expo-ai-kit 0.4.0 → 0.5.0

package/build/index.js

@@ -7,7 +7,104 @@ export * from './models';
 const DEFAULT_SYSTEM_PROMPT = 'You are a helpful, friendly assistant. Answer the user directly and concisely.';
 let streamIdCounter = 0;
 function generateSessionId() {
-    return `stream_${Date.now()}_${++streamIdCounter}`;
+    return `gen_${Date.now()}_${++streamIdCounter}`;
+}
+// The set of codes the native layer encodes in error messages as "CODE:modelId:reason".
+const KNOWN_ERROR_CODES = new Set([
+    'MODEL_NOT_FOUND',
+    'MODEL_NOT_DOWNLOADED',
+    'DOWNLOAD_FAILED',
+    'DOWNLOAD_CORRUPT',
+    'DOWNLOAD_STORAGE_FULL',
+    'DOWNLOAD_CANCELLED',
+    'INFERENCE_OOM',
+    'INFERENCE_FAILED',
+    'INFERENCE_BUSY',
+    'INFERENCE_CANCELLED',
+    'MODEL_LOAD_FAILED',
+    'DEVICE_NOT_SUPPORTED',
+]);
+/**
+ * Normalize an error from the native layer into a {@link ModelError}.
+ *
+ * The native modules format failures as "CODE:modelId:reason" (see the
+ * GemmaError/GemmaInferenceClient contract). Expo surfaces that string as the
+ * error's message, so we parse it here and rethrow a typed ModelError with a
+ * reliable `.code` and `.modelId`. Anything unrecognized becomes UNKNOWN.
+ */
+function toModelError(e) {
+    if (e instanceof ModelError)
+        throw e;
+    const message = String(e?.message ?? e ?? '');
+    const match = /^([A-Z_]+):([^:]*):([\s\S]*)$/.exec(message);
+    if (match && KNOWN_ERROR_CODES.has(match[1])) {
+        throw new ModelError(match[1], match[2], match[3]);
+    }
+    throw new ModelError('UNKNOWN', '', message);
+}
+/** Run a native promise, normalizing any rejection into a ModelError. */
+async function wrapNative(run) {
+    try {
+        return await run();
+    }
+    catch (e) {
+        toModelError(e);
+    }
+}
+// ---------------------------------------------------------------------------
+// Single-flight inference guard
+// ---------------------------------------------------------------------------
+// On-device models are backed by a single native context + KV cache that is not
+// safe for concurrent decodes (interleaving can corrupt the cache and crash the
+// native side). JS is single-threaded, so a synchronous check-and-set of this
+// flag before any `await` is race-free. The flag is shared by sendMessage and
+// streamMessage and is held until the *native* call settles — not until an
+// early abort — so a detached-but-still-running generation still blocks a new one.
+let inferenceInFlight = false;
+function acquireInference() {
+    if (inferenceInFlight) {
+        throw new ModelError('INFERENCE_BUSY', '', 'A generation is already in flight. Wait for it to finish, or stop the active stream first.');
+    }
+    inferenceInFlight = true;
+}
+/**
+ * Map the public GenerationConfig to the native shape, dropping undefined fields
+ * and validating ranges up front so callers get a clear error instead of an
+ * opaque native MODEL_LOAD_FAILED from the sampler.
+ */
+function toNativeGeneration(g) {
+    const out = {};
+    if (g?.temperature != null) {
+        if (g.temperature < 0) {
+            throw new Error('generation.temperature must be >= 0');
+        }
+        out.temperature = g.temperature;
+    }
+    if (g?.topK != null) {
+        if (!Number.isInteger(g.topK) || g.topK <= 0) {
+            throw new Error('generation.topK must be a positive integer');
+        }
+        out.topK = g.topK;
+    }
+    if (g?.topP != null) {
+        if (g.topP < 0 || g.topP > 1) {
+            throw new Error('generation.topP must be within [0, 1]');
+        }
+        out.topP = g.topP;
+    }
+    if (g?.seed != null) {
+        if (!Number.isInteger(g.seed)) {
+            throw new Error('generation.seed must be an integer');
+        }
+        out.seed = g.seed;
+    }
+    if (g?.maxTokens != null) {
+        if (!Number.isInteger(g.maxTokens) || g.maxTokens <= 0) {
+            throw new Error('generation.maxTokens must be a positive integer');
+        }
+        out.maxTokens = g.maxTokens;
+    }
+    return out;
 }
 // ============================================================================
 // Inference API
@@ -64,12 +161,58 @@ export async function sendMessage(messages, options) {
     if (!messages || messages.length === 0) {
         throw new Error('messages array cannot be empty');
     }
+    if (options?.signal?.aborted) {
+        throw new ModelError('INFERENCE_CANCELLED', '', 'Aborted before start');
+    }
     // Determine system prompt: use from messages array if present, else options, else default
     const hasSystemMessage = messages.some((m) => m.role === 'system');
     const systemPrompt = hasSystemMessage
         ? '' // Native will extract from messages
         : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
-    return ExpoAiKitModule.sendMessage(messages, systemPrompt);
+    acquireInference(); // throws INFERENCE_BUSY if a generation is already running
+    const sessionId = generateSessionId();
+    // Hold the single-flight flag until the NATIVE call settles — even if the
+    // caller aborts early — because the model may keep computing in the background.
+    const native = ExpoAiKitModule.sendMessage(messages, systemPrompt, sessionId);
+    const release = () => {
+        inferenceInFlight = false;
+    };
+    native.then(release, release);
+    const signal = options?.signal;
+    if (!signal) {
+        try {
+            return await native;
+        }
+        catch (e) {
+            toModelError(e);
+        }
+    }
+    // Race the native result against the abort signal. On abort we unblock the
+    // caller immediately and best-effort ask native to cancel; the flag stays
+    // held (via `release` above) until the native call actually finishes.
+    return await new Promise((resolve, reject) => {
+        let done = false;
+        const finish = (action) => {
+            if (done)
+                return;
+            done = true;
+            signal.removeEventListener('abort', onAbort);
+            action();
+        };
+        function onAbort() {
+            ExpoAiKitModule.stopStreaming(sessionId).catch(() => { });
+            finish(() => reject(new ModelError('INFERENCE_CANCELLED', '', 'Aborted by caller')));
+        }
+        signal.addEventListener('abort', onAbort);
+        native.then((r) => finish(() => resolve(r)), (e) => finish(() => {
+            try {
+                toModelError(e);
+            }
+            catch (me) {
+                reject(me);
+            }
+        }));
+    });
 }
 /**
  * Stream messages to the on-device LLM and receive progressive token updates.
@@ -118,42 +261,60 @@ export function streamMessage(messages, onToken, options) {
             stop: () => { },
         };
     }
+    if (inferenceInFlight) {
+        return {
+            promise: Promise.reject(new ModelError('INFERENCE_BUSY', '', 'A generation is already in flight. Stop the active stream first.')),
+            stop: () => { },
+        };
+    }
+    inferenceInFlight = true; // set synchronously — race-free with other JS
     const sessionId = generateSessionId();
-    let finalText = '';
-    let stopped = false;
     // Determine system prompt: use from messages array if present, else options, else default
     const hasSystemMessage = messages.some((m) => m.role === 'system');
     const systemPrompt = hasSystemMessage
         ? '' // Native will extract from messages
         : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
+    let finalText = '';
+    let settled = false;
+    let subscription;
+    let resolveOuter;
+    let rejectOuter;
+    // Settle exactly once: remove the listener and release the single-flight flag.
+    const settle = (action) => {
+        if (settled)
+            return;
+        settled = true;
+        subscription?.remove();
+        inferenceInFlight = false;
+        action();
+    };
     const promise = new Promise((resolve, reject) => {
-        // Subscribe to stream events
-        const subscription = ExpoAiKitModule.addListener('onStreamToken', (event) => {
-            // Only process events for this session
-            if (event.sessionId !== sessionId)
-                return;
-            finalText = event.accumulatedText;
-            // Call the user's callback
-            onToken(event);
-            // If done, clean up and resolve
-            if (event.isDone) {
-                subscription.remove();
-                resolve({ text: finalText });
+        resolveOuter = resolve;
+        rejectOuter = reject;
+    });
+    subscription = ExpoAiKitModule.addListener('onStreamToken', (event) => {
+        if (event.sessionId !== sessionId)
+            return;
+        finalText = event.accumulatedText;
+        onToken(event);
+        if (event.isDone)
+            settle(() => resolveOuter({ text: finalText }));
+    });
+    ExpoAiKitModule.startStreaming(messages, systemPrompt, sessionId).catch((error) => {
+        settle(() => {
+            try {
+                toModelError(error);
+            }
+            catch (me) {
+                rejectOuter(me);
             }
-        });
-        // Start streaming on native side
-        ExpoAiKitModule.startStreaming(messages, systemPrompt, sessionId).catch((error) => {
-            subscription.remove();
-            reject(error);
         });
     });
     const stop = () => {
-        if (stopped)
-            return;
-        stopped = true;
-        ExpoAiKitModule.stopStreaming(sessionId).catch(() => {
-            // Ignore errors when stopping
-        });
+        // Best-effort native cancel (native also emits a terminal isDone on cancel),
+        // but resolve immediately with the text so far so `promise` can never hang.
+        ExpoAiKitModule.stopStreaming(sessionId).catch(() => { });
+        settle(() => resolveOuter({ text: finalText }));
     };
     return { promise, stop };
 }
@@ -194,8 +355,10 @@ export async function getDownloadableModels() {
     catch {
         // Native call unavailable -- default to 0 (all models will show meetsRequirements: false)
     }
-    return platformModels.map((entry) => {
-        const status = ExpoAiKitModule.getDownloadableModelStatus(entry.id);
+    return Promise.all(platformModels.map(async (entry) => {
+        // Await: on iOS this bridges as a Promise (reads actor state); on Android
+        // it's synchronous and awaiting a plain value is a no-op.
+        const status = await ExpoAiKitModule.getDownloadableModelStatus(entry.id);
         return {
             id: entry.id,
             name: entry.name,
@@ -206,7 +369,34 @@ export async function getDownloadableModels() {
             meetsRequirements: deviceRamBytes >= entry.minRamBytes,
             status,
         };
-    });
+    }));
+}
+/**
+ * Pick the best downloadable model the current device can run.
+ *
+ * Returns the most capable model (largest, by RAM requirement) whose
+ * `meetsRequirements` is true — e.g. Gemma 4 E4B on high-spec phones, falling
+ * back to E2B on more constrained ones — or `null` if the device can't run any.
+ *
+ * This is a convenience over {@link getDownloadableModels}; the caller still
+ * downloads + activates explicitly. Pass `platform` is implicit (current OS).
+ *
+ * @example
+ * ```ts
+ * const best = await getRecommendedModel();
+ * if (best) {
+ *   await downloadModel(best.id, { onProgress });
+ *   await setModel(best.id);
+ * }
+ * ```
+ */
+export async function getRecommendedModel() {
+    const models = await getDownloadableModels();
+    const runnable = models.filter((m) => m.meetsRequirements);
+    if (runnable.length === 0)
+        return null;
+    // Higher RAM requirement ⇒ larger/more capable model. Prefer the biggest that fits.
+    return runnable.sort((a, b) => b.minRamBytes - a.minRamBytes)[0];
 }
 /**
  * Download a model to the device.
@@ -251,12 +441,26 @@ export async function downloadModel(modelId, options) {
         });
     }
     try {
-        await ExpoAiKitModule.downloadModel(modelId, entry.downloadUrl, entry.sha256);
+        await wrapNative(() => ExpoAiKitModule.downloadModel(modelId, entry.downloadUrl, entry.sha256));
     }
     finally {
         subscription?.remove();
     }
 }
+/**
+ * Cancel an in-flight download for a model.
+ *
+ * The in-progress {@link downloadModel} promise rejects with a
+ * DOWNLOAD_CANCELLED {@link ModelError}. No-op if the model isn't downloading.
+ *
+ * @param modelId - ID of the model whose download should be cancelled
+ */
+export async function cancelDownload(modelId) {
+    if (Platform.OS !== 'ios' && Platform.OS !== 'android') {
+        return;
+    }
+    await wrapNative(() => ExpoAiKitModule.cancelDownload(modelId));
+}
 /**
  * Delete a downloaded model from the device.
  *
@@ -270,7 +474,7 @@ export async function deleteModel(modelId) {
     if (!entry) {
         throw new ModelError('MODEL_NOT_FOUND', modelId);
     }
-    await ExpoAiKitModule.deleteModel(modelId);
+    await wrapNative(() => ExpoAiKitModule.deleteModel(modelId));
 }
 /**
  * Set the active model for inference.
@@ -299,7 +503,8 @@ export async function setModel(modelId, options) {
     const entry = getRegistryEntry(modelId);
     const minRamBytes = entry?.minRamBytes ?? 0;
     const backend = options?.backend ?? 'auto';
-    await ExpoAiKitModule.setModel(modelId, minRamBytes, backend);
+    const generation = toNativeGeneration(options?.generation);
+    await wrapNative(() => ExpoAiKitModule.setModel(modelId, minRamBytes, backend, generation));
 }
 /**
  * Get the ID of the currently active model.
@@ -316,6 +521,6 @@ export function getActiveModel() {
  * No-op if no downloadable model is currently loaded.
  */
 export async function unloadModel() {
-    await ExpoAiKitModule.unloadModel();
+    await wrapNative(() => ExpoAiKitModule.unloadModel());
 }
 //# sourceMappingURL=index.js.map

package/build/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,eAAe,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EASL,UAAU,GAEX,MAAM,SAAS,CAAC;AACjB,OAAO,EAAE,cAAc,EAAE,gBAAgB,EAAE,MAAM,UAAU,CAAC;AAE5D,cAAc,SAAS,CAAC;AACxB,cAAc,UAAU,CAAC;AAEzB,MAAM,qBAAqB,GACzB,gFAAgF,CAAC;AAEnF,IAAI,eAAe,GAAG,CAAC,CAAC;AACxB,SAAS,iBAAiB;IACxB,OAAO,UAAU,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,CAAC;AACrD,CAAC;AAED,+EAA+E;AAC/E,gBAAgB;AAChB,+EAA+E;AAE/E;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,KAAK,CAAC;IACf,CAAC;IACD,OAAO,eAAe,CAAC,WAAW,EAAE,CAAC;AACvC,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAC/B,QAAsB,EACtB,OAAwB;IAExB,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;IACtB,CAAC;IAED,IAAI,CAAC,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvC,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;IACpD,CAAC;IAED,0FAA0F;IAC1F,MAAM,gBAAgB,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IACnE,MAAM,YAAY,GAAG,gBAAgB;QACnC,CAAC,CAAC,EAAE,CAAC,oCAAoC;QACzC,CAAC,CAAC,OAAO,EAAE,YAAY,IAAI,qBAAqB,CAAC;IAEnD,OAAO,eAAe,CAAC,WAAW,CAAC,QAAQ,EAAE,YAAY,CAAC,CAAC;AAC7D,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,MAAM,UAAU,aAAa,CAC3B,QAAsB,EACtB,OAA0B,EAC1B,OAA0B;IAE1B,+BAA+B;IAC/B,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;YACtC,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC;SACf,CAAC;IACJ,CAAC;IAED,IAAI,CAAC,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvC,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;YACpE,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC;SACf,CAAC;IACJ,CAAC;IAED,MAAM,SAAS,GAAG,iBAAiB,EAAE,CAAC;IACtC,IAAI,SAAS,GAAG,EAAE,CAAC;IACnB,IAAI,OAAO,GAAG,KAAK,CAAC;IAEpB,0FAA0F;IAC1F,MAAM,gBAAgB,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IACnE,MAAM,YAAY,GAAG,gBAAgB;QACnC,CAAC,CAAC,EAAE,CAAC,oCAAoC;QACzC,CAAC,CAAC,OAAO,EAAE,YAAY,IAAI,qBAAqB,CAAC;IAEnD,MAAM,OAAO,GAAG,IAAI,OAAO,CAAc,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QAC3D,6BAA6B;QAC7B,MAAM,YAAY,GAAG,eAAe,CAAC,WAAW,CAC9C,eAAe,EACf,CAAC,KAAqB,EAAE,EAAE;YACxB,uCAAuC;YACvC,IAAI,KAAK,CAAC,SAAS,KAAK,SAAS;gBAAE,OAAO;YAE1C,SAAS,GAAG,KAAK,CAAC,eAAe,CAAC;YAElC,2BAA2B;YAC3B,OAAO,CAAC,KAAK,CAAC,CAAC;YAEf,gCAAgC;YAChC,IAAI,KAAK,CAAC,MAAM,EAAE,CAAC;gBACjB,YAAY,CAAC,MAAM,EAAE,CAAC;gBACtB,OAAO,CAAC,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;YAC/B,CAAC;QACH,CAAC,CACF,CAAC;QAEF,iCAAiC;QACjC,eAAe,CAAC,cAAc,CAAC,QAAQ,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC,KAAK,CACrE,CAAC,KAAK,EAAE,EAAE;YACR,YAAY,CAAC,MAAM,EAAE,CAAC;YACtB,MAAM,CAAC,KAAK,CAAC,CAAC;QAChB,CAAC,CACF,CAAC;IACJ,CAAC,CAAC,CAAC;IAEH,MAAM,IAAI,GAAG,GAAG,EAAE;QAChB,IAAI,OAAO;YAAE,OAAO;QACpB,OAAO,GAAG,IAAI,CAAC;QACf,eAAe,CAAC,aAAa,CAAC,SAAS,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE;YAClD,8BAA8B;QAChC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC;IAEF,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;AAC3B,CAAC;AAED,+EAA+E;AAC/E,uBAAuB;AACvB,+EAA+E;AAE/E;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB;IACpC,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,CAAC;IACZ,CAAC;IACD,OAAO,eAAe,CAAC,gBAAgB,EAAE,CAAC;AAC5C,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB;IACzC,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,CAAC;IACZ,CAAC;IAED,MAAM,cAAc,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CACrD,KAAK,CAAC,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAuB,CAAC,CACpE,CAAC;IAEF,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,IAAI,CAAC;QACH,cAAc,GAAG,eAAe,CAAC,iBAAiB,EAAE,CAAC;IACvD,CAAC;IAAC,MAAM,CAAC;QACP,0FAA0F;IAC5F,CAAC;IAED,OAAO,cAAc,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE;QAClC,MAAM,MAAM,GAAG,eAAe,CAAC,0BAA0B,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QACpE,OAAO;YACL,EAAE,EAAE,KAAK,CAAC,EAAE;YACZ,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,cAAc,EAAE,KAAK,CAAC,cAAc;YACpC,SAAS,EAAE,KAAK,CAAC,SAAS;YAC1B,aAAa,EAAE,KAAK,CAAC,aAAa;YAClC,WAAW,EAAE,KAAK,CAAC,WAAW;YAC9B,iBAAiB,EAAE,cAAc,IAAI,KAAK,CAAC,WAAW;YACtD,MAAM;SACP,CAAC;IACJ,CAAC,CAAC,CAAC;AACL,CAAC;AAED;;;;;;;;;;;;;;GAcG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,OAAe,EACf,OAAqD;IAErD,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,CAAC,KAAK,EAAE,CAAC;QACX,MAAM,IAAI,UAAU,CAAC,iBAAiB,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAuB,CAAC,EAAE,CAAC;QACzE,MAAM,IAAI,UAAU,CAClB,sBAAsB,EACtB,OAAO,EACP,SAAS,OAAO,wBAAwB,QAAQ,CAAC,EAAE,EAAE,CACtD,CAAC;IACJ,CAAC;IAED,IAAI,CAAC;QACH,MAAM,cAAc,GAAG,eAAe,CAAC,iBAAiB,EAAE,CAAC;QAC3D,IAAI,cAAc,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;YACvC,MAAM,IAAI,UAAU,CAClB,sBAAsB,EACtB,OAAO,EACP,cAAc,IAAI,CAAC,KAAK,CAAC,cAAc,GAAG,GAAG,CAAC,0BAA0B,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,WAAW,GAAG,GAAG,CAAC,IAAI,CAChH,CAAC;QACJ,CAAC;IACH,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,IAAI,CAAC,YAAY,UAAU;YAAE,MAAM,CAAC,CAAC;QACrC,sDAAsD;IACxD,CAAC;IAED,IAAI,YAAwE,CAAC;IAC7E,IAAI,OAAO,EAAE,UAAU,EAAE,CAAC;QACxB,YAAY,GAAG,eAAe,CAAC,WAAW,CACxC,oBAAoB,EACpB,CAAC,KAAK,EAAE,EAAE;YACR,IAAI,KAAK,CAAC,OAAO,KAAK,OAAO,EAAE,CAAC;gBAC9B,OAAO,CAAC,UAAW,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;YACtC,CAAC;QACH,CAAC,CACF,CAAC;IACJ,CAAC;IAED,IAAI,CAAC;QACH,MAAM,eAAe,CAAC,aAAa,CACjC,OAAO,EACP,KAAK,CAAC,WAAW,EACjB,KAAK,CAAC,MAAM,CACb,CAAC;IACJ,CAAC;YAAS,CAAC;QACT,YAAY,EAAE,MAAM,EAAE,CAAC;IACzB,CAAC;AACH,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,OAAe;IAC/C,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,CAAC,KAAK,EAAE,CAAC;QACX,MAAM,IAAI,UAAU,CAAC,iBAAiB,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,MAAM,eAAe,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC;AAC7C,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,MAAM,CAAC,KAAK,UAAU,QAAQ,CAAC,OAAe,EAAE,OAAyB;IACvE,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,WAAW,GAAG,KAAK,EAAE,WAAW,IAAI,CAAC,CAAC;IAC5C,MAAM,OAAO,GAAG,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;IAC3C,MAAM,eAAe,CAAC,QAAQ,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,CAAC,CAAC;AAChE,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc;IAC5B,OAAO,eAAe,CAAC,cAAc,EAAE,CAAC;AAC1C,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,MAAM,eAAe,CAAC,WAAW,EAAE,CAAC;AACtC,CAAC","sourcesContent":["import ExpoAiKitModule from './ExpoAiKitModule';\nimport { Platform } from 'react-native';\nimport {\n  LLMMessage,\n  LLMSendOptions,\n  LLMResponse,\n  LLMStreamOptions,\n  LLMStreamEvent,\n  LLMStreamCallback,\n  BuiltInModel,\n  DownloadableModel,\n  ModelError,\n  SetModelOptions,\n} from './types';\nimport { MODEL_REGISTRY, getRegistryEntry } from './models';\n\nexport * from './types';\nexport * from './models';\n\nconst DEFAULT_SYSTEM_PROMPT =\n  'You are a helpful, friendly assistant. Answer the user directly and concisely.';\n\nlet streamIdCounter = 0;\nfunction generateSessionId(): string {\n  return `stream_${Date.now()}_${++streamIdCounter}`;\n}\n\n// ============================================================================\n// Inference API\n// ============================================================================\n\n/**\n * Check if on-device AI is available on the current device.\n * Returns false on unsupported platforms (web, etc.).\n */\nexport async function isAvailable(): Promise<boolean> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return false;\n  }\n  return ExpoAiKitModule.isAvailable();\n}\n\n/**\n * Send messages to the on-device LLM and get a response.\n *\n * @param messages - Array of messages representing the conversation\n * @param options - Optional settings (systemPrompt fallback)\n * @returns Promise with the generated response\n *\n * @example\n * ```ts\n * const response = await sendMessage([\n *   { role: 'user', content: 'What is 2 + 2?' }\n * ]);\n * console.log(response.text); // \"4\"\n * ```\n *\n * @example\n * ```ts\n * // With system prompt\n * const response = await sendMessage(\n *   [{ role: 'user', content: 'Hello!' }],\n *   { systemPrompt: 'You are a pirate. Respond in pirate speak.' }\n * );\n * ```\n *\n * @example\n * ```ts\n * // Multi-turn conversation\n * const response = await sendMessage([\n *   { role: 'system', content: 'You are a helpful assistant.' },\n *   { role: 'user', content: 'My name is Alice.' },\n *   { role: 'assistant', content: 'Nice to meet you, Alice!' },\n *   { role: 'user', content: 'What is my name?' }\n * ]);\n * ```\n */\nexport async function sendMessage(\n  messages: LLMMessage[],\n  options?: LLMSendOptions\n): Promise<LLMResponse> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return { text: '' };\n  }\n\n  if (!messages || messages.length === 0) {\n    throw new Error('messages array cannot be empty');\n  }\n\n  // Determine system prompt: use from messages array if present, else options, else default\n  const hasSystemMessage = messages.some((m) => m.role === 'system');\n  const systemPrompt = hasSystemMessage\n    ? '' // Native will extract from messages\n    : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;\n\n  return ExpoAiKitModule.sendMessage(messages, systemPrompt);\n}\n\n/**\n * Stream messages to the on-device LLM and receive progressive token updates.\n *\n * @param messages - Array of messages representing the conversation\n * @param onToken - Callback function called for each token/chunk received\n * @param options - Optional settings (systemPrompt fallback)\n * @returns Object with stop() function to cancel streaming and promise that resolves when complete\n *\n * @example\n * ```ts\n * // Basic streaming\n * const { promise } = streamMessage(\n *   [{ role: 'user', content: 'Tell me a story' }],\n *   (event) => {\n *     console.log(event.token); // Each token as it arrives\n *     console.log(event.accumulatedText); // Full text so far\n *   }\n * );\n * await promise;\n * ```\n *\n * @example\n * ```ts\n * // With cancellation\n * const { promise, stop } = streamMessage(\n *   [{ role: 'user', content: 'Write a long essay' }],\n *   (event) => setText(event.accumulatedText)\n * );\n *\n * // Cancel after 5 seconds\n * setTimeout(() => stop(), 5000);\n * ```\n */\nexport function streamMessage(\n  messages: LLMMessage[],\n  onToken: LLMStreamCallback,\n  options?: LLMStreamOptions\n): { promise: Promise<LLMResponse>; stop: () => void } {\n  // Handle unsupported platforms\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return {\n      promise: Promise.resolve({ text: '' }),\n      stop: () => {},\n    };\n  }\n\n  if (!messages || messages.length === 0) {\n    return {\n      promise: Promise.reject(new Error('messages array cannot be empty')),\n      stop: () => {},\n    };\n  }\n\n  const sessionId = generateSessionId();\n  let finalText = '';\n  let stopped = false;\n\n  // Determine system prompt: use from messages array if present, else options, else default\n  const hasSystemMessage = messages.some((m) => m.role === 'system');\n  const systemPrompt = hasSystemMessage\n    ? '' // Native will extract from messages\n    : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;\n\n  const promise = new Promise<LLMResponse>((resolve, reject) => {\n    // Subscribe to stream events\n    const subscription = ExpoAiKitModule.addListener(\n      'onStreamToken',\n      (event: LLMStreamEvent) => {\n        // Only process events for this session\n        if (event.sessionId !== sessionId) return;\n\n        finalText = event.accumulatedText;\n\n        // Call the user's callback\n        onToken(event);\n\n        // If done, clean up and resolve\n        if (event.isDone) {\n          subscription.remove();\n          resolve({ text: finalText });\n        }\n      }\n    );\n\n    // Start streaming on native side\n    ExpoAiKitModule.startStreaming(messages, systemPrompt, sessionId).catch(\n      (error) => {\n        subscription.remove();\n        reject(error);\n      }\n    );\n  });\n\n  const stop = () => {\n    if (stopped) return;\n    stopped = true;\n    ExpoAiKitModule.stopStreaming(sessionId).catch(() => {\n      // Ignore errors when stopping\n    });\n  };\n\n  return { promise, stop };\n}\n\n// ============================================================================\n// Model Management API\n// ============================================================================\n\n/**\n * Get all built-in models available on the current platform.\n *\n * Built-in models are provided by the OS and require no download.\n * On iOS this returns Apple Foundation Models; on Android, ML Kit.\n *\n * @returns Array of built-in models with availability status\n */\nexport async function getBuiltInModels(): Promise<BuiltInModel[]> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return [];\n  }\n  return ExpoAiKitModule.getBuiltInModels();\n}\n\n/**\n * Get all downloadable models from the registry, enriched with on-device status.\n *\n * Reads from the hardcoded MODEL_REGISTRY and queries the native layer\n * for the current download/load status of each model.\n *\n * @returns Array of downloadable models with their current status\n */\nexport async function getDownloadableModels(): Promise<DownloadableModel[]> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return [];\n  }\n\n  const platformModels = MODEL_REGISTRY.filter((entry) =>\n    entry.supportedPlatforms.includes(Platform.OS as 'ios' | 'android')\n  );\n\n  let deviceRamBytes = 0;\n  try {\n    deviceRamBytes = ExpoAiKitModule.getDeviceRamBytes();\n  } catch {\n    // Native call unavailable -- default to 0 (all models will show meetsRequirements: false)\n  }\n\n  return platformModels.map((entry) => {\n    const status = ExpoAiKitModule.getDownloadableModelStatus(entry.id);\n    return {\n      id: entry.id,\n      name: entry.name,\n      parameterCount: entry.parameterCount,\n      sizeBytes: entry.sizeBytes,\n      contextWindow: entry.contextWindow,\n      minRamBytes: entry.minRamBytes,\n      meetsRequirements: deviceRamBytes >= entry.minRamBytes,\n      status,\n    };\n  });\n}\n\n/**\n * Download a model to the device.\n *\n * Looks up the model in the registry, validates platform support and\n * device requirements, then initiates the download with integrity verification.\n *\n * @param modelId - ID of the model to download (e.g. 'gemma-e2b')\n * @param options - Optional download configuration\n * @param options.onProgress - Callback with download progress (0-1)\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is not in the registry\n * @throws {ModelError} DEVICE_NOT_SUPPORTED if platform is not supported\n * @throws {ModelError} DOWNLOAD_FAILED on network error\n * @throws {ModelError} DOWNLOAD_STORAGE_FULL if insufficient disk space\n * @throws {ModelError} DOWNLOAD_CORRUPT if SHA256 hash doesn't match\n */\nexport async function downloadModel(\n  modelId: string,\n  options?: { onProgress?: (progress: number) => void }\n): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  if (!entry) {\n    throw new ModelError('MODEL_NOT_FOUND', modelId);\n  }\n\n  if (!entry.supportedPlatforms.includes(Platform.OS as 'ios' | 'android')) {\n    throw new ModelError(\n      'DEVICE_NOT_SUPPORTED',\n      modelId,\n      `Model ${modelId} is not supported on ${Platform.OS}`\n    );\n  }\n\n  try {\n    const deviceRamBytes = ExpoAiKitModule.getDeviceRamBytes();\n    if (deviceRamBytes < entry.minRamBytes) {\n      throw new ModelError(\n        'DEVICE_NOT_SUPPORTED',\n        modelId,\n        `Device has ${Math.round(deviceRamBytes / 1e9)}GB RAM, model requires ${Math.round(entry.minRamBytes / 1e9)}GB`\n      );\n    }\n  } catch (e) {\n    if (e instanceof ModelError) throw e;\n    // If getDeviceRamBytes is unavailable, skip the check\n  }\n\n  let subscription: ReturnType<typeof ExpoAiKitModule.addListener> | undefined;\n  if (options?.onProgress) {\n    subscription = ExpoAiKitModule.addListener(\n      'onDownloadProgress',\n      (event) => {\n        if (event.modelId === modelId) {\n          options.onProgress!(event.progress);\n        }\n      }\n    );\n  }\n\n  try {\n    await ExpoAiKitModule.downloadModel(\n      modelId,\n      entry.downloadUrl,\n      entry.sha256\n    );\n  } finally {\n    subscription?.remove();\n  }\n}\n\n/**\n * Delete a downloaded model from the device.\n *\n * If the model is currently loaded, it will be unloaded first.\n *\n * @param modelId - ID of the model to delete\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is not in the registry\n */\nexport async function deleteModel(modelId: string): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  if (!entry) {\n    throw new ModelError('MODEL_NOT_FOUND', modelId);\n  }\n\n  await ExpoAiKitModule.deleteModel(modelId);\n}\n\n/**\n * Set the active model for inference.\n *\n * This is the sole gatekeeper for model validity. If setModel succeeds,\n * the model is loaded and ready -- sendMessage never needs its own check.\n *\n * For downloadable models, this loads the model into memory (status\n * transitions: loading -> ready). Only one downloadable model can be\n * loaded at a time; the previous one is auto-unloaded.\n *\n * For built-in models, this simply switches the active backend.\n *\n * If setModel was never called, sendMessage uses the platform built-in\n * model (today's behavior, no error).\n *\n * @param modelId - ID of the model to activate (e.g. 'gemma-e2b', 'apple-fm', 'mlkit')\n * @param options - Optional configuration for model loading\n * @param options.backend - Hardware backend: 'auto' (default, GPU with CPU fallback), 'gpu', or 'cpu'\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is invalid\n * @throws {ModelError} MODEL_NOT_DOWNLOADED if the downloadable model file is not on disk\n * @throws {ModelError} MODEL_LOAD_FAILED if loading into memory fails\n * @throws {ModelError} INFERENCE_OOM if device can't fit model in memory\n */\nexport async function setModel(modelId: string, options?: SetModelOptions): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  const minRamBytes = entry?.minRamBytes ?? 0;\n  const backend = options?.backend ?? 'auto';\n  await ExpoAiKitModule.setModel(modelId, minRamBytes, backend);\n}\n\n/**\n * Get the ID of the currently active model.\n *\n * @returns The active model ID (e.g. 'apple-fm', 'mlkit', 'gemma-e2b')\n */\nexport function getActiveModel(): string {\n  return ExpoAiKitModule.getActiveModel();\n}\n\n/**\n * Explicitly unload the current downloadable model from memory.\n *\n * Frees memory and reverts to the platform built-in model.\n * No-op if no downloadable model is currently loaded.\n */\nexport async function unloadModel(): Promise<void> {\n  await ExpoAiKitModule.unloadModel();\n}\n\n"]}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,eAAgD,MAAM,mBAAmB,CAAC;AACjF,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAWL,UAAU,GAGX,MAAM,SAAS,CAAC;AACjB,OAAO,EAAE,cAAc,EAAE,gBAAgB,EAAE,MAAM,UAAU,CAAC;AAE5D,cAAc,SAAS,CAAC;AACxB,cAAc,UAAU,CAAC;AAEzB,MAAM,qBAAqB,GACzB,gFAAgF,CAAC;AAEnF,IAAI,eAAe,GAAG,CAAC,CAAC;AACxB,SAAS,iBAAiB;IACxB,OAAO,OAAO,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,CAAC;AAClD,CAAC;AAED,wFAAwF;AACxF,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAAiB;IAChD,iBAAiB;IACjB,sBAAsB;IACtB,iBAAiB;IACjB,kBAAkB;IAClB,uBAAuB;IACvB,oBAAoB;IACpB,eAAe;IACf,kBAAkB;IAClB,gBAAgB;IAChB,qBAAqB;IACrB,mBAAmB;IACnB,sBAAsB;CACvB,CAAC,CAAC;AAEH;;;;;;;GAOG;AACH,SAAS,YAAY,CAAC,CAAU;IAC9B,IAAI,CAAC,YAAY,UAAU;QAAE,MAAM,CAAC,CAAC;IACrC,MAAM,OAAO,GAAG,MAAM,CAAE,CAAS,EAAE,OAAO,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC;IACvD,MAAM,KAAK,GAAG,+BAA+B,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;IAC5D,IAAI,KAAK,IAAI,iBAAiB,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAmB,CAAC,EAAE,CAAC;QAC/D,MAAM,IAAI,UAAU,CAAC,KAAK,CAAC,CAAC,CAAmB,EAAE,KAAK,CAAC,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACvE,CAAC;IACD,MAAM,IAAI,UAAU,CAAC,SAAS,EAAE,EAAE,EAAE,OAAO,CAAC,CAAC;AAC/C,CAAC;AAED,yEAAyE;AACzE,KAAK,UAAU,UAAU,CAAI,GAAqB;IAChD,IAAI,CAAC;QACH,OAAO,MAAM,GAAG,EAAE,CAAC;IACrB,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,YAAY,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC;AACH,CAAC;AAED,8EAA8E;AAC9E,gCAAgC;AAChC,8EAA8E;AAC9E,gFAAgF;AAChF,gFAAgF;AAChF,8EAA8E;AAC9E,8EAA8E;AAC9E,2EAA2E;AAC3E,mFAAmF;AACnF,IAAI,iBAAiB,GAAG,KAAK,CAAC;AAE9B,SAAS,gBAAgB;IACvB,IAAI,iBAAiB,EAAE,CAAC;QACtB,MAAM,IAAI,UAAU,CAClB,gBAAgB,EAChB,EAAE,EACF,4FAA4F,CAC7F,CAAC;IACJ,CAAC;IACD,iBAAiB,GAAG,IAAI,CAAC;AAC3B,CAAC;AAED;;;;GAIG;AACH,SAAS,kBAAkB,CAAC,CAAoB;IAC9C,MAAM,GAAG,GAA2B,EAAE,CAAC;IACvC,IAAI,CAAC,EAAE,WAAW,IAAI,IAAI,EAAE,CAAC;QAC3B,IAAI,CAAC,CAAC,WAAW,GAAG,CAAC,EAAE,CAAC;YACtB,MAAM,IAAI,KAAK,CAAC,qCAAqC,CAAC,CAAC;QACzD,CAAC;QACD,GAAG,CAAC,WAAW,GAAG,CAAC,CAAC,WAAW,CAAC;IAClC,CAAC;IACD,IAAI,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC;QACpB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,IAAI,CAAC,EAAE,CAAC;YAC7C,MAAM,IAAI,KAAK,CAAC,4CAA4C,CAAC,CAAC;QAChE,CAAC;QACD,GAAG,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC;IACpB,CAAC;IACD,IAAI,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC;QACpB,IAAI,CAAC,CAAC,IAAI,GAAG,CAAC,IAAI,CAAC,CAAC,IAAI,GAAG,CAAC,EAAE,CAAC;YAC7B,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;QAC3D,CAAC;QACD,GAAG,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC;IACpB,CAAC;IACD,IAAI,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC;QACpB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YAC9B,MAAM,IAAI,KAAK,CAAC,oCAAoC,CAAC,CAAC;QACxD,CAAC;QACD,GAAG,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC;IACpB,CAAC;IACD,IAAI,CAAC,EAAE,SAAS,IAAI,IAAI,EAAE,CAAC;QACzB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,SAAS,IAAI,CAAC,EAAE,CAAC;YACvD,MAAM,IAAI,KAAK,CAAC,iDAAiD,CAAC,CAAC;QACrE,CAAC;QACD,GAAG,CAAC,SAAS,GAAG,CAAC,CAAC,SAAS,CAAC;IAC9B,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,+EAA+E;AAC/E,gBAAgB;AAChB,+EAA+E;AAE/E;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,KAAK,CAAC;IACf,CAAC;IACD,OAAO,eAAe,CAAC,WAAW,EAAE,CAAC;AACvC,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAC/B,QAAsB,EACtB,OAAwB;IAExB,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;IACtB,CAAC;IAED,IAAI,CAAC,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvC,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;IACpD,CAAC;IAED,IAAI,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;QAC7B,MAAM,IAAI,UAAU,CAAC,qBAAqB,EAAE,EAAE,EAAE,sBAAsB,CAAC,CAAC;IAC1E,CAAC;IAED,0FAA0F;IAC1F,MAAM,gBAAgB,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IACnE,MAAM,YAAY,GAAG,gBAAgB;QACnC,CAAC,CAAC,EAAE,CAAC,oCAAoC;QACzC,CAAC,CAAC,OAAO,EAAE,YAAY,IAAI,qBAAqB,CAAC;IAEnD,gBAAgB,EAAE,CAAC,CAAC,2DAA2D;IAC/E,MAAM,SAAS,GAAG,iBAAiB,EAAE,CAAC;IAEtC,0EAA0E;IAC1E,gFAAgF;IAChF,MAAM,MAAM,GAAG,eAAe,CAAC,WAAW,CAAC,QAAQ,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;IAC9E,MAAM,OAAO,GAAG,GAAG,EAAE;QACnB,iBAAiB,GAAG,KAAK,CAAC;IAC5B,CAAC,CAAC;IACF,MAAM,CAAC,IAAI,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;IAE9B,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,CAAC;IAC/B,IAAI,CAAC,MAAM,EAAE,CAAC;QACZ,IAAI,CAAC;YACH,OAAO,MAAM,MAAM,CAAC;QACtB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,YAAY,CAAC,CAAC,CAAC,CAAC;QAClB,CAAC;IACH,CAAC;IAED,2EAA2E;IAC3E,0EAA0E;IAC1E,sEAAsE;IACtE,OAAO,MAAM,IAAI,OAAO,CAAc,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACxD,IAAI,IAAI,GAAG,KAAK,CAAC;QACjB,MAAM,MAAM,GAAG,CAAC,MAAkB,EAAE,EAAE;YACpC,IAAI,IAAI;gBAAE,OAAO;YACjB,IAAI,GAAG,IAAI,CAAC;YACZ,MAAM,CAAC,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC7C,MAAM,EAAE,CAAC;QACX,CAAC,CAAC;QACF,SAAS,OAAO;YACd,eAAe,CAAC,aAAa,CAAC,SAAS,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YACzD,MAAM,CAAC,GAAG,EAAE,CAAC,MAAM,CAAC,IAAI,UAAU,CAAC,qBAAqB,EAAE,EAAE,EAAE,mBAAmB,CAAC,CAAC,CAAC,CAAC;QACvF,CAAC;QACD,MAAM,CAAC,gBAAgB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QAC1C,MAAM,CAAC,IAAI,CACT,CAAC,CAAC,EAAE,EAAE,CAAC,MAAM,CAAC,GAAG,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAC/B,CAAC,CAAC,EAAE,EAAE,CACJ,MAAM,CAAC,GAAG,EAAE;YACV,IAAI,CAAC;gBACH,YAAY,CAAC,CAAC,CAAC,CAAC;YAClB,CAAC;YAAC,OAAO,EAAE,EAAE,CAAC;gBACZ,MAAM,CAAC,EAAE,CAAC,CAAC;YACb,CAAC;QACH,CAAC,CAAC,CACL,CAAC;IACJ,CAAC,CAAC,CAAC;AACL,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,MAAM,UAAU,aAAa,CAC3B,QAAsB,EACtB,OAA0B,EAC1B,OAA0B;IAE1B,+BAA+B;IAC/B,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;YACtC,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC;SACf,CAAC;IACJ,CAAC;IAED,IAAI,CAAC,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvC,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;YACpE,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC;SACf,CAAC;IACJ,CAAC;IAED,IAAI,iBAAiB,EAAE,CAAC;QACtB,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,MAAM,CACrB,IAAI,UAAU,CACZ,gBAAgB,EAChB,EAAE,EACF,kEAAkE,CACnE,CACF;YACD,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC;SACf,CAAC;IACJ,CAAC;IACD,iBAAiB,GAAG,IAAI,CAAC,CAAC,8CAA8C;IAExE,MAAM,SAAS,GAAG,iBAAiB,EAAE,CAAC;IAEtC,0FAA0F;IAC1F,MAAM,gBAAgB,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IACnE,MAAM,YAAY,GAAG,gBAAgB;QACnC,CAAC,CAAC,EAAE,CAAC,oCAAoC;QACzC,CAAC,CAAC,OAAO,EAAE,YAAY,IAAI,qBAAqB,CAAC;IAEnD,IAAI,SAAS,GAAG,EAAE,CAAC;IACnB,IAAI,OAAO,GAAG,KAAK,CAAC;IACpB,IAAI,YAAwE,CAAC;IAC7E,IAAI,YAAuC,CAAC;IAC5C,IAAI,WAAkC,CAAC;IAEvC,+EAA+E;IAC/E,MAAM,MAAM,GAAG,CAAC,MAAkB,EAAE,EAAE;QACpC,IAAI,OAAO;YAAE,OAAO;QACpB,OAAO,GAAG,IAAI,CAAC;QACf,YAAY,EAAE,MAAM,EAAE,CAAC;QACvB,iBAAiB,GAAG,KAAK,CAAC;QAC1B,MAAM,EAAE,CAAC;IACX,CAAC,CAAC;IAEF,MAAM,OAAO,GAAG,IAAI,OAAO,CAAc,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QAC3D,YAAY,GAAG,OAAO,CAAC;QACvB,WAAW,GAAG,MAAM,CAAC;IACvB,CAAC,CAAC,CAAC;IAEH,YAAY,GAAG,eAAe,CAAC,WAAW,CACxC,eAAe,EACf,CAAC,KAAqB,EAAE,EAAE;QACxB,IAAI,KAAK,CAAC,SAAS,KAAK,SAAS;YAAE,OAAO;QAC1C,SAAS,GAAG,KAAK,CAAC,eAAe,CAAC;QAClC,OAAO,CAAC,KAAK,CAAC,CAAC;QACf,IAAI,KAAK,CAAC,MAAM;YAAE,MAAM,CAAC,GAAG,EAAE,CAAC,YAAY,CAAC,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC;IACpE,CAAC,CACF,CAAC;IAEF,eAAe,CAAC,cAAc,CAAC,QAAQ,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC,KAAK,CACrE,CAAC,KAAK,EAAE,EAAE;QACR,MAAM,CAAC,GAAG,EAAE;YACV,IAAI,CAAC;gBACH,YAAY,CAAC,KAAK,CAAC,CAAC;YACtB,CAAC;YAAC,OAAO,EAAE,EAAE,CAAC;gBACZ,WAAW,CAAC,EAAE,CAAC,CAAC;YAClB,CAAC;QACH,CAAC,CAAC,CAAC;IACL,CAAC,CACF,CAAC;IAEF,MAAM,IAAI,GAAG,GAAG,EAAE;QAChB,6EAA6E;QAC7E,4EAA4E;QAC5E,eAAe,CAAC,aAAa,CAAC,SAAS,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;QACzD,MAAM,CAAC,GAAG,EAAE,CAAC,YAAY,CAAC,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC;IAClD,CAAC,CAAC;IAEF,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;AAC3B,CAAC;AAED,+EAA+E;AAC/E,uBAAuB;AACvB,+EAA+E;AAE/E;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB;IACpC,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,CAAC;IACZ,CAAC;IACD,OAAO,eAAe,CAAC,gBAAgB,EAAE,CAAC;AAC5C,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB;IACzC,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO,EAAE,CAAC;IACZ,CAAC;IAED,MAAM,cAAc,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CACrD,KAAK,CAAC,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAuB,CAAC,CACpE,CAAC;IAEF,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,IAAI,CAAC;QACH,cAAc,GAAG,eAAe,CAAC,iBAAiB,EAAE,CAAC;IACvD,CAAC;IAAC,MAAM,CAAC;QACP,0FAA0F;IAC5F,CAAC;IAED,OAAO,OAAO,CAAC,GAAG,CAChB,cAAc,CAAC,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,EAAE;QACjC,0EAA0E;QAC1E,0DAA0D;QAC1D,MAAM,MAAM,GAAG,MAAM,eAAe,CAAC,0BAA0B,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QAC1E,OAAO;YACL,EAAE,EAAE,KAAK,CAAC,EAAE;YACZ,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,cAAc,EAAE,KAAK,CAAC,cAAc;YACpC,SAAS,EAAE,KAAK,CAAC,SAAS;YAC1B,aAAa,EAAE,KAAK,CAAC,aAAa;YAClC,WAAW,EAAE,KAAK,CAAC,WAAW;YAC9B,iBAAiB,EAAE,cAAc,IAAI,KAAK,CAAC,WAAW;YACtD,MAAM;SACP,CAAC;IACJ,CAAC,CAAC,CACH,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,CAAC,KAAK,UAAU,mBAAmB;IACvC,MAAM,MAAM,GAAG,MAAM,qBAAqB,EAAE,CAAC;IAC7C,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC;IAC3D,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACvC,oFAAoF;IACpF,OAAO,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,GAAG,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;AACnE,CAAC;AAED;;;;;;;;;;;;;;GAcG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,OAAe,EACf,OAAqD;IAErD,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,CAAC,KAAK,EAAE,CAAC;QACX,MAAM,IAAI,UAAU,CAAC,iBAAiB,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,IAAI,CAAC,KAAK,CAAC,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAuB,CAAC,EAAE,CAAC;QACzE,MAAM,IAAI,UAAU,CAClB,sBAAsB,EACtB,OAAO,EACP,SAAS,OAAO,wBAAwB,QAAQ,CAAC,EAAE,EAAE,CACtD,CAAC;IACJ,CAAC;IAED,IAAI,CAAC;QACH,MAAM,cAAc,GAAG,eAAe,CAAC,iBAAiB,EAAE,CAAC;QAC3D,IAAI,cAAc,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;YACvC,MAAM,IAAI,UAAU,CAClB,sBAAsB,EACtB,OAAO,EACP,cAAc,IAAI,CAAC,KAAK,CAAC,cAAc,GAAG,GAAG,CAAC,0BAA0B,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,WAAW,GAAG,GAAG,CAAC,IAAI,CAChH,CAAC;QACJ,CAAC;IACH,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,IAAI,CAAC,YAAY,UAAU;YAAE,MAAM,CAAC,CAAC;QACrC,sDAAsD;IACxD,CAAC;IAED,IAAI,YAAwE,CAAC;IAC7E,IAAI,OAAO,EAAE,UAAU,EAAE,CAAC;QACxB,YAAY,GAAG,eAAe,CAAC,WAAW,CACxC,oBAAoB,EACpB,CAAC,KAAK,EAAE,EAAE;YACR,IAAI,KAAK,CAAC,OAAO,KAAK,OAAO,EAAE,CAAC;gBAC9B,OAAO,CAAC,UAAW,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;YACtC,CAAC;QACH,CAAC,CACF,CAAC;IACJ,CAAC;IAED,IAAI,CAAC;QACH,MAAM,UAAU,CAAC,GAAG,EAAE,CACpB,eAAe,CAAC,aAAa,CAAC,OAAO,EAAE,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,MAAM,CAAC,CACxE,CAAC;IACJ,CAAC;YAAS,CAAC;QACT,YAAY,EAAE,MAAM,EAAE,CAAC;IACzB,CAAC;AACH,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,OAAe;IAClD,IAAI,QAAQ,CAAC,EAAE,KAAK,KAAK,IAAI,QAAQ,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;QACvD,OAAO;IACT,CAAC;IACD,MAAM,UAAU,CAAC,GAAG,EAAE,CAAC,eAAe,CAAC,cAAc,CAAC,OAAO,CAAC,CAAC,CAAC;AAClE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,OAAe;IAC/C,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,CAAC,KAAK,EAAE,CAAC;QACX,MAAM,IAAI,UAAU,CAAC,iBAAiB,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,MAAM,UAAU,CAAC,GAAG,EAAE,CAAC,eAAe,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC;AAC/D,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,MAAM,CAAC,KAAK,UAAU,QAAQ,CAAC,OAAe,EAAE,OAAyB;IACvE,MAAM,KAAK,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,WAAW,GAAG,KAAK,EAAE,WAAW,IAAI,CAAC,CAAC;IAC5C,MAAM,OAAO,GAAG,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;IAC3C,MAAM,UAAU,GAAG,kBAAkB,CAAC,OAAO,EAAE,UAAU,CAAC,CAAC;IAC3D,MAAM,UAAU,CAAC,GAAG,EAAE,CACpB,eAAe,CAAC,QAAQ,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,UAAU,CAAC,CACpE,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc;IAC5B,OAAO,eAAe,CAAC,cAAc,EAAE,CAAC;AAC1C,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,MAAM,UAAU,CAAC,GAAG,EAAE,CAAC,eAAe,CAAC,WAAW,EAAE,CAAC,CAAC;AACxD,CAAC","sourcesContent":["import ExpoAiKitModule, { type NativeGenerationConfig } from './ExpoAiKitModule';\nimport { Platform } from 'react-native';\nimport {\n  LLMMessage,\n  LLMSendOptions,\n  LLMResponse,\n  LLMStreamOptions,\n  LLMStreamEvent,\n  LLMStreamCallback,\n  LLMStreamHandle,\n  BuiltInModel,\n  DownloadableModel,\n  GenerationConfig,\n  ModelError,\n  ModelErrorCode,\n  SetModelOptions,\n} from './types';\nimport { MODEL_REGISTRY, getRegistryEntry } from './models';\n\nexport * from './types';\nexport * from './models';\n\nconst DEFAULT_SYSTEM_PROMPT =\n  'You are a helpful, friendly assistant. Answer the user directly and concisely.';\n\nlet streamIdCounter = 0;\nfunction generateSessionId(): string {\n  return `gen_${Date.now()}_${++streamIdCounter}`;\n}\n\n// The set of codes the native layer encodes in error messages as \"CODE:modelId:reason\".\nconst KNOWN_ERROR_CODES = new Set<ModelErrorCode>([\n  'MODEL_NOT_FOUND',\n  'MODEL_NOT_DOWNLOADED',\n  'DOWNLOAD_FAILED',\n  'DOWNLOAD_CORRUPT',\n  'DOWNLOAD_STORAGE_FULL',\n  'DOWNLOAD_CANCELLED',\n  'INFERENCE_OOM',\n  'INFERENCE_FAILED',\n  'INFERENCE_BUSY',\n  'INFERENCE_CANCELLED',\n  'MODEL_LOAD_FAILED',\n  'DEVICE_NOT_SUPPORTED',\n]);\n\n/**\n * Normalize an error from the native layer into a {@link ModelError}.\n *\n * The native modules format failures as \"CODE:modelId:reason\" (see the\n * GemmaError/GemmaInferenceClient contract). Expo surfaces that string as the\n * error's message, so we parse it here and rethrow a typed ModelError with a\n * reliable `.code` and `.modelId`. Anything unrecognized becomes UNKNOWN.\n */\nfunction toModelError(e: unknown): never {\n  if (e instanceof ModelError) throw e;\n  const message = String((e as any)?.message ?? e ?? '');\n  const match = /^([A-Z_]+):([^:]*):([\\s\\S]*)$/.exec(message);\n  if (match && KNOWN_ERROR_CODES.has(match[1] as ModelErrorCode)) {\n    throw new ModelError(match[1] as ModelErrorCode, match[2], match[3]);\n  }\n  throw new ModelError('UNKNOWN', '', message);\n}\n\n/** Run a native promise, normalizing any rejection into a ModelError. */\nasync function wrapNative<T>(run: () => Promise<T>): Promise<T> {\n  try {\n    return await run();\n  } catch (e) {\n    toModelError(e);\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Single-flight inference guard\n// ---------------------------------------------------------------------------\n// On-device models are backed by a single native context + KV cache that is not\n// safe for concurrent decodes (interleaving can corrupt the cache and crash the\n// native side). JS is single-threaded, so a synchronous check-and-set of this\n// flag before any `await` is race-free. The flag is shared by sendMessage and\n// streamMessage and is held until the *native* call settles — not until an\n// early abort — so a detached-but-still-running generation still blocks a new one.\nlet inferenceInFlight = false;\n\nfunction acquireInference(): void {\n  if (inferenceInFlight) {\n    throw new ModelError(\n      'INFERENCE_BUSY',\n      '',\n      'A generation is already in flight. Wait for it to finish, or stop the active stream first.'\n    );\n  }\n  inferenceInFlight = true;\n}\n\n/**\n * Map the public GenerationConfig to the native shape, dropping undefined fields\n * and validating ranges up front so callers get a clear error instead of an\n * opaque native MODEL_LOAD_FAILED from the sampler.\n */\nfunction toNativeGeneration(g?: GenerationConfig): NativeGenerationConfig {\n  const out: NativeGenerationConfig = {};\n  if (g?.temperature != null) {\n    if (g.temperature < 0) {\n      throw new Error('generation.temperature must be >= 0');\n    }\n    out.temperature = g.temperature;\n  }\n  if (g?.topK != null) {\n    if (!Number.isInteger(g.topK) || g.topK <= 0) {\n      throw new Error('generation.topK must be a positive integer');\n    }\n    out.topK = g.topK;\n  }\n  if (g?.topP != null) {\n    if (g.topP < 0 || g.topP > 1) {\n      throw new Error('generation.topP must be within [0, 1]');\n    }\n    out.topP = g.topP;\n  }\n  if (g?.seed != null) {\n    if (!Number.isInteger(g.seed)) {\n      throw new Error('generation.seed must be an integer');\n    }\n    out.seed = g.seed;\n  }\n  if (g?.maxTokens != null) {\n    if (!Number.isInteger(g.maxTokens) || g.maxTokens <= 0) {\n      throw new Error('generation.maxTokens must be a positive integer');\n    }\n    out.maxTokens = g.maxTokens;\n  }\n  return out;\n}\n\n// ============================================================================\n// Inference API\n// ============================================================================\n\n/**\n * Check if on-device AI is available on the current device.\n * Returns false on unsupported platforms (web, etc.).\n */\nexport async function isAvailable(): Promise<boolean> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return false;\n  }\n  return ExpoAiKitModule.isAvailable();\n}\n\n/**\n * Send messages to the on-device LLM and get a response.\n *\n * @param messages - Array of messages representing the conversation\n * @param options - Optional settings (systemPrompt fallback)\n * @returns Promise with the generated response\n *\n * @example\n * ```ts\n * const response = await sendMessage([\n *   { role: 'user', content: 'What is 2 + 2?' }\n * ]);\n * console.log(response.text); // \"4\"\n * ```\n *\n * @example\n * ```ts\n * // With system prompt\n * const response = await sendMessage(\n *   [{ role: 'user', content: 'Hello!' }],\n *   { systemPrompt: 'You are a pirate. Respond in pirate speak.' }\n * );\n * ```\n *\n * @example\n * ```ts\n * // Multi-turn conversation\n * const response = await sendMessage([\n *   { role: 'system', content: 'You are a helpful assistant.' },\n *   { role: 'user', content: 'My name is Alice.' },\n *   { role: 'assistant', content: 'Nice to meet you, Alice!' },\n *   { role: 'user', content: 'What is my name?' }\n * ]);\n * ```\n */\nexport async function sendMessage(\n  messages: LLMMessage[],\n  options?: LLMSendOptions\n): Promise<LLMResponse> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return { text: '' };\n  }\n\n  if (!messages || messages.length === 0) {\n    throw new Error('messages array cannot be empty');\n  }\n\n  if (options?.signal?.aborted) {\n    throw new ModelError('INFERENCE_CANCELLED', '', 'Aborted before start');\n  }\n\n  // Determine system prompt: use from messages array if present, else options, else default\n  const hasSystemMessage = messages.some((m) => m.role === 'system');\n  const systemPrompt = hasSystemMessage\n    ? '' // Native will extract from messages\n    : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;\n\n  acquireInference(); // throws INFERENCE_BUSY if a generation is already running\n  const sessionId = generateSessionId();\n\n  // Hold the single-flight flag until the NATIVE call settles — even if the\n  // caller aborts early — because the model may keep computing in the background.\n  const native = ExpoAiKitModule.sendMessage(messages, systemPrompt, sessionId);\n  const release = () => {\n    inferenceInFlight = false;\n  };\n  native.then(release, release);\n\n  const signal = options?.signal;\n  if (!signal) {\n    try {\n      return await native;\n    } catch (e) {\n      toModelError(e);\n    }\n  }\n\n  // Race the native result against the abort signal. On abort we unblock the\n  // caller immediately and best-effort ask native to cancel; the flag stays\n  // held (via `release` above) until the native call actually finishes.\n  return await new Promise<LLMResponse>((resolve, reject) => {\n    let done = false;\n    const finish = (action: () => void) => {\n      if (done) return;\n      done = true;\n      signal.removeEventListener('abort', onAbort);\n      action();\n    };\n    function onAbort() {\n      ExpoAiKitModule.stopStreaming(sessionId).catch(() => {});\n      finish(() => reject(new ModelError('INFERENCE_CANCELLED', '', 'Aborted by caller')));\n    }\n    signal.addEventListener('abort', onAbort);\n    native.then(\n      (r) => finish(() => resolve(r)),\n      (e) =>\n        finish(() => {\n          try {\n            toModelError(e);\n          } catch (me) {\n            reject(me);\n          }\n        })\n    );\n  });\n}\n\n/**\n * Stream messages to the on-device LLM and receive progressive token updates.\n *\n * @param messages - Array of messages representing the conversation\n * @param onToken - Callback function called for each token/chunk received\n * @param options - Optional settings (systemPrompt fallback)\n * @returns Object with stop() function to cancel streaming and promise that resolves when complete\n *\n * @example\n * ```ts\n * // Basic streaming\n * const { promise } = streamMessage(\n *   [{ role: 'user', content: 'Tell me a story' }],\n *   (event) => {\n *     console.log(event.token); // Each token as it arrives\n *     console.log(event.accumulatedText); // Full text so far\n *   }\n * );\n * await promise;\n * ```\n *\n * @example\n * ```ts\n * // With cancellation\n * const { promise, stop } = streamMessage(\n *   [{ role: 'user', content: 'Write a long essay' }],\n *   (event) => setText(event.accumulatedText)\n * );\n *\n * // Cancel after 5 seconds\n * setTimeout(() => stop(), 5000);\n * ```\n */\nexport function streamMessage(\n  messages: LLMMessage[],\n  onToken: LLMStreamCallback,\n  options?: LLMStreamOptions\n): LLMStreamHandle {\n  // Handle unsupported platforms\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return {\n      promise: Promise.resolve({ text: '' }),\n      stop: () => {},\n    };\n  }\n\n  if (!messages || messages.length === 0) {\n    return {\n      promise: Promise.reject(new Error('messages array cannot be empty')),\n      stop: () => {},\n    };\n  }\n\n  if (inferenceInFlight) {\n    return {\n      promise: Promise.reject(\n        new ModelError(\n          'INFERENCE_BUSY',\n          '',\n          'A generation is already in flight. Stop the active stream first.'\n        )\n      ),\n      stop: () => {},\n    };\n  }\n  inferenceInFlight = true; // set synchronously — race-free with other JS\n\n  const sessionId = generateSessionId();\n\n  // Determine system prompt: use from messages array if present, else options, else default\n  const hasSystemMessage = messages.some((m) => m.role === 'system');\n  const systemPrompt = hasSystemMessage\n    ? '' // Native will extract from messages\n    : options?.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;\n\n  let finalText = '';\n  let settled = false;\n  let subscription: ReturnType<typeof ExpoAiKitModule.addListener> | undefined;\n  let resolveOuter!: (r: LLMResponse) => void;\n  let rejectOuter!: (e: unknown) => void;\n\n  // Settle exactly once: remove the listener and release the single-flight flag.\n  const settle = (action: () => void) => {\n    if (settled) return;\n    settled = true;\n    subscription?.remove();\n    inferenceInFlight = false;\n    action();\n  };\n\n  const promise = new Promise<LLMResponse>((resolve, reject) => {\n    resolveOuter = resolve;\n    rejectOuter = reject;\n  });\n\n  subscription = ExpoAiKitModule.addListener(\n    'onStreamToken',\n    (event: LLMStreamEvent) => {\n      if (event.sessionId !== sessionId) return;\n      finalText = event.accumulatedText;\n      onToken(event);\n      if (event.isDone) settle(() => resolveOuter({ text: finalText }));\n    }\n  );\n\n  ExpoAiKitModule.startStreaming(messages, systemPrompt, sessionId).catch(\n    (error) => {\n      settle(() => {\n        try {\n          toModelError(error);\n        } catch (me) {\n          rejectOuter(me);\n        }\n      });\n    }\n  );\n\n  const stop = () => {\n    // Best-effort native cancel (native also emits a terminal isDone on cancel),\n    // but resolve immediately with the text so far so `promise` can never hang.\n    ExpoAiKitModule.stopStreaming(sessionId).catch(() => {});\n    settle(() => resolveOuter({ text: finalText }));\n  };\n\n  return { promise, stop };\n}\n\n// ============================================================================\n// Model Management API\n// ============================================================================\n\n/**\n * Get all built-in models available on the current platform.\n *\n * Built-in models are provided by the OS and require no download.\n * On iOS this returns Apple Foundation Models; on Android, ML Kit.\n *\n * @returns Array of built-in models with availability status\n */\nexport async function getBuiltInModels(): Promise<BuiltInModel[]> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return [];\n  }\n  return ExpoAiKitModule.getBuiltInModels();\n}\n\n/**\n * Get all downloadable models from the registry, enriched with on-device status.\n *\n * Reads from the hardcoded MODEL_REGISTRY and queries the native layer\n * for the current download/load status of each model.\n *\n * @returns Array of downloadable models with their current status\n */\nexport async function getDownloadableModels(): Promise<DownloadableModel[]> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return [];\n  }\n\n  const platformModels = MODEL_REGISTRY.filter((entry) =>\n    entry.supportedPlatforms.includes(Platform.OS as 'ios' | 'android')\n  );\n\n  let deviceRamBytes = 0;\n  try {\n    deviceRamBytes = ExpoAiKitModule.getDeviceRamBytes();\n  } catch {\n    // Native call unavailable -- default to 0 (all models will show meetsRequirements: false)\n  }\n\n  return Promise.all(\n    platformModels.map(async (entry) => {\n      // Await: on iOS this bridges as a Promise (reads actor state); on Android\n      // it's synchronous and awaiting a plain value is a no-op.\n      const status = await ExpoAiKitModule.getDownloadableModelStatus(entry.id);\n      return {\n        id: entry.id,\n        name: entry.name,\n        parameterCount: entry.parameterCount,\n        sizeBytes: entry.sizeBytes,\n        contextWindow: entry.contextWindow,\n        minRamBytes: entry.minRamBytes,\n        meetsRequirements: deviceRamBytes >= entry.minRamBytes,\n        status,\n      };\n    })\n  );\n}\n\n/**\n * Pick the best downloadable model the current device can run.\n *\n * Returns the most capable model (largest, by RAM requirement) whose\n * `meetsRequirements` is true — e.g. Gemma 4 E4B on high-spec phones, falling\n * back to E2B on more constrained ones — or `null` if the device can't run any.\n *\n * This is a convenience over {@link getDownloadableModels}; the caller still\n * downloads + activates explicitly. Pass `platform` is implicit (current OS).\n *\n * @example\n * ```ts\n * const best = await getRecommendedModel();\n * if (best) {\n *   await downloadModel(best.id, { onProgress });\n *   await setModel(best.id);\n * }\n * ```\n */\nexport async function getRecommendedModel(): Promise<DownloadableModel | null> {\n  const models = await getDownloadableModels();\n  const runnable = models.filter((m) => m.meetsRequirements);\n  if (runnable.length === 0) return null;\n  // Higher RAM requirement ⇒ larger/more capable model. Prefer the biggest that fits.\n  return runnable.sort((a, b) => b.minRamBytes - a.minRamBytes)[0];\n}\n\n/**\n * Download a model to the device.\n *\n * Looks up the model in the registry, validates platform support and\n * device requirements, then initiates the download with integrity verification.\n *\n * @param modelId - ID of the model to download (e.g. 'gemma-e2b')\n * @param options - Optional download configuration\n * @param options.onProgress - Callback with download progress (0-1)\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is not in the registry\n * @throws {ModelError} DEVICE_NOT_SUPPORTED if platform is not supported\n * @throws {ModelError} DOWNLOAD_FAILED on network error\n * @throws {ModelError} DOWNLOAD_STORAGE_FULL if insufficient disk space\n * @throws {ModelError} DOWNLOAD_CORRUPT if SHA256 hash doesn't match\n */\nexport async function downloadModel(\n  modelId: string,\n  options?: { onProgress?: (progress: number) => void }\n): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  if (!entry) {\n    throw new ModelError('MODEL_NOT_FOUND', modelId);\n  }\n\n  if (!entry.supportedPlatforms.includes(Platform.OS as 'ios' | 'android')) {\n    throw new ModelError(\n      'DEVICE_NOT_SUPPORTED',\n      modelId,\n      `Model ${modelId} is not supported on ${Platform.OS}`\n    );\n  }\n\n  try {\n    const deviceRamBytes = ExpoAiKitModule.getDeviceRamBytes();\n    if (deviceRamBytes < entry.minRamBytes) {\n      throw new ModelError(\n        'DEVICE_NOT_SUPPORTED',\n        modelId,\n        `Device has ${Math.round(deviceRamBytes / 1e9)}GB RAM, model requires ${Math.round(entry.minRamBytes / 1e9)}GB`\n      );\n    }\n  } catch (e) {\n    if (e instanceof ModelError) throw e;\n    // If getDeviceRamBytes is unavailable, skip the check\n  }\n\n  let subscription: ReturnType<typeof ExpoAiKitModule.addListener> | undefined;\n  if (options?.onProgress) {\n    subscription = ExpoAiKitModule.addListener(\n      'onDownloadProgress',\n      (event) => {\n        if (event.modelId === modelId) {\n          options.onProgress!(event.progress);\n        }\n      }\n    );\n  }\n\n  try {\n    await wrapNative(() =>\n      ExpoAiKitModule.downloadModel(modelId, entry.downloadUrl, entry.sha256)\n    );\n  } finally {\n    subscription?.remove();\n  }\n}\n\n/**\n * Cancel an in-flight download for a model.\n *\n * The in-progress {@link downloadModel} promise rejects with a\n * DOWNLOAD_CANCELLED {@link ModelError}. No-op if the model isn't downloading.\n *\n * @param modelId - ID of the model whose download should be cancelled\n */\nexport async function cancelDownload(modelId: string): Promise<void> {\n  if (Platform.OS !== 'ios' && Platform.OS !== 'android') {\n    return;\n  }\n  await wrapNative(() => ExpoAiKitModule.cancelDownload(modelId));\n}\n\n/**\n * Delete a downloaded model from the device.\n *\n * If the model is currently loaded, it will be unloaded first.\n *\n * @param modelId - ID of the model to delete\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is not in the registry\n */\nexport async function deleteModel(modelId: string): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  if (!entry) {\n    throw new ModelError('MODEL_NOT_FOUND', modelId);\n  }\n\n  await wrapNative(() => ExpoAiKitModule.deleteModel(modelId));\n}\n\n/**\n * Set the active model for inference.\n *\n * This is the sole gatekeeper for model validity. If setModel succeeds,\n * the model is loaded and ready -- sendMessage never needs its own check.\n *\n * For downloadable models, this loads the model into memory (status\n * transitions: loading -> ready). Only one downloadable model can be\n * loaded at a time; the previous one is auto-unloaded.\n *\n * For built-in models, this simply switches the active backend.\n *\n * If setModel was never called, sendMessage uses the platform built-in\n * model (today's behavior, no error).\n *\n * @param modelId - ID of the model to activate (e.g. 'gemma-e2b', 'apple-fm', 'mlkit')\n * @param options - Optional configuration for model loading\n * @param options.backend - Hardware backend: 'auto' (default, GPU with CPU fallback), 'gpu', or 'cpu'\n * @throws {ModelError} MODEL_NOT_FOUND if modelId is invalid\n * @throws {ModelError} MODEL_NOT_DOWNLOADED if the downloadable model file is not on disk\n * @throws {ModelError} MODEL_LOAD_FAILED if loading into memory fails\n * @throws {ModelError} INFERENCE_OOM if device can't fit model in memory\n */\nexport async function setModel(modelId: string, options?: SetModelOptions): Promise<void> {\n  const entry = getRegistryEntry(modelId);\n  const minRamBytes = entry?.minRamBytes ?? 0;\n  const backend = options?.backend ?? 'auto';\n  const generation = toNativeGeneration(options?.generation);\n  await wrapNative(() =>\n    ExpoAiKitModule.setModel(modelId, minRamBytes, backend, generation)\n  );\n}\n\n/**\n * Get the ID of the currently active model.\n *\n * @returns The active model ID (e.g. 'apple-fm', 'mlkit', 'gemma-e2b')\n */\nexport function getActiveModel(): string {\n  return ExpoAiKitModule.getActiveModel();\n}\n\n/**\n * Explicitly unload the current downloadable model from memory.\n *\n * Frees memory and reverts to the platform built-in model.\n * No-op if no downloadable model is currently loaded.\n */\nexport async function unloadModel(): Promise<void> {\n  await wrapNative(() => ExpoAiKitModule.unloadModel());\n}\n\n"]}

package/build/models.js

@@ -12,13 +12,13 @@ export const MODEL_REGISTRY = [
         parameterCount: '2.3B',
         quantization: 'mixed-2/4/8-bit',
         downloadUrl: 'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',
-        sha256: '', // TODO: Fill with actual hash once model file is verified
-        sizeBytes: 2_580_000_000, // ~2.58GB
+        sha256: '181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c',
+        sizeBytes: 2_588_147_712, // 2.59GB (exact, HF LFS)
         // Conservative limit for 4GB RAM devices.
         // TODO: Benchmark during Phase 2 testing.
         contextWindow: 8_000,
         minRamBytes: 2_000_000_000, // 2GB — LiteRT-LM memory-maps weights, actual RSS ~1.5GB
-        supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
+        supportedPlatforms: ['ios', 'android'],
     },
     {
         id: 'gemma-e4b',
@@ -26,11 +26,11 @@ export const MODEL_REGISTRY = [
         parameterCount: '4.5B',
         quantization: 'mixed-4/8-bit',
         downloadUrl: 'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',
-        sha256: '', // TODO: Fill with actual hash once model file is verified
-        sizeBytes: 3_650_000_000, // ~3.65GB
+        sha256: '0b2a8980ce155fd97673d8e820b4d29d9c7d99b8fa6806f425d969b145bd52e0',
+        sizeBytes: 3_659_530_240, // 3.66GB (exact, HF LFS)
         contextWindow: 16_000,
         minRamBytes: 3_000_000_000, // 3GB — LiteRT-LM memory-maps weights
-        supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
+        supportedPlatforms: ['ios', 'android'],
     },
 ];
 /**

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,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,UAAU;QACpC,0CAA0C;QAC1C,0CAA0C;QAC1C,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,yDAAyD;QACrF,kBAAkB,EAAE,CAAC,SAAS,CAAC,EAAE,uCAAuC;KACzE;IACD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,eAAe;QAC7B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,UAAU;QACpC,aAAa,EAAE,MAAM;QACrB,WAAW,EAAE,aAAa,EAAE,sCAAsC;QAClE,kBAAkB,EAAE,CAAC,SAAS,CAAC,EAAE,uCAAuC;KACzE;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: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 2_580_000_000, // ~2.58GB\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: ['android'], // iOS waiting for LiteRT-LM Swift APIs\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: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 3_650_000_000, // ~3.65GB\n    contextWindow: 16_000,\n    minRamBytes: 3_000_000_000, // 3GB — LiteRT-LM memory-maps weights\n    supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs\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;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"]}

package/build/types.d.ts

@@ -6,12 +6,53 @@
  * - 'cpu': Force CPU — slower (~2-5 tok/s) but works on low-RAM devices
  */
 export type InferenceBackend = 'auto' | 'gpu' | 'cpu';
+/**
+ * Sampling / generation parameters applied to a model session.
+ *
+ * Support is per-backend (on-device runtimes expose different knobs), so these
+ * are best-effort — unsupported fields are ignored rather than erroring:
+ *
+ * | field       | Gemma (LiteRT-LM) | Apple Foundation Models | ML Kit |
+ * |-------------|:-----------------:|:-----------------------:|:------:|
+ * | temperature | ✓                 | ✓                       | —      |
+ * | topK        | ✓                 | —                       | —      |
+ * | topP        | ✓                 | —                       | —      |
+ * | seed        | ✓ (iOS only)      | —                       | —      |
+ * | maxTokens   | —                 | ✓ (max output)          | —      |
+ *
+ * Notes:
+ * - Gemma/LiteRT-LM has no per-generation output-token cap (its `maxNumTokens`
+ *   is the total KV-cache size, set at load), so `maxTokens` is not honored
+ *   there. Its sampler (topK/topP/temperature[/seed]) is fixed at conversation
+ *   creation, which is why generation config lives on setModel() rather than
+ *   per-call. `seed` is currently wired on iOS only.
+ * - The ML Kit built-in (Android default) does not yet apply generation config;
+ *   it uses its own defaults.
+ */
+export type GenerationConfig = {
+    /** Sampling temperature. Lower = more deterministic. Typically 0.0–2.0. */
+    temperature?: number;
+    /** Nucleus sampling: number of top logits to consider. Must be > 0. */
+    topK?: number;
+    /** Nucleus sampling: cumulative probability threshold in [0, 1]. */
+    topP?: number;
+    /** RNG seed for reproducible sampling (Gemma only). */
+    seed?: number;
+    /** Maximum number of output tokens to generate (Apple FM / ML Kit only). */
+    maxTokens?: number;
+};
 /**
  * Options for setModel.
  */
 export type SetModelOptions = {
     /** Hardware backend to use for inference. Defaults to 'auto'. */
     backend?: InferenceBackend;
+    /**
+     * Default sampling parameters for this model session. Applied when the model
+     * is activated and used for all subsequent sendMessage/streamMessage calls
+     * until the next setModel(). See {@link GenerationConfig} for per-backend support.
+     */
+    generation?: GenerationConfig;
 };
 /**
  * Role in a conversation message.
@@ -33,6 +74,17 @@ export type LLMSendOptions = {
      * If a system message exists in the array, this is ignored.
      */
     systemPrompt?: string;
+    /**
+     * Abort the request. When the signal fires, the returned promise rejects with
+     * an INFERENCE_CANCELLED {@link ModelError}.
+     *
+     * Note: on-device, non-streaming generation cannot always be interrupted
+     * mid-decode — abort always unblocks the caller, but the model may keep
+     * computing in the background until it finishes, during which a new
+     * sendMessage/streamMessage will throw INFERENCE_BUSY. To truly interrupt a
+     * long generation, prefer streamMessage().stop().
+     */
+    signal?: AbortSignal;
 };
 /**
  * Response from sendMessage.
@@ -51,6 +103,15 @@ export type LLMStreamOptions = {
      */
     systemPrompt?: string;
 };
+/**
+ * Handle returned by streamMessage.
+ */
+export type LLMStreamHandle = {
+    /** Resolves with the final text when streaming completes or is stopped. */
+    promise: Promise<LLMResponse>;
+    /** Stop streaming. Resolves `promise` with the text accumulated so far. */
+    stop: () => void;
+};
 /**
  * Event payload for streaming tokens.
  */
@@ -111,14 +172,17 @@ export type DownloadableModel = {
  *
  * - 'not-downloaded': Model file is not on disk
  * - 'downloading': Model file is being downloaded
+ * - 'downloaded': Model file is on disk but not loaded into memory. Call
+ *   setModel() to load it. This survives app restarts, so use it to decide
+ *   whether a (re-)download is needed.
  * - 'loading': File is on disk, model is being loaded into memory for inference
  * - 'ready': Model is loaded in memory and ready for inference
  */
-export type DownloadableModelStatus = 'not-downloaded' | 'downloading' | 'loading' | 'ready';
+export type DownloadableModelStatus = 'not-downloaded' | 'downloading' | 'downloaded' | 'loading' | 'ready';
 /**
  * Error codes for model-related operations.
  */
-export type ModelErrorCode = 'MODEL_NOT_FOUND' | 'MODEL_NOT_DOWNLOADED' | 'DOWNLOAD_FAILED' | 'DOWNLOAD_CORRUPT' | 'DOWNLOAD_STORAGE_FULL' | 'INFERENCE_OOM' | 'INFERENCE_FAILED' | 'MODEL_LOAD_FAILED' | 'DEVICE_NOT_SUPPORTED';
+export type ModelErrorCode = 'MODEL_NOT_FOUND' | 'MODEL_NOT_DOWNLOADED' | 'DOWNLOAD_FAILED' | 'DOWNLOAD_CORRUPT' | 'DOWNLOAD_STORAGE_FULL' | 'DOWNLOAD_CANCELLED' | 'INFERENCE_OOM' | 'INFERENCE_FAILED' | 'INFERENCE_BUSY' | 'INFERENCE_CANCELLED' | 'MODEL_LOAD_FAILED' | 'DEVICE_NOT_SUPPORTED' | 'UNKNOWN';
 /**
  * Structured error for model operations.
  */