'You have access to the following tools. Use one only when it helps answer the request:',
+        '',
+    ];
+    for (const name of names) {
+        const tool = tools[name];
+        lines.push(`- ${name}: ${tool.description}`);
+        lines.push(`  arguments JSON Schema: ${JSON.stringify(tool.parameters)}`);
+    }
+    lines.push('', 'To call a tool, respond with ONLY a JSON object of this exact form and nothing else:', '{"tool": "<tool name>", "arguments": { ... }}', '', 'Rules:', '- Call at most one tool per response.', `- "tool" must be exactly one of: ${names.join(', ')}.`, '- "arguments" must conform to that tool\'s arguments JSON Schema.', '- If you do not need a tool, answer the user directly in plain text with no JSON.', '- After you receive a tool result, use it to answer; do not repeat the same call.');
+    return lines.join('\n');
+}
+/**
+ * Detect a tool call in model output.
+ *
+ * A response is a tool call when it contains a JSON object (possibly wrapped in
+ * prose or a ```json fence) with a string `tool` field. We tolerate `arguments`
+ * or `args` for the payload. If `tool` names something not in `toolNames` it's
+ * reported as `unknown-tool` so the loop can re-prompt instead of leaking the
+ * raw JSON as an answer. Anything else is plain `text`.
+ */
+export function parseToolCall(text, toolNames) {
+    const parsed = extractJson(text);
+    if (!parsed.ok)
+        return { kind: 'text' };
+    const value = parsed.value;
+    if (!isPlainObject(value) || typeof value.tool !== 'string') {
+        return { kind: 'text' };
+    }
+    const toolName = value.tool;
+    if (!toolNames.includes(toolName)) {
+        return { kind: 'unknown-tool', toolName };
+    }
+    const rawArgs = 'arguments' in value ? value.arguments : value.args;
+    return { kind: 'tool', toolName, args: rawArgs ?? {} };
+}
+/** Follow-up prompt when the model named a tool that doesn't exist. */
+export function buildUnknownToolRepair(toolName, toolNames) {
+    return (`The tool "${toolName}" does not exist. ` +
+        `Available tools are: ${toolNames.join(', ')}. ` +
+        'Either call one of these tools using the required JSON form, or answer in plain text.');
+}
+/** Follow-up prompt when a tool's proposed arguments failed schema validation. */
+export function buildToolArgsRepair(toolName, errors) {
+    const detail = errors.slice(0, 8).join('; ');
+    return (`The arguments for "${toolName}" did not match its schema: ${detail}. ` +
+        'Respond again with ONLY the corrected {"tool": "' +
+        toolName +
+        '", "arguments": { ... }} JSON — no prose, no markdown code fences.');
+}
+/**
+ * Render a tool result as the user-turn text fed back to the model.
+ * Non-string results are JSON-encoded; strings pass through as-is.
+ */
+export function formatToolResult(toolName, result) {
+    const body = typeof result === 'string' ? result : safeStringify(result);
+    return `Result of calling the tool "${toolName}":\n${body}`;
+}
+function safeStringify(value) {
+    try {
+        return JSON.stringify(value) ?? String(value);
+    }
+    catch {
+        return String(value);
+    }
+}
+function isPlainObject(value) {
+    return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+//# sourceMappingURL=tools.js.map

package/build/tools.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tools.js","sourceRoot":"","sources":["../src/tools.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AAE3C,8EAA8E;AAC9E,gDAAgD;AAChD,EAAE;AACF,+EAA+E;AAC/E,+EAA+E;AAC/E,yCAAyC;AACzC,EAAE;AACF,gFAAgF;AAChF,iFAAiF;AACjF,gFAAgF;AAChF,2EAA2E;AAC3E,8EAA8E;AAE9E,kFAAkF;AAClF,MAAM,UAAU,oBAAoB,CAAC,KAAc;IACjD,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IACjC,MAAM,KAAK,GAAa;QACtB,wFAAwF;QACxF,EAAE;KACH,CAAC;IACF,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC;QACzB,KAAK,CAAC,IAAI,CAAC,KAAK,IAAI,KAAK,IAAI,CAAC,WAAW,EAAE,CAAC,CAAC;QAC7C,KAAK,CAAC,IAAI,CAAC,4BAA4B,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC,CAAC;IAC5E,CAAC;IACD,KAAK,CAAC,IAAI,CACR,EAAE,EACF,sFAAsF,EACtF,+CAA+C,EAC/C,EAAE,EACF,QAAQ,EACR,uCAAuC,EACvC,oCAAoC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EACvD,mEAAmE,EACnE,mFAAmF,EACnF,mFAAmF,CACpF,CAAC;IACF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC;AAcD;;;;;;;;GAQG;AACH,MAAM,UAAU,aAAa,CAAC,IAAY,EAAE,SAAmB;IAC7D,MAAM,MAAM,GAAG,WAAW,CAAC,IAAI,CAAC,CAAC;IACjC,IAAI,CAAC,MAAM,CAAC,EAAE;QAAE,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;IAExC,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC;IAC3B,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC5D,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;IAC1B,CAAC;IAED,MAAM,QAAQ,GAAG,KAAK,CAAC,IAAI,CAAC;IAC5B,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;QAClC,OAAO,EAAE,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,CAAC;IAC5C,CAAC;IAED,MAAM,OAAO,GAAG,WAAW,IAAI,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC;IACpE,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,IAAI,EAAE,OAAO,IAAI,EAAE,EAAE,CAAC;AACzD,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,sBAAsB,CAAC,QAAgB,EAAE,SAAmB;IAC1E,OAAO,CACL,aAAa,QAAQ,oBAAoB;QACzC,wBAAwB,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI;QAChD,uFAAuF,CACxF,CAAC;AACJ,CAAC;AAED,kFAAkF;AAClF,MAAM,UAAU,mBAAmB,CAAC,QAAgB,EAAE,MAAgB;IACpE,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7C,OAAO,CACL,sBAAsB,QAAQ,+BAA+B,MAAM,IAAI;QACvE,kDAAkD;QAClD,QAAQ;QACR,oEAAoE,CACrE,CAAC;AACJ,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,QAAgB,EAAE,MAAe;IAChE,MAAM,IAAI,GACR,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;IAC9D,OAAO,+BAA+B,QAAQ,OAAO,IAAI,EAAE,CAAC;AAC9D,CAAC;AAED,SAAS,aAAa,CAAC,KAAc;IACnC,IAAI,CAAC;QACH,OAAO,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,CAAC;IAChD,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC;IACvB,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,KAAc;IACnC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC9E,CAAC","sourcesContent":["import type { JSONSchema, ToolSet } from './types';\nimport { extractJson } from './structured';\n\n// ---------------------------------------------------------------------------\n// Pure helpers for generateText() tool calling.\n//\n// Like structured.ts, this module imports no native module so its logic can be\n// unit-tested in plain Node. generateText() (in index.ts) drives the inference\n// + tool-execution loop on top of these.\n//\n// On-device backends have no native tool-call channel, so we define a tiny text\n// protocol: the model emits a JSON object `{\"tool\": \"<name>\", \"arguments\": {…}}`\n// to request a call, or plain text to answer. We parse that back out tolerantly\n// (reusing extractJson), validate the name + args, and run the tool in JS.\n// ---------------------------------------------------------------------------\n\n/** Build the instruction appended to the system prompt to enable tool calling. */\nexport function buildToolInstruction(tools: ToolSet): string {\n  const names = Object.keys(tools);\n  const lines: string[] = [\n    'You have access to the following tools. Use one only when it helps answer the request:',\n    '',\n  ];\n  for (const name of names) {\n    const tool = tools[name];\n    lines.push(`- ${name}: ${tool.description}`);\n    lines.push(`  arguments JSON Schema: ${JSON.stringify(tool.parameters)}`);\n  }\n  lines.push(\n    '',\n    'To call a tool, respond with ONLY a JSON object of this exact form and nothing else:',\n    '{\"tool\": \"<tool name>\", \"arguments\": { ... }}',\n    '',\n    'Rules:',\n    '- Call at most one tool per response.',\n    `- \"tool\" must be exactly one of: ${names.join(', ')}.`,\n    '- \"arguments\" must conform to that tool\\'s arguments JSON Schema.',\n    '- If you do not need a tool, answer the user directly in plain text with no JSON.',\n    '- After you receive a tool result, use it to answer; do not repeat the same call.'\n  );\n  return lines.join('\\n');\n}\n\n/**\n * What {@link parseToolCall} found in a model response.\n *\n * - `tool`: a well-formed call to a known tool (args still need schema validation).\n * - `unknown-tool`: looked like a tool call but the name isn't in the tool set.\n * - `text`: no tool call — treat the response as the final answer.\n */\nexport type ParsedToolCall =\n  | { kind: 'tool'; toolName: string; args: unknown }\n  | { kind: 'unknown-tool'; toolName: string }\n  | { kind: 'text' };\n\n/**\n * Detect a tool call in model output.\n *\n * A response is a tool call when it contains a JSON object (possibly wrapped in\n * prose or a ```json fence) with a string `tool` field. We tolerate `arguments`\n * or `args` for the payload. If `tool` names something not in `toolNames` it's\n * reported as `unknown-tool` so the loop can re-prompt instead of leaking the\n * raw JSON as an answer. Anything else is plain `text`.\n */\nexport function parseToolCall(text: string, toolNames: string[]): ParsedToolCall {\n  const parsed = extractJson(text);\n  if (!parsed.ok) return { kind: 'text' };\n\n  const value = parsed.value;\n  if (!isPlainObject(value) || typeof value.tool !== 'string') {\n    return { kind: 'text' };\n  }\n\n  const toolName = value.tool;\n  if (!toolNames.includes(toolName)) {\n    return { kind: 'unknown-tool', toolName };\n  }\n\n  const rawArgs = 'arguments' in value ? value.arguments : value.args;\n  return { kind: 'tool', toolName, args: rawArgs ?? {} };\n}\n\n/** Follow-up prompt when the model named a tool that doesn't exist. */\nexport function buildUnknownToolRepair(toolName: string, toolNames: string[]): string {\n  return (\n    `The tool \"${toolName}\" does not exist. ` +\n    `Available tools are: ${toolNames.join(', ')}. ` +\n    'Either call one of these tools using the required JSON form, or answer in plain text.'\n  );\n}\n\n/** Follow-up prompt when a tool's proposed arguments failed schema validation. */\nexport function buildToolArgsRepair(toolName: string, errors: string[]): string {\n  const detail = errors.slice(0, 8).join('; ');\n  return (\n    `The arguments for \"${toolName}\" did not match its schema: ${detail}. ` +\n    'Respond again with ONLY the corrected {\"tool\": \"' +\n    toolName +\n    '\", \"arguments\": { ... }} JSON — no prose, no markdown code fences.'\n  );\n}\n\n/**\n * Render a tool result as the user-turn text fed back to the model.\n * Non-string results are JSON-encoded; strings pass through as-is.\n */\nexport function formatToolResult(toolName: string, result: unknown): string {\n  const body =\n    typeof result === 'string' ? result : safeStringify(result);\n  return `Result of calling the tool \"${toolName}\":\\n${body}`;\n}\n\nfunction safeStringify(value: unknown): string {\n  try {\n    return JSON.stringify(value) ?? String(value);\n  } catch {\n    return String(value);\n  }\n}\n\nfunction isPlainObject(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value);\n}\n\n// Re-export for callers that want the schema type alongside tool helpers.\nexport type { JSONSchema };\n"]}

package/build/types.d.ts

@@ -191,6 +191,117 @@ export type GenerateObjectResult<T = unknown> = {
     /** The raw model output that produced `object` (useful for debugging). */
     text: string;
 };
+/**
+ * A tool (function) the model may call to fetch data or take an action.
+ *
+ * The model never runs anything itself — it *proposes* a call (a name + JSON
+ * arguments), {@link generateText} validates the arguments against `parameters`,
+ * and only then invokes your `execute`. The result is fed back into the
+ * conversation so the model can use it to produce its final answer.
+ *
+ * @typeParam TArgs - Shape of the validated arguments passed to `execute`.
+ * @typeParam TResult - What `execute` returns (serialized back to the model).
+ */
+export type Tool<TArgs = any, TResult = any> = {
+    /**
+     * What the tool does and when to use it. This is how the model decides
+     * whether a request matches this tool — make it specific and action-oriented.
+     */
+    description: string;
+    /**
+     * JSON Schema for the tool's arguments. The model is told to conform to it,
+     * and the args it proposes are validated against it (same pragmatic subset as
+     * {@link generateObject}) before `execute` runs. Keep it small and shallow.
+     */
+    parameters: JSONSchema;
+    /**
+     * Runs the tool with the validated arguments and returns a result.
+     *
+     * **Optional on purpose.** If you omit it, {@link generateText} does not run
+     * anything — it stops with `finishReason: 'tool-calls'` and hands you the
+     * proposed call so you can confirm, gate, or execute it yourself.
+     */
+    execute?: (args: TArgs) => TResult | Promise<TResult>;
+};
+/** A map of tool name → {@link Tool}, passed to {@link generateText}. */
+export type ToolSet = Record<string, Tool>;
+/** A tool invocation the model proposed (name + validated arguments). */
+export type ToolCall = {
+    /** The tool's key in the {@link ToolSet}. */
+    toolName: string;
+    /** Arguments the model supplied, validated against the tool's `parameters`. */
+    args: unknown;
+};
+/** The outcome of running a {@link ToolCall} via its `execute`. */
+export type ToolResult = {
+    /** The tool's key in the {@link ToolSet}. */
+    toolName: string;
+    /** The arguments that were passed to `execute`. */
+    args: unknown;
+    /** Whatever `execute` returned (or `{ error }` if it threw). */
+    result: unknown;
+};
+/** One model round-trip in the {@link generateText} loop. */
+export type StepResult = {
+    /** Assistant text produced this step (empty when the step only called a tool). */
+    text: string;
+    /** Tool calls proposed this step (at most one in the current protocol). */
+    toolCalls: ToolCall[];
+    /** Results of the tool calls executed this step. */
+    toolResults: ToolResult[];
+};
+/**
+ * Why {@link generateText} stopped.
+ *
+ * - `'stop'`: the model produced a final text answer.
+ * - `'tool-calls'`: stopped because a proposed tool has no `execute` — the call
+ *   is returned for you to handle (human-in-the-loop).
+ * - `'max-steps'`: hit the `maxSteps` cap while still calling tools. Raise the cap.
+ */
+export type GenerateTextFinishReason = 'stop' | 'tool-calls' | 'max-steps';
+/**
+ * Options for {@link generateText}.
+ */
+export type GenerateTextOptions = {
+    /** Tools the model may call. Omit (or pass `{}`) for a plain text generation. */
+    tools?: ToolSet;
+    /**
+     * Maximum number of model round-trips (each call + tool execution is one step).
+     * Bounds the tool-calling chain so it can't run away. Defaults to 5.
+     */
+    maxSteps?: number;
+    /**
+     * System prompt used when the messages array has no system message. The tool
+     * instructions are appended to it (or to the array's system message if present).
+     */
+    systemPrompt?: string;
+    /**
+     * Abort the request. Behaves like {@link LLMSendOptions.signal} — the returned
+     * promise rejects with an INFERENCE_CANCELLED {@link ModelError}.
+     */
+    signal?: AbortSignal;
+    /**
+     * How many times to re-prompt within a step when the model emits a malformed
+     * tool call, an unknown tool name, or arguments that fail schema validation.
+     * Defaults to 2. If it still can't comply, `generateText` throws INFERENCE_FAILED.
+     */
+    maxRepairAttempts?: number;
+};
+/**
+ * Result of {@link generateText}.
+ */
+export type GenerateTextResult = {
+    /** The final assistant text (empty if it stopped on a tool call without `execute`). */
+    text: string;
+    /** Every step taken, in order — useful for tracing or debugging. */
+    steps: StepResult[];
+    /** All tool calls across every step, flattened. */
+    toolCalls: ToolCall[];
+    /** All tool results across every step, flattened. */
+    toolResults: ToolResult[];
+    /** Why generation stopped. See {@link GenerateTextFinishReason}. */
+    finishReason: GenerateTextFinishReason;
+};
 /**
  * A built-in model provided by the OS (e.g. Apple Foundation Models, ML Kit).
  * These are always available on supported devices -- no download needed.
@@ -218,6 +329,8 @@ export type DownloadableModel = {
     name: string;
     /** Parameter count label (e.g. '2.3B') */
     parameterCount: string;
+    /** License the weights are distributed under (e.g. 'Apache-2.0', 'MIT', 'Gemma'). */
+    license: string;
     /** Download file size in bytes */
     sizeBytes: number;
     /** Maximum context window in tokens */

package/build/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,MAAM,MAAM,gBAAgB,GAAG,MAAM,GAAG,KAAK,GAAG,KAAK,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,MAAM,MAAM,gBAAgB,GAAG;IAC7B,2EAA2E;IAC3E,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,uEAAuE;IACvE,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,oEAAoE;IACpE,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,uDAAuD;IACvD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,eAAe,GAAG;IAC5B,iEAAiE;IACjE,OAAO,CAAC,EAAE,gBAAgB,CAAC;IAC3B;;;;OAIG;IACH,UAAU,CAAC,EAAE,gBAAgB,CAAC;CAC/B,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,OAAO,GAAG,QAAQ,GAAG,MAAM,GAAG,WAAW,CAAC;AAEtD;;GAEG;AACH,MAAM,MAAM,UAAU,GAAG;IACvB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;CACjB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,cAAc,GAAG;IAC3B;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;;;;;;;OASG;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;CACtB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,WAAW,GAAG;IACxB,kCAAkC;IAClC,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,gBAAgB,GAAG;IAC7B;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,eAAe,GAAG;IAC5B,2EAA2E;IAC3E,OAAO,EAAE,OAAO,CAAC,WAAW,CAAC,CAAC;IAC9B,2EAA2E;IAC3E,IAAI,EAAE,MAAM,IAAI,CAAC;CAClB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,cAAc,GAAG;IAC3B,mDAAmD;IACnD,SAAS,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,KAAK,EAAE,MAAM,CAAC;IACd,8BAA8B;IAC9B,eAAe,EAAE,MAAM,CAAC;IACxB,sCAAsC;IACtC,MAAM,EAAE,OAAO,CAAC;CACjB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,iBAAiB,GAAG,CAAC,KAAK,EAAE,cAAc,KAAK,IAAI,CAAC;AAOhE;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,QAAQ,GACR,OAAO,GACP,QAAQ,GACR,QAAQ,GACR,SAAS,GACT,SAAS,GACT,MAAM,CAAC;AAEX;;;;;;;;;GASG;AACH,MAAM,MAAM,UAAU,GAAG;IACvB,gDAAgD;IAChD,IAAI,CAAC,EAAE,cAAc,GAAG,cAAc,EAAE,CAAC;IACzC,uDAAuD;IACvD,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,+DAA+D;IAC/D,UAAU,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,CAAC,CAAC;IACxC,iEAAiE;IACjE,QAAQ,CAAC,EAAE,MAAM,EAAE,CAAC;IACpB,iEAAiE;IACjE,KAAK,CAAC,EAAE,UAAU,CAAC;IACnB,kDAAkD;IAClD,IAAI,CAAC,EAAE,aAAa,CAAC,MAAM,GAAG,MAAM,GAAG,OAAO,GAAG,IAAI,CAAC,CAAC;IACvD,0EAA0E;IAC1E,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;CACxB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,qBAAqB,GAAG;IAClC;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,oBAAoB,CAAC,CAAC,GAAG,OAAO,IAAI;IAC9C,sDAAsD;IACtD,MAAM,EAAE,CAAC,CAAC;IACV,0EAA0E;IAC1E,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAOF;;;GAGG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,yDAAyD;IACzD,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,+DAA+D;IAC/D,SAAS,EAAE,OAAO,CAAC;IACnB,6CAA6C;IAC7C,QAAQ,EAAE,KAAK,GAAG,SAAS,CAAC;IAC5B,uCAAuC;IACvC,aAAa,EAAE,MAAM,CAAC;CACvB,CAAC;AAEF;;;GAGG;AACH,MAAM,MAAM,iBAAiB,GAAG;IAC9B,8DAA8D;IAC9D,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,0CAA0C;IAC1C,cAAc,EAAE,MAAM,CAAC;IACvB,kCAAkC;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,aAAa,EAAE,MAAM,CAAC;IACtB,kDAAkD;IAClD,WAAW,EAAE,MAAM,CAAC;IACpB,oEAAoE;IACpE,iBAAiB,EAAE,OAAO,CAAC;IAC3B,+BAA+B;IAC/B,MAAM,EAAE,uBAAuB,CAAC;CACjC,CAAC;AAEF;;;;;;;;;;GAUG;AACH,MAAM,MAAM,uBAAuB,GAC/B,gBAAgB,GAChB,aAAa,GACb,YAAY,GACZ,SAAS,GACT,OAAO,CAAC;AAEZ;;GAEG;AACH,MAAM,MAAM,cAAc,GACtB,iBAAiB,GACjB,sBAAsB,GACtB,iBAAiB,GACjB,kBAAkB,GAClB,uBAAuB,GACvB,oBAAoB,GACpB,eAAe,GACf,kBAAkB,GAClB,gBAAgB,GAChB,qBAAqB,GACrB,mBAAmB,GACnB,sBAAsB,GACtB,SAAS,CAAC;AAEd;;GAEG;AACH,qBAAa,UAAW,SAAQ,KAAK;IACnC,IAAI,EAAE,cAAc,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;gBAEJ,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM;CAMpE;AAED;;GAEG;AACH,MAAM,MAAM,0BAA0B,GAAG;IACvC,6BAA6B;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,QAAQ,EAAE,MAAM,CAAC;CAClB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,qBAAqB,GAAG;IAClC,gCAAgC;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,iBAAiB;IACjB,MAAM,EAAE,uBAAuB,CAAC;CACjC,CAAC"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,MAAM,MAAM,gBAAgB,GAAG,MAAM,GAAG,KAAK,GAAG,KAAK,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,MAAM,MAAM,gBAAgB,GAAG;IAC7B,2EAA2E;IAC3E,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,uEAAuE;IACvE,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,oEAAoE;IACpE,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,uDAAuD;IACvD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,eAAe,GAAG;IAC5B,iEAAiE;IACjE,OAAO,CAAC,EAAE,gBAAgB,CAAC;IAC3B;;;;OAIG;IACH,UAAU,CAAC,EAAE,gBAAgB,CAAC;CAC/B,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,OAAO,GAAG,QAAQ,GAAG,MAAM,GAAG,WAAW,CAAC;AAEtD;;GAEG;AACH,MAAM,MAAM,UAAU,GAAG;IACvB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;CACjB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,cAAc,GAAG;IAC3B;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;;;;;;;OASG;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;CACtB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,WAAW,GAAG;IACxB,kCAAkC;IAClC,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,gBAAgB,GAAG;IAC7B;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,eAAe,GAAG;IAC5B,2EAA2E;IAC3E,OAAO,EAAE,OAAO,CAAC,WAAW,CAAC,CAAC;IAC9B,2EAA2E;IAC3E,IAAI,EAAE,MAAM,IAAI,CAAC;CAClB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,cAAc,GAAG;IAC3B,mDAAmD;IACnD,SAAS,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,KAAK,EAAE,MAAM,CAAC;IACd,8BAA8B;IAC9B,eAAe,EAAE,MAAM,CAAC;IACxB,sCAAsC;IACtC,MAAM,EAAE,OAAO,CAAC;CACjB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,iBAAiB,GAAG,CAAC,KAAK,EAAE,cAAc,KAAK,IAAI,CAAC;AAOhE;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,QAAQ,GACR,OAAO,GACP,QAAQ,GACR,QAAQ,GACR,SAAS,GACT,SAAS,GACT,MAAM,CAAC;AAEX;;;;;;;;;GASG;AACH,MAAM,MAAM,UAAU,GAAG;IACvB,gDAAgD;IAChD,IAAI,CAAC,EAAE,cAAc,GAAG,cAAc,EAAE,CAAC;IACzC,uDAAuD;IACvD,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,+DAA+D;IAC/D,UAAU,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,CAAC,CAAC;IACxC,iEAAiE;IACjE,QAAQ,CAAC,EAAE,MAAM,EAAE,CAAC;IACpB,iEAAiE;IACjE,KAAK,CAAC,EAAE,UAAU,CAAC;IACnB,kDAAkD;IAClD,IAAI,CAAC,EAAE,aAAa,CAAC,MAAM,GAAG,MAAM,GAAG,OAAO,GAAG,IAAI,CAAC,CAAC;IACvD,0EAA0E;IAC1E,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;CACxB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,qBAAqB,GAAG;IAClC;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,oBAAoB,CAAC,CAAC,GAAG,OAAO,IAAI;IAC9C,sDAAsD;IACtD,MAAM,EAAE,CAAC,CAAC;IACV,0EAA0E;IAC1E,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAOF;;;;;;;;;;GAUG;AACH,MAAM,MAAM,IAAI,CAAC,KAAK,GAAG,GAAG,EAAE,OAAO,GAAG,GAAG,IAAI;IAC7C;;;OAGG;IACH,WAAW,EAAE,MAAM,CAAC;IACpB;;;;OAIG;IACH,UAAU,EAAE,UAAU,CAAC;IACvB;;;;;;OAMG;IACH,OAAO,CAAC,EAAE,CAAC,IAAI,EAAE,KAAK,KAAK,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;CACvD,CAAC;AAEF,yEAAyE;AACzE,MAAM,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;AAE3C,yEAAyE;AACzE,MAAM,MAAM,QAAQ,GAAG;IACrB,6CAA6C;IAC7C,QAAQ,EAAE,MAAM,CAAC;IACjB,+EAA+E;IAC/E,IAAI,EAAE,OAAO,CAAC;CACf,CAAC;AAEF,mEAAmE;AACnE,MAAM,MAAM,UAAU,GAAG;IACvB,6CAA6C;IAC7C,QAAQ,EAAE,MAAM,CAAC;IACjB,mDAAmD;IACnD,IAAI,EAAE,OAAO,CAAC;IACd,gEAAgE;IAChE,MAAM,EAAE,OAAO,CAAC;CACjB,CAAC;AAEF,6DAA6D;AAC7D,MAAM,MAAM,UAAU,GAAG;IACvB,kFAAkF;IAClF,IAAI,EAAE,MAAM,CAAC;IACb,2EAA2E;IAC3E,SAAS,EAAE,QAAQ,EAAE,CAAC;IACtB,oDAAoD;IACpD,WAAW,EAAE,UAAU,EAAE,CAAC;CAC3B,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,MAAM,wBAAwB,GAAG,MAAM,GAAG,YAAY,GAAG,WAAW,CAAC;AAE3E;;GAEG;AACH,MAAM,MAAM,mBAAmB,GAAG;IAChC,iFAAiF;IACjF,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB;;;OAGG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,kBAAkB,GAAG;IAC/B,uFAAuF;IACvF,IAAI,EAAE,MAAM,CAAC;IACb,oEAAoE;IACpE,KAAK,EAAE,UAAU,EAAE,CAAC;IACpB,mDAAmD;IACnD,SAAS,EAAE,QAAQ,EAAE,CAAC;IACtB,qDAAqD;IACrD,WAAW,EAAE,UAAU,EAAE,CAAC;IAC1B,oEAAoE;IACpE,YAAY,EAAE,wBAAwB,CAAC;CACxC,CAAC;AAOF;;;GAGG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,yDAAyD;IACzD,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,+DAA+D;IAC/D,SAAS,EAAE,OAAO,CAAC;IACnB,6CAA6C;IAC7C,QAAQ,EAAE,KAAK,GAAG,SAAS,CAAC;IAC5B,uCAAuC;IACvC,aAAa,EAAE,MAAM,CAAC;CACvB,CAAC;AAEF;;;GAGG;AACH,MAAM,MAAM,iBAAiB,GAAG;IAC9B,8DAA8D;IAC9D,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,0CAA0C;IAC1C,cAAc,EAAE,MAAM,CAAC;IACvB,qFAAqF;IACrF,OAAO,EAAE,MAAM,CAAC;IAChB,kCAAkC;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,aAAa,EAAE,MAAM,CAAC;IACtB,kDAAkD;IAClD,WAAW,EAAE,MAAM,CAAC;IACpB,oEAAoE;IACpE,iBAAiB,EAAE,OAAO,CAAC;IAC3B,+BAA+B;IAC/B,MAAM,EAAE,uBAAuB,CAAC;CACjC,CAAC;AAEF;;;;;;;;;;GAUG;AACH,MAAM,MAAM,uBAAuB,GAC/B,gBAAgB,GAChB,aAAa,GACb,YAAY,GACZ,SAAS,GACT,OAAO,CAAC;AAEZ;;GAEG;AACH,MAAM,MAAM,cAAc,GACtB,iBAAiB,GACjB,sBAAsB,GACtB,iBAAiB,GACjB,kBAAkB,GAClB,uBAAuB,GACvB,oBAAoB,GACpB,eAAe,GACf,kBAAkB,GAClB,gBAAgB,GAChB,qBAAqB,GACrB,mBAAmB,GACnB,sBAAsB,GACtB,SAAS,CAAC;AAEd;;GAEG;AACH,qBAAa,UAAW,SAAQ,KAAK;IACnC,IAAI,EAAE,cAAc,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;gBAEJ,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM;CAMpE;AAED;;GAEG;AACH,MAAM,MAAM,0BAA0B,GAAG;IACvC,6BAA6B;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,QAAQ,EAAE,MAAM,CAAC;CAClB,CAAC;AAEF;;GAEG;AACH,MAAM,MAAM,qBAAqB,GAAG;IAClC,gCAAgC;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,iBAAiB;IACjB,MAAM,EAAE,uBAAuB,CAAC;CACjC,CAAC"}

package/build/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AA6SA;;GAEG;AACH,MAAM,OAAO,UAAW,SAAQ,KAAK;IACnC,IAAI,CAAiB;IACrB,OAAO,CAAS;IAEhB,YAAY,IAAoB,EAAE,OAAe,EAAE,OAAgB;QACjE,KAAK,CAAC,OAAO,IAAI,GAAG,IAAI,KAAK,OAAO,EAAE,CAAC,CAAC;QACxC,IAAI,CAAC,IAAI,GAAG,YAAY,CAAC;QACzB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;IACzB,CAAC;CACF","sourcesContent":["/**\n * Hardware backend for on-device model inference.\n *\n * - 'auto': Try GPU first, fall back to CPU (default)\n * - 'gpu': Force GPU — faster (~40-50 tok/s) but needs more memory\n * - 'cpu': Force CPU — slower (~2-5 tok/s) but works on low-RAM devices\n */\nexport type InferenceBackend = 'auto' | 'gpu' | 'cpu';\n\n/**\n * Sampling / generation parameters applied to a model session.\n *\n * Support is per-backend (on-device runtimes expose different knobs), so these\n * are best-effort — unsupported fields are ignored rather than erroring:\n *\n * | field       | Gemma (LiteRT-LM) | Apple Foundation Models | ML Kit |\n * |-------------|:-----------------:|:-----------------------:|:------:|\n * | temperature | ✓                 | ✓                       | —      |\n * | topK        | ✓                 | —                       | —      |\n * | topP        | ✓                 | —                       | —      |\n * | seed        | ✓ (iOS only)      | —                       | —      |\n * | maxTokens   | —                 | ✓ (max output)          | —      |\n *\n * Notes:\n * - Gemma/LiteRT-LM has no per-generation output-token cap (its `maxNumTokens`\n *   is the total KV-cache size, set at load), so `maxTokens` is not honored\n *   there. Its sampler (topK/topP/temperature[/seed]) is fixed at conversation\n *   creation, which is why generation config lives on setModel() rather than\n *   per-call. `seed` is currently wired on iOS only.\n * - The ML Kit built-in (Android default) does not yet apply generation config;\n *   it uses its own defaults.\n */\nexport type GenerationConfig = {\n  /** Sampling temperature. Lower = more deterministic. Typically 0.0–2.0. */\n  temperature?: number;\n  /** Nucleus sampling: number of top logits to consider. Must be > 0. */\n  topK?: number;\n  /** Nucleus sampling: cumulative probability threshold in [0, 1]. */\n  topP?: number;\n  /** RNG seed for reproducible sampling (Gemma only). */\n  seed?: number;\n  /** Maximum number of output tokens to generate (Apple FM / ML Kit only). */\n  maxTokens?: number;\n};\n\n/**\n * Options for setModel.\n */\nexport type SetModelOptions = {\n  /** Hardware backend to use for inference. Defaults to 'auto'. */\n  backend?: InferenceBackend;\n  /**\n   * Default sampling parameters for this model session. Applied when the model\n   * is activated and used for all subsequent sendMessage/streamMessage calls\n   * until the next setModel(). See {@link GenerationConfig} for per-backend support.\n   */\n  generation?: GenerationConfig;\n};\n\n/**\n * Role in a conversation message.\n */\nexport type LLMRole = 'system' | 'user' | 'assistant';\n\n/**\n * A single message in a conversation.\n */\nexport type LLMMessage = {\n  role: LLMRole;\n  content: string;\n};\n\n/**\n * Options for sendMessage.\n */\nexport type LLMSendOptions = {\n  /**\n   * Default system prompt to use if no system message is provided in the messages array.\n   * If a system message exists in the array, this is ignored.\n   */\n  systemPrompt?: string;\n  /**\n   * Abort the request. When the signal fires, the returned promise rejects with\n   * an INFERENCE_CANCELLED {@link ModelError}.\n   *\n   * Note: on-device, non-streaming generation cannot always be interrupted\n   * mid-decode — abort always unblocks the caller, but the model may keep\n   * computing in the background until it finishes, during which a new\n   * sendMessage/streamMessage will throw INFERENCE_BUSY. To truly interrupt a\n   * long generation, prefer streamMessage().stop().\n   */\n  signal?: AbortSignal;\n};\n\n/**\n * Response from sendMessage.\n */\nexport type LLMResponse = {\n  /** The generated response text */\n  text: string;\n};\n\n/**\n * Options for streamMessage.\n */\nexport type LLMStreamOptions = {\n  /**\n   * Default system prompt to use if no system message is provided in the messages array.\n   * If a system message exists in the array, this is ignored.\n   */\n  systemPrompt?: string;\n};\n\n/**\n * Handle returned by streamMessage.\n */\nexport type LLMStreamHandle = {\n  /** Resolves with the final text when streaming completes or is stopped. */\n  promise: Promise<LLMResponse>;\n  /** Stop streaming. Resolves `promise` with the text accumulated so far. */\n  stop: () => void;\n};\n\n/**\n * Event payload for streaming tokens.\n */\nexport type LLMStreamEvent = {\n  /** Unique identifier for this streaming session */\n  sessionId: string;\n  /** The token/chunk of text received */\n  token: string;\n  /** Accumulated text so far */\n  accumulatedText: string;\n  /** Whether this is the final chunk */\n  isDone: boolean;\n};\n\n/**\n * Callback function for streaming events.\n */\nexport type LLMStreamCallback = (event: LLMStreamEvent) => void;\n\n\n// ============================================================================\n// Structured Output (generateObject)\n// ============================================================================\n\n/**\n * The set of JSON Schema primitive `type` values understood by\n * {@link generateObject}'s local validator.\n */\nexport type JSONSchemaType =\n  | 'object'\n  | 'array'\n  | 'string'\n  | 'number'\n  | 'integer'\n  | 'boolean'\n  | 'null';\n\n/**\n * A JSON Schema describing the shape you want {@link generateObject} to return.\n *\n * A pragmatic subset is enforced locally — `type`, `properties`, `required`,\n * `items`, and `enum` — which covers most extraction shapes. Any other JSON\n * Schema keywords you include (e.g. `description`, `minLength`) are still sent\n * to the model to guide it, but are not validated on-device. Keep schemas small:\n * on-device models follow flat, shallow shapes far more reliably than deeply\n * nested ones.\n */\nexport type JSONSchema = {\n  /** Expected JSON type (or a union of types). */\n  type?: JSONSchemaType | JSONSchemaType[];\n  /** Human-readable hint passed through to the model. */\n  description?: string;\n  /** For `object` schemas: the schema of each named property. */\n  properties?: Record<string, JSONSchema>;\n  /** For `object` schemas: property names that must be present. */\n  required?: string[];\n  /** For `array` schemas: the schema each element must satisfy. */\n  items?: JSONSchema;\n  /** Restrict the value to this set of literals. */\n  enum?: ReadonlyArray<string | number | boolean | null>;\n  /** Other JSON Schema keywords are accepted and forwarded to the model. */\n  [key: string]: unknown;\n};\n\n/**\n * Options for {@link generateObject}.\n */\nexport type GenerateObjectOptions = {\n  /**\n   * System prompt used when the messages array has no system message. Defaults\n   * to a structured-output-oriented instruction. If a system message is present\n   * in the array, this is ignored (the schema instruction is appended to it).\n   */\n  systemPrompt?: string;\n  /**\n   * Abort the request. Behaves like {@link LLMSendOptions.signal} — the returned\n   * promise rejects with an INFERENCE_CANCELLED {@link ModelError}.\n   */\n  signal?: AbortSignal;\n  /**\n   * How many times to re-prompt the model when its output is not valid JSON or\n   * does not match the schema. Each repair feeds the error back to the model.\n   * Defaults to 2 (i.e. up to 3 generations total).\n   */\n  maxRepairAttempts?: number;\n};\n\n/**\n * Result of {@link generateObject}.\n */\nexport type GenerateObjectResult<T = unknown> = {\n  /** The parsed value, validated against the schema. */\n  object: T;\n  /** The raw model output that produced `object` (useful for debugging). */\n  text: string;\n};\n\n\n// ============================================================================\n// Model Types\n// ============================================================================\n\n/**\n * A built-in model provided by the OS (e.g. Apple Foundation Models, ML Kit).\n * These are always available on supported devices -- no download needed.\n */\nexport type BuiltInModel = {\n  /** Unique model identifier (e.g. 'apple-fm', 'mlkit') */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Whether this model is available on the current device/OS */\n  available: boolean;\n  /** Platform this model is associated with */\n  platform: 'ios' | 'android';\n  /** Maximum context window in tokens */\n  contextWindow: number;\n};\n\n/**\n * A downloadable model that the user manages (download, load, delete).\n * These require explicit download before use.\n */\nexport type DownloadableModel = {\n  /** Unique model identifier (e.g. 'gemma-e2b', 'gemma-e4b') */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label (e.g. '2.3B') */\n  parameterCount: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /** Maximum context window in tokens */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run */\n  minRamBytes: number;\n  /** Whether this device meets the model's minimum RAM requirement */\n  meetsRequirements: boolean;\n  /** Current lifecycle status */\n  status: DownloadableModelStatus;\n};\n\n/**\n * Lifecycle status of a downloadable model.\n *\n * - 'not-downloaded': Model file is not on disk\n * - 'downloading': Model file is being downloaded\n * - 'downloaded': Model file is on disk but not loaded into memory. Call\n *   setModel() to load it. This survives app restarts, so use it to decide\n *   whether a (re-)download is needed.\n * - 'loading': File is on disk, model is being loaded into memory for inference\n * - 'ready': Model is loaded in memory and ready for inference\n */\nexport type DownloadableModelStatus =\n  | 'not-downloaded'\n  | 'downloading'\n  | 'downloaded'\n  | 'loading'\n  | 'ready';\n\n/**\n * Error codes for model-related operations.\n */\nexport type ModelErrorCode =\n  | 'MODEL_NOT_FOUND'\n  | 'MODEL_NOT_DOWNLOADED'\n  | 'DOWNLOAD_FAILED'\n  | 'DOWNLOAD_CORRUPT'\n  | 'DOWNLOAD_STORAGE_FULL'\n  | 'DOWNLOAD_CANCELLED'\n  | 'INFERENCE_OOM'\n  | 'INFERENCE_FAILED'\n  | 'INFERENCE_BUSY'\n  | 'INFERENCE_CANCELLED'\n  | 'MODEL_LOAD_FAILED'\n  | 'DEVICE_NOT_SUPPORTED'\n  | 'UNKNOWN';\n\n/**\n * Structured error for model operations.\n */\nexport class ModelError extends Error {\n  code: ModelErrorCode;\n  modelId: string;\n\n  constructor(code: ModelErrorCode, modelId: string, message?: string) {\n    super(message ?? `${code}: ${modelId}`);\n    this.name = 'ModelError';\n    this.code = code;\n    this.modelId = modelId;\n  }\n}\n\n/**\n * Event payload for model download progress.\n */\nexport type ModelDownloadProgressEvent = {\n  /** Model being downloaded */\n  modelId: string;\n  /** Download progress from 0 to 1 */\n  progress: number;\n};\n\n/**\n * Event payload for model state changes.\n */\nexport type ModelStateChangeEvent = {\n  /** Model whose state changed */\n  modelId: string;\n  /** New status */\n  status: DownloadableModelStatus;\n};\n"]}
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AA2aA;;GAEG;AACH,MAAM,OAAO,UAAW,SAAQ,KAAK;IACnC,IAAI,CAAiB;IACrB,OAAO,CAAS;IAEhB,YAAY,IAAoB,EAAE,OAAe,EAAE,OAAgB;QACjE,KAAK,CAAC,OAAO,IAAI,GAAG,IAAI,KAAK,OAAO,EAAE,CAAC,CAAC;QACxC,IAAI,CAAC,IAAI,GAAG,YAAY,CAAC;QACzB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;IACzB,CAAC;CACF","sourcesContent":["/**\n * Hardware backend for on-device model inference.\n *\n * - 'auto': Try GPU first, fall back to CPU (default)\n * - 'gpu': Force GPU — faster (~40-50 tok/s) but needs more memory\n * - 'cpu': Force CPU — slower (~2-5 tok/s) but works on low-RAM devices\n */\nexport type InferenceBackend = 'auto' | 'gpu' | 'cpu';\n\n/**\n * Sampling / generation parameters applied to a model session.\n *\n * Support is per-backend (on-device runtimes expose different knobs), so these\n * are best-effort — unsupported fields are ignored rather than erroring:\n *\n * | field       | Gemma (LiteRT-LM) | Apple Foundation Models | ML Kit |\n * |-------------|:-----------------:|:-----------------------:|:------:|\n * | temperature | ✓                 | ✓                       | —      |\n * | topK        | ✓                 | —                       | —      |\n * | topP        | ✓                 | —                       | —      |\n * | seed        | ✓ (iOS only)      | —                       | —      |\n * | maxTokens   | —                 | ✓ (max output)          | —      |\n *\n * Notes:\n * - Gemma/LiteRT-LM has no per-generation output-token cap (its `maxNumTokens`\n *   is the total KV-cache size, set at load), so `maxTokens` is not honored\n *   there. Its sampler (topK/topP/temperature[/seed]) is fixed at conversation\n *   creation, which is why generation config lives on setModel() rather than\n *   per-call. `seed` is currently wired on iOS only.\n * - The ML Kit built-in (Android default) does not yet apply generation config;\n *   it uses its own defaults.\n */\nexport type GenerationConfig = {\n  /** Sampling temperature. Lower = more deterministic. Typically 0.0–2.0. */\n  temperature?: number;\n  /** Nucleus sampling: number of top logits to consider. Must be > 0. */\n  topK?: number;\n  /** Nucleus sampling: cumulative probability threshold in [0, 1]. */\n  topP?: number;\n  /** RNG seed for reproducible sampling (Gemma only). */\n  seed?: number;\n  /** Maximum number of output tokens to generate (Apple FM / ML Kit only). */\n  maxTokens?: number;\n};\n\n/**\n * Options for setModel.\n */\nexport type SetModelOptions = {\n  /** Hardware backend to use for inference. Defaults to 'auto'. */\n  backend?: InferenceBackend;\n  /**\n   * Default sampling parameters for this model session. Applied when the model\n   * is activated and used for all subsequent sendMessage/streamMessage calls\n   * until the next setModel(). See {@link GenerationConfig} for per-backend support.\n   */\n  generation?: GenerationConfig;\n};\n\n/**\n * Role in a conversation message.\n */\nexport type LLMRole = 'system' | 'user' | 'assistant';\n\n/**\n * A single message in a conversation.\n */\nexport type LLMMessage = {\n  role: LLMRole;\n  content: string;\n};\n\n/**\n * Options for sendMessage.\n */\nexport type LLMSendOptions = {\n  /**\n   * Default system prompt to use if no system message is provided in the messages array.\n   * If a system message exists in the array, this is ignored.\n   */\n  systemPrompt?: string;\n  /**\n   * Abort the request. When the signal fires, the returned promise rejects with\n   * an INFERENCE_CANCELLED {@link ModelError}.\n   *\n   * Note: on-device, non-streaming generation cannot always be interrupted\n   * mid-decode — abort always unblocks the caller, but the model may keep\n   * computing in the background until it finishes, during which a new\n   * sendMessage/streamMessage will throw INFERENCE_BUSY. To truly interrupt a\n   * long generation, prefer streamMessage().stop().\n   */\n  signal?: AbortSignal;\n};\n\n/**\n * Response from sendMessage.\n */\nexport type LLMResponse = {\n  /** The generated response text */\n  text: string;\n};\n\n/**\n * Options for streamMessage.\n */\nexport type LLMStreamOptions = {\n  /**\n   * Default system prompt to use if no system message is provided in the messages array.\n   * If a system message exists in the array, this is ignored.\n   */\n  systemPrompt?: string;\n};\n\n/**\n * Handle returned by streamMessage.\n */\nexport type LLMStreamHandle = {\n  /** Resolves with the final text when streaming completes or is stopped. */\n  promise: Promise<LLMResponse>;\n  /** Stop streaming. Resolves `promise` with the text accumulated so far. */\n  stop: () => void;\n};\n\n/**\n * Event payload for streaming tokens.\n */\nexport type LLMStreamEvent = {\n  /** Unique identifier for this streaming session */\n  sessionId: string;\n  /** The token/chunk of text received */\n  token: string;\n  /** Accumulated text so far */\n  accumulatedText: string;\n  /** Whether this is the final chunk */\n  isDone: boolean;\n};\n\n/**\n * Callback function for streaming events.\n */\nexport type LLMStreamCallback = (event: LLMStreamEvent) => void;\n\n\n// ============================================================================\n// Structured Output (generateObject)\n// ============================================================================\n\n/**\n * The set of JSON Schema primitive `type` values understood by\n * {@link generateObject}'s local validator.\n */\nexport type JSONSchemaType =\n  | 'object'\n  | 'array'\n  | 'string'\n  | 'number'\n  | 'integer'\n  | 'boolean'\n  | 'null';\n\n/**\n * A JSON Schema describing the shape you want {@link generateObject} to return.\n *\n * A pragmatic subset is enforced locally — `type`, `properties`, `required`,\n * `items`, and `enum` — which covers most extraction shapes. Any other JSON\n * Schema keywords you include (e.g. `description`, `minLength`) are still sent\n * to the model to guide it, but are not validated on-device. Keep schemas small:\n * on-device models follow flat, shallow shapes far more reliably than deeply\n * nested ones.\n */\nexport type JSONSchema = {\n  /** Expected JSON type (or a union of types). */\n  type?: JSONSchemaType | JSONSchemaType[];\n  /** Human-readable hint passed through to the model. */\n  description?: string;\n  /** For `object` schemas: the schema of each named property. */\n  properties?: Record<string, JSONSchema>;\n  /** For `object` schemas: property names that must be present. */\n  required?: string[];\n  /** For `array` schemas: the schema each element must satisfy. */\n  items?: JSONSchema;\n  /** Restrict the value to this set of literals. */\n  enum?: ReadonlyArray<string | number | boolean | null>;\n  /** Other JSON Schema keywords are accepted and forwarded to the model. */\n  [key: string]: unknown;\n};\n\n/**\n * Options for {@link generateObject}.\n */\nexport type GenerateObjectOptions = {\n  /**\n   * System prompt used when the messages array has no system message. Defaults\n   * to a structured-output-oriented instruction. If a system message is present\n   * in the array, this is ignored (the schema instruction is appended to it).\n   */\n  systemPrompt?: string;\n  /**\n   * Abort the request. Behaves like {@link LLMSendOptions.signal} — the returned\n   * promise rejects with an INFERENCE_CANCELLED {@link ModelError}.\n   */\n  signal?: AbortSignal;\n  /**\n   * How many times to re-prompt the model when its output is not valid JSON or\n   * does not match the schema. Each repair feeds the error back to the model.\n   * Defaults to 2 (i.e. up to 3 generations total).\n   */\n  maxRepairAttempts?: number;\n};\n\n/**\n * Result of {@link generateObject}.\n */\nexport type GenerateObjectResult<T = unknown> = {\n  /** The parsed value, validated against the schema. */\n  object: T;\n  /** The raw model output that produced `object` (useful for debugging). */\n  text: string;\n};\n\n\n// ============================================================================\n// Tool / Function Calling (generateText)\n// ============================================================================\n\n/**\n * A tool (function) the model may call to fetch data or take an action.\n *\n * The model never runs anything itself — it *proposes* a call (a name + JSON\n * arguments), {@link generateText} validates the arguments against `parameters`,\n * and only then invokes your `execute`. The result is fed back into the\n * conversation so the model can use it to produce its final answer.\n *\n * @typeParam TArgs - Shape of the validated arguments passed to `execute`.\n * @typeParam TResult - What `execute` returns (serialized back to the model).\n */\nexport type Tool<TArgs = any, TResult = any> = {\n  /**\n   * What the tool does and when to use it. This is how the model decides\n   * whether a request matches this tool — make it specific and action-oriented.\n   */\n  description: string;\n  /**\n   * JSON Schema for the tool's arguments. The model is told to conform to it,\n   * and the args it proposes are validated against it (same pragmatic subset as\n   * {@link generateObject}) before `execute` runs. Keep it small and shallow.\n   */\n  parameters: JSONSchema;\n  /**\n   * Runs the tool with the validated arguments and returns a result.\n   *\n   * **Optional on purpose.** If you omit it, {@link generateText} does not run\n   * anything — it stops with `finishReason: 'tool-calls'` and hands you the\n   * proposed call so you can confirm, gate, or execute it yourself.\n   */\n  execute?: (args: TArgs) => TResult | Promise<TResult>;\n};\n\n/** A map of tool name → {@link Tool}, passed to {@link generateText}. */\nexport type ToolSet = Record<string, Tool>;\n\n/** A tool invocation the model proposed (name + validated arguments). */\nexport type ToolCall = {\n  /** The tool's key in the {@link ToolSet}. */\n  toolName: string;\n  /** Arguments the model supplied, validated against the tool's `parameters`. */\n  args: unknown;\n};\n\n/** The outcome of running a {@link ToolCall} via its `execute`. */\nexport type ToolResult = {\n  /** The tool's key in the {@link ToolSet}. */\n  toolName: string;\n  /** The arguments that were passed to `execute`. */\n  args: unknown;\n  /** Whatever `execute` returned (or `{ error }` if it threw). */\n  result: unknown;\n};\n\n/** One model round-trip in the {@link generateText} loop. */\nexport type StepResult = {\n  /** Assistant text produced this step (empty when the step only called a tool). */\n  text: string;\n  /** Tool calls proposed this step (at most one in the current protocol). */\n  toolCalls: ToolCall[];\n  /** Results of the tool calls executed this step. */\n  toolResults: ToolResult[];\n};\n\n/**\n * Why {@link generateText} stopped.\n *\n * - `'stop'`: the model produced a final text answer.\n * - `'tool-calls'`: stopped because a proposed tool has no `execute` — the call\n *   is returned for you to handle (human-in-the-loop).\n * - `'max-steps'`: hit the `maxSteps` cap while still calling tools. Raise the cap.\n */\nexport type GenerateTextFinishReason = 'stop' | 'tool-calls' | 'max-steps';\n\n/**\n * Options for {@link generateText}.\n */\nexport type GenerateTextOptions = {\n  /** Tools the model may call. Omit (or pass `{}`) for a plain text generation. */\n  tools?: ToolSet;\n  /**\n   * Maximum number of model round-trips (each call + tool execution is one step).\n   * Bounds the tool-calling chain so it can't run away. Defaults to 5.\n   */\n  maxSteps?: number;\n  /**\n   * System prompt used when the messages array has no system message. The tool\n   * instructions are appended to it (or to the array's system message if present).\n   */\n  systemPrompt?: string;\n  /**\n   * Abort the request. Behaves like {@link LLMSendOptions.signal} — the returned\n   * promise rejects with an INFERENCE_CANCELLED {@link ModelError}.\n   */\n  signal?: AbortSignal;\n  /**\n   * How many times to re-prompt within a step when the model emits a malformed\n   * tool call, an unknown tool name, or arguments that fail schema validation.\n   * Defaults to 2. If it still can't comply, `generateText` throws INFERENCE_FAILED.\n   */\n  maxRepairAttempts?: number;\n};\n\n/**\n * Result of {@link generateText}.\n */\nexport type GenerateTextResult = {\n  /** The final assistant text (empty if it stopped on a tool call without `execute`). */\n  text: string;\n  /** Every step taken, in order — useful for tracing or debugging. */\n  steps: StepResult[];\n  /** All tool calls across every step, flattened. */\n  toolCalls: ToolCall[];\n  /** All tool results across every step, flattened. */\n  toolResults: ToolResult[];\n  /** Why generation stopped. See {@link GenerateTextFinishReason}. */\n  finishReason: GenerateTextFinishReason;\n};\n\n\n// ============================================================================\n// Model Types\n// ============================================================================\n\n/**\n * A built-in model provided by the OS (e.g. Apple Foundation Models, ML Kit).\n * These are always available on supported devices -- no download needed.\n */\nexport type BuiltInModel = {\n  /** Unique model identifier (e.g. 'apple-fm', 'mlkit') */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Whether this model is available on the current device/OS */\n  available: boolean;\n  /** Platform this model is associated with */\n  platform: 'ios' | 'android';\n  /** Maximum context window in tokens */\n  contextWindow: number;\n};\n\n/**\n * A downloadable model that the user manages (download, load, delete).\n * These require explicit download before use.\n */\nexport type DownloadableModel = {\n  /** Unique model identifier (e.g. 'gemma-e2b', 'gemma-e4b') */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label (e.g. '2.3B') */\n  parameterCount: string;\n  /** License the weights are distributed under (e.g. 'Apache-2.0', 'MIT', 'Gemma'). */\n  license: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /** Maximum context window in tokens */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run */\n  minRamBytes: number;\n  /** Whether this device meets the model's minimum RAM requirement */\n  meetsRequirements: boolean;\n  /** Current lifecycle status */\n  status: DownloadableModelStatus;\n};\n\n/**\n * Lifecycle status of a downloadable model.\n *\n * - 'not-downloaded': Model file is not on disk\n * - 'downloading': Model file is being downloaded\n * - 'downloaded': Model file is on disk but not loaded into memory. Call\n *   setModel() to load it. This survives app restarts, so use it to decide\n *   whether a (re-)download is needed.\n * - 'loading': File is on disk, model is being loaded into memory for inference\n * - 'ready': Model is loaded in memory and ready for inference\n */\nexport type DownloadableModelStatus =\n  | 'not-downloaded'\n  | 'downloading'\n  | 'downloaded'\n  | 'loading'\n  | 'ready';\n\n/**\n * Error codes for model-related operations.\n */\nexport type ModelErrorCode =\n  | 'MODEL_NOT_FOUND'\n  | 'MODEL_NOT_DOWNLOADED'\n  | 'DOWNLOAD_FAILED'\n  | 'DOWNLOAD_CORRUPT'\n  | 'DOWNLOAD_STORAGE_FULL'\n  | 'DOWNLOAD_CANCELLED'\n  | 'INFERENCE_OOM'\n  | 'INFERENCE_FAILED'\n  | 'INFERENCE_BUSY'\n  | 'INFERENCE_CANCELLED'\n  | 'MODEL_LOAD_FAILED'\n  | 'DEVICE_NOT_SUPPORTED'\n  | 'UNKNOWN';\n\n/**\n * Structured error for model operations.\n */\nexport class ModelError extends Error {\n  code: ModelErrorCode;\n  modelId: string;\n\n  constructor(code: ModelErrorCode, modelId: string, message?: string) {\n    super(message ?? `${code}: ${modelId}`);\n    this.name = 'ModelError';\n    this.code = code;\n    this.modelId = modelId;\n  }\n}\n\n/**\n * Event payload for model download progress.\n */\nexport type ModelDownloadProgressEvent = {\n  /** Model being downloaded */\n  modelId: string;\n  /** Download progress from 0 to 1 */\n  progress: number;\n};\n\n/**\n * Event payload for model state changes.\n */\nexport type ModelStateChangeEvent = {\n  /** Model whose state changed */\n  modelId: string;\n  /** New status */\n  status: DownloadableModelStatus;\n};\n"]}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "expo-ai-kit",
-  "version": "0.6.0",
-  "description": "On-device AI for Expo — run Gemma 4, Apple Foundation Models & ML Kit locally: streaming, structured output, zero API keys",
+  "version": "0.8.0",
+  "description": "On-device AI for Expo — run Gemma 4, Apple Foundation Models & ML Kit locally: streaming, structured output, tool calling, zero API keys",
   "main": "build/index.js",
   "types": "build/index.d.ts",
   "files": [
@@ -38,6 +38,9 @@
     "expo",
     "gemma",
     "gemma-4",
+    "qwen",
+    "qwen3",
+    "phi-4",
     "on-device-ai",
     "llm",
     "litert",
@@ -45,6 +48,8 @@
     "ml-kit",
     "local-inference",
     "structured-output",
+    "tool-calling",
+    "function-calling",
     "json-schema",
     "expo-ai-kit",
     "ExpoAiKit"