expo-ai-kit 0.9.0 → 0.11.0

package/build/ai/index.js

@@ -0,0 +1,273 @@
+import { embed, getActiveModel, sendMessage, setModel, streamMessage } from '../index';
+import { ModelError } from '../types';
+import { convertCallOptions, extractOutput, EMPTY_USAGE, newPartId } from './convert';
+// ---------------------------------------------------------------------------
+// Vercel AI SDK provider for expo-ai-kit — `import { expoAiKit } from 'expo-ai-kit/ai'`.
+//
+// Implements the LanguageModelV3 spec (AI SDK 6; also accepted by AI SDK 7),
+// wrapping the same core sendMessage/streamMessage/embed calls as the rest of
+// the library — the single-flight INFERENCE_BUSY guard, stateless-model
+// semantics, and typed ModelError contract all apply unchanged.
+//
+// Zero-dependency by construction: everything imported from '@ai-sdk/provider'
+// is a *type* (erased at compile time). The package is only needed by consumers
+// who use this entry point, as an optional peer for its typings.
+// ---------------------------------------------------------------------------
+const PROVIDER = 'expo-ai-kit';
+/**
+ * Sentinel model id meaning "whatever model is currently active" — the OS
+ * built-in by default, or whatever the app last activated via setModel().
+ * The provider never switches models for this id.
+ */
+export const AUTO_MODEL_ID = 'auto';
+/** Model id for the built-in embedding model (Apple NLContextualEmbedding, iOS 17+). */
+export const EMBEDDING_MODEL_ID = 'apple-nl-contextual';
+/**
+ * Create an expo-ai-kit provider for the Vercel AI SDK.
+ *
+ * ```ts
+ * import { generateText, streamText } from 'ai';
+ * import { expoAiKit } from 'expo-ai-kit/ai';
+ *
+ * const { text } = await generateText({
+ *   model: expoAiKit(),               // the active on-device model
+ *   prompt: 'Capital of France?',
+ * });
+ *
+ * const result = streamText({
+ *   model: expoAiKit('gemma-e2b'),    // activates gemma-e2b first (must be downloaded)
+ *   prompt: 'Write a short story',
+ * });
+ * ```
+ *
+ * On-device caveats (all documented in the guide):
+ * - One generation at a time — concurrent calls reject with INFERENCE_BUSY.
+ * - Per-call sampling (temperature, topK, …) is reported as an unsupported-
+ *   setting warning; sampling is fixed when the model is activated.
+ * - Tool calling and JSON output ride the same prompt protocol as
+ *   generateText()/generateObject() — single-shot here, since the AI SDK owns
+ *   the loop. Streaming buffers when tools/JSON are requested (the envelope
+ *   must be parsed whole, not surfaced as text deltas).
+ * - embeddingModel() is iOS-only for now; Android throws DEVICE_NOT_SUPPORTED.
+ */
+export function createExpoAiKit() {
+    const languageModel = (modelId = AUTO_MODEL_ID, settings) => createLanguageModel(modelId, settings);
+    const embeddingModel = (modelId = EMBEDDING_MODEL_ID) => {
+        if (modelId !== EMBEDDING_MODEL_ID) {
+            throw new ModelError('MODEL_NOT_FOUND', modelId, `expo-ai-kit has one embedding model: "${EMBEDDING_MODEL_ID}" (Apple NLContextualEmbedding).`);
+        }
+        return createEmbeddingModel();
+    };
+    const provider = languageModel;
+    provider.languageModel = languageModel;
+    provider.embeddingModel = embeddingModel;
+    provider.textEmbeddingModel = embeddingModel;
+    provider.imageModel = (modelId) => {
+        throw new ModelError('MODEL_NOT_FOUND', modelId, 'expo-ai-kit does not provide image models.');
+    };
+    return provider;
+}
+/** The default provider instance. */
+export const expoAiKit = createExpoAiKit();
+// ---------------------------------------------------------------------------
+// Language model
+// ---------------------------------------------------------------------------
+function createLanguageModel(modelId, settings) {
+    /**
+     * Make sure this instance's model is the active one. 'auto' never switches.
+     * Deliberately skipped when already active: reloading a multi-GB model per
+     * call would be catastrophic, so settings apply on activation only.
+     */
+    const ensureModel = async () => {
+        if (modelId === AUTO_MODEL_ID) {
+            try {
+                return getActiveModel();
+            }
+            catch {
+                return AUTO_MODEL_ID; // native module unavailable (e.g. web) — let the call itself fail
+            }
+        }
+        let active;
+        try {
+            active = getActiveModel();
+        }
+        catch {
+            active = undefined;
+        }
+        if (active !== modelId) {
+            await setModel(modelId, settings);
+        }
+        return modelId;
+    };
+    return {
+        specificationVersion: 'v3',
+        provider: PROVIDER,
+        modelId,
+        supportedUrls: {}, // no URLs are consumed natively — text-only models
+        async doGenerate(options) {
+            const call = convertCallOptions(options);
+            const activeModelId = await ensureModel();
+            const { text } = await sendMessage(call.messages, { signal: options.abortSignal });
+            const output = extractOutput(text, call.toolNames, call.jsonMode);
+            const content = output.kind === 'tool-call'
+                ? [
+                    {
+                        type: 'tool-call',
+                        toolCallId: newPartId('call'),
+                        toolName: output.toolName,
+                        input: output.input,
+                    },
+                ]
+                : [{ type: 'text', text: output.text }];
+            return {
+                content,
+                finishReason: {
+                    unified: output.kind === 'tool-call' ? 'tool-calls' : 'stop',
+                    raw: undefined,
+                },
+                usage: EMPTY_USAGE,
+                warnings: call.warnings,
+                response: { modelId: activeModelId, timestamp: new Date() },
+            };
+        },
+        async doStream(options) {
+            if (typeof ReadableStream === 'undefined') {
+                throw new Error('expo-ai-kit/ai: streaming needs a ReadableStream polyfill in React Native ' +
+                    '(e.g. web-streams-polyfill) — the AI SDK requires the same polyfills. See the docs.');
+            }
+            const call = convertCallOptions(options);
+            const activeModelId = await ensureModel();
+            const abortSignal = options.abortSignal;
+            // Tool calling / JSON mode: the envelope must be parsed as a whole — a
+            // half-streamed `{"tool": …` would surface as garbage text deltas — so
+            // those runs buffer and emit the parsed result in one go.
+            if (call.toolNames.length > 0 || call.jsonMode) {
+                const stream = new ReadableStream({
+                    start: async (controller) => {
+                        controller.enqueue({ type: 'stream-start', warnings: call.warnings });
+                        controller.enqueue({
+                            type: 'response-metadata',
+                            modelId: activeModelId,
+                            timestamp: new Date(),
+                        });
+                        try {
+                            const { text } = await sendMessage(call.messages, { signal: abortSignal });
+                            const output = extractOutput(text, call.toolNames, call.jsonMode);
+                            if (output.kind === 'tool-call') {
+                                const id = newPartId('call');
+                                controller.enqueue({ type: 'tool-input-start', id, toolName: output.toolName });
+                                controller.enqueue({ type: 'tool-input-delta', id, delta: output.input });
+                                controller.enqueue({ type: 'tool-input-end', id });
+                                controller.enqueue({
+                                    type: 'tool-call',
+                                    toolCallId: id,
+                                    toolName: output.toolName,
+                                    input: output.input,
+                                });
+                                controller.enqueue({
+                                    type: 'finish',
+                                    usage: EMPTY_USAGE,
+                                    finishReason: { unified: 'tool-calls', raw: undefined },
+                                });
+                            }
+                            else {
+                                const id = newPartId('text');
+                                controller.enqueue({ type: 'text-start', id });
+                                controller.enqueue({ type: 'text-delta', id, delta: output.text });
+                                controller.enqueue({ type: 'text-end', id });
+                                controller.enqueue({
+                                    type: 'finish',
+                                    usage: EMPTY_USAGE,
+                                    finishReason: { unified: 'stop', raw: undefined },
+                                });
+                            }
+                            controller.close();
+                        }
+                        catch (error) {
+                            controller.error(error);
+                        }
+                    },
+                });
+                return { stream };
+            }
+            // Plain text: real token streaming over streamMessage().
+            let handle;
+            let settled = false;
+            const stream = new ReadableStream({
+                start: (controller) => {
+                    controller.enqueue({ type: 'stream-start', warnings: call.warnings });
+                    controller.enqueue({
+                        type: 'response-metadata',
+                        modelId: activeModelId,
+                        timestamp: new Date(),
+                    });
+                    const textId = newPartId('text');
+                    let textStarted = false;
+                    const onAbort = () => {
+                        if (settled)
+                            return;
+                        settled = true;
+                        handle?.stop(); // stops native generation; promise resolves with partial text
+                        controller.error(new ModelError('INFERENCE_CANCELLED', '', 'Aborted by caller'));
+                    };
+                    abortSignal?.addEventListener('abort', onAbort, { once: true });
+                    handle = streamMessage(call.messages, (event) => {
+                        if (settled)
+                            return;
+                        if (event.token !== '') {
+                            if (!textStarted) {
+                                textStarted = true;
+                                controller.enqueue({ type: 'text-start', id: textId });
+                            }
+                            controller.enqueue({ type: 'text-delta', id: textId, delta: event.token });
+                        }
+                        if (event.isDone) {
+                            settled = true;
+                            abortSignal?.removeEventListener('abort', onAbort);
+                            if (textStarted)
+                                controller.enqueue({ type: 'text-end', id: textId });
+                            controller.enqueue({
+                                type: 'finish',
+                                usage: EMPTY_USAGE,
+                                finishReason: { unified: 'stop', raw: undefined },
+                            });
+                            controller.close();
+                        }
+                    });
+                    handle.promise.catch((error) => {
+                        if (settled)
+                            return;
+                        settled = true;
+                        abortSignal?.removeEventListener('abort', onAbort);
+                        controller.error(error);
+                    });
+                },
+                cancel: () => {
+                    settled = true;
+                    handle?.stop();
+                },
+            });
+            return { stream };
+        },
+    };
+}
+// ---------------------------------------------------------------------------
+// Embedding model
+// ---------------------------------------------------------------------------
+function createEmbeddingModel() {
+    return {
+        specificationVersion: 'v3',
+        provider: PROVIDER,
+        modelId: EMBEDDING_MODEL_ID,
+        maxEmbeddingsPerCall: Infinity, // embed() batches internally; no API limit
+        supportsParallelCalls: true, // embeddings bypass the single-flight inference guard
+        async doEmbed({ values, abortSignal }) {
+            if (abortSignal?.aborted) {
+                throw new ModelError('INFERENCE_CANCELLED', '', 'Aborted before start');
+            }
+            const { embeddings } = await embed(values);
+            return { embeddings, warnings: [] };
+        },
+    };
+}
+//# sourceMappingURL=index.js.map

package/build/ai/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/ai/index.ts"],"names":[],"mappings":"AASA,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE,WAAW,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACvF,OAAO,EAAE,UAAU,EAAwB,MAAM,UAAU,CAAC;AAC5D,OAAO,EAAE,kBAAkB,EAAE,aAAa,EAAE,WAAW,EAAE,SAAS,EAAE,MAAM,WAAW,CAAC;AAEtF,8EAA8E;AAC9E,yFAAyF;AACzF,EAAE;AACF,6EAA6E;AAC7E,8EAA8E;AAC9E,wEAAwE;AACxE,gEAAgE;AAChE,EAAE;AACF,+EAA+E;AAC/E,gFAAgF;AAChF,iEAAiE;AACjE,8EAA8E;AAE9E,MAAM,QAAQ,GAAG,aAAa,CAAC;AAE/B;;;;GAIG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG,MAAM,CAAC;AAEpC,wFAAwF;AACxF,MAAM,CAAC,MAAM,kBAAkB,GAAG,qBAAqB,CAAC;AA4BxD;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,MAAM,UAAU,eAAe;IAC7B,MAAM,aAAa,GAAG,CACpB,UAAkB,aAAa,EAC/B,QAAiC,EAChB,EAAE,CAAC,mBAAmB,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;IAE7D,MAAM,cAAc,GAAG,CAAC,UAAkB,kBAAkB,EAAoB,EAAE;QAChF,IAAI,OAAO,KAAK,kBAAkB,EAAE,CAAC;YACnC,MAAM,IAAI,UAAU,CAClB,iBAAiB,EACjB,OAAO,EACP,yCAAyC,kBAAkB,kCAAkC,CAC9F,CAAC;QACJ,CAAC;QACD,OAAO,oBAAoB,EAAE,CAAC;IAChC,CAAC,CAAC;IAEF,MAAM,QAAQ,GAAG,aAAkC,CAAC;IACpD,QAAQ,CAAC,aAAa,GAAG,aAAa,CAAC;IACvC,QAAQ,CAAC,cAAc,GAAG,cAAc,CAAC;IACzC,QAAQ,CAAC,kBAAkB,GAAG,cAAc,CAAC;IAC7C,QAAQ,CAAC,UAAU,GAAG,CAAC,OAAe,EAAS,EAAE;QAC/C,MAAM,IAAI,UAAU,CAAC,iBAAiB,EAAE,OAAO,EAAE,4CAA4C,CAAC,CAAC;IACjG,CAAC,CAAC;IACF,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,qCAAqC;AACrC,MAAM,CAAC,MAAM,SAAS,GAAsB,eAAe,EAAE,CAAC;AAE9D,8EAA8E;AAC9E,iBAAiB;AACjB,8EAA8E;AAE9E,SAAS,mBAAmB,CAC1B,OAAe,EACf,QAAiC;IAEjC;;;;OAIG;IACH,MAAM,WAAW,GAAG,KAAK,IAAqB,EAAE;QAC9C,IAAI,OAAO,KAAK,aAAa,EAAE,CAAC;YAC9B,IAAI,CAAC;gBACH,OAAO,cAAc,EAAE,CAAC;YAC1B,CAAC;YAAC,MAAM,CAAC;gBACP,OAAO,aAAa,CAAC,CAAC,kEAAkE;YAC1F,CAAC;QACH,CAAC;QACD,IAAI,MAA0B,CAAC;QAC/B,IAAI,CAAC;YACH,MAAM,GAAG,cAAc,EAAE,CAAC;QAC5B,CAAC;QAAC,MAAM,CAAC;YACP,MAAM,GAAG,SAAS,CAAC;QACrB,CAAC;QACD,IAAI,MAAM,KAAK,OAAO,EAAE,CAAC;YACvB,MAAM,QAAQ,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;QACpC,CAAC;QACD,OAAO,OAAO,CAAC;IACjB,CAAC,CAAC;IAEF,OAAO;QACL,oBAAoB,EAAE,IAAI;QAC1B,QAAQ,EAAE,QAAQ;QAClB,OAAO;QACP,aAAa,EAAE,EAAE,EAAE,mDAAmD;QAEtE,KAAK,CAAC,UAAU,CAAC,OAAmC;YAClD,MAAM,IAAI,GAAG,kBAAkB,CAAC,OAAO,CAAC,CAAC;YACzC,MAAM,aAAa,GAAG,MAAM,WAAW,EAAE,CAAC;YAE1C,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,WAAW,CAAC,IAAI,CAAC,QAAQ,EAAE,EAAE,MAAM,EAAE,OAAO,CAAC,WAAW,EAAE,CAAC,CAAC;YACnF,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;YAElE,MAAM,OAAO,GACX,MAAM,CAAC,IAAI,KAAK,WAAW;gBACzB,CAAC,CAAC;oBACE;wBACE,IAAI,EAAE,WAAW;wBACjB,UAAU,EAAE,SAAS,CAAC,MAAM,CAAC;wBAC7B,QAAQ,EAAE,MAAM,CAAC,QAAQ;wBACzB,KAAK,EAAE,MAAM,CAAC,KAAK;qBACpB;iBACF;gBACH,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC;YAE5C,OAAO;gBACL,OAAO;gBACP,YAAY,EAAE;oBACZ,OAAO,EAAE,MAAM,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM;oBAC5D,GAAG,EAAE,SAAS;iBACf;gBACD,KAAK,EAAE,WAAW;gBAClB,QAAQ,EAAE,IAAI,CAAC,QAAQ;gBACvB,QAAQ,EAAE,EAAE,OAAO,EAAE,aAAa,EAAE,SAAS,EAAE,IAAI,IAAI,EAAE,EAAE;aAC5D,CAAC;QACJ,CAAC;QAED,KAAK,CAAC,QAAQ,CAAC,OAAmC;YAChD,IAAI,OAAO,cAAc,KAAK,WAAW,EAAE,CAAC;gBAC1C,MAAM,IAAI,KAAK,CACb,4EAA4E;oBAC1E,qFAAqF,CACxF,CAAC;YACJ,CAAC;YAED,MAAM,IAAI,GAAG,kBAAkB,CAAC,OAAO,CAAC,CAAC;YACzC,MAAM,aAAa,GAAG,MAAM,WAAW,EAAE,CAAC;YAC1C,MAAM,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC;YAExC,uEAAuE;YACvE,uEAAuE;YACvE,0DAA0D;YAC1D,IAAI,IAAI,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAC/C,MAAM,MAAM,GAAG,IAAI,cAAc,CAA4B;oBAC3D,KAAK,EAAE,KAAK,EAAE,UAAU,EAAE,EAAE;wBAC1B,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC;wBACtE,UAAU,CAAC,OAAO,CAAC;4BACjB,IAAI,EAAE,mBAAmB;4BACzB,OAAO,EAAE,aAAa;4BACtB,SAAS,EAAE,IAAI,IAAI,EAAE;yBACtB,CAAC,CAAC;wBACH,IAAI,CAAC;4BACH,MAAM,EAAE,IAAI,EAAE,GAAG,MAAM,WAAW,CAAC,IAAI,CAAC,QAAQ,EAAE,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,CAAC;4BAC3E,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;4BAClE,IAAI,MAAM,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gCAChC,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC;gCAC7B,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,kBAAkB,EAAE,EAAE,EAAE,QAAQ,EAAE,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;gCAChF,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,kBAAkB,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,CAAC,CAAC;gCAC1E,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,EAAE,EAAE,CAAC,CAAC;gCACnD,UAAU,CAAC,OAAO,CAAC;oCACjB,IAAI,EAAE,WAAW;oCACjB,UAAU,EAAE,EAAE;oCACd,QAAQ,EAAE,MAAM,CAAC,QAAQ;oCACzB,KAAK,EAAE,MAAM,CAAC,KAAK;iCACpB,CAAC,CAAC;gCACH,UAAU,CAAC,OAAO,CAAC;oCACjB,IAAI,EAAE,QAAQ;oCACd,KAAK,EAAE,WAAW;oCAClB,YAAY,EAAE,EAAE,OAAO,EAAE,YAAY,EAAE,GAAG,EAAE,SAAS,EAAE;iCACxD,CAAC,CAAC;4BACL,CAAC;iCAAM,CAAC;gCACN,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC;gCAC7B,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,EAAE,EAAE,CAAC,CAAC;gCAC/C,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC;gCACnE,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC,CAAC;gCAC7C,UAAU,CAAC,OAAO,CAAC;oCACjB,IAAI,EAAE,QAAQ;oCACd,KAAK,EAAE,WAAW;oCAClB,YAAY,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,EAAE,SAAS,EAAE;iCAClD,CAAC,CAAC;4BACL,CAAC;4BACD,UAAU,CAAC,KAAK,EAAE,CAAC;wBACrB,CAAC;wBAAC,OAAO,KAAK,EAAE,CAAC;4BACf,UAAU,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;wBAC1B,CAAC;oBACH,CAAC;iBACF,CAAC,CAAC;gBACH,OAAO,EAAE,MAAM,EAAE,CAAC;YACpB,CAAC;YAED,yDAAyD;YACzD,IAAI,MAAoD,CAAC;YACzD,IAAI,OAAO,GAAG,KAAK,CAAC;YACpB,MAAM,MAAM,GAAG,IAAI,cAAc,CAA4B;gBAC3D,KAAK,EAAE,CAAC,UAAU,EAAE,EAAE;oBACpB,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC;oBACtE,UAAU,CAAC,OAAO,CAAC;wBACjB,IAAI,EAAE,mBAAmB;wBACzB,OAAO,EAAE,aAAa;wBACtB,SAAS,EAAE,IAAI,IAAI,EAAE;qBACtB,CAAC,CAAC;oBAEH,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC;oBACjC,IAAI,WAAW,GAAG,KAAK,CAAC;oBAExB,MAAM,OAAO,GAAG,GAAG,EAAE;wBACnB,IAAI,OAAO;4BAAE,OAAO;wBACpB,OAAO,GAAG,IAAI,CAAC;wBACf,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,8DAA8D;wBAC9E,UAAU,CAAC,KAAK,CAAC,IAAI,UAAU,CAAC,qBAAqB,EAAE,EAAE,EAAE,mBAAmB,CAAC,CAAC,CAAC;oBACnF,CAAC,CAAC;oBACF,WAAW,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;oBAEhE,MAAM,GAAG,aAAa,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC,KAAK,EAAE,EAAE;wBAC9C,IAAI,OAAO;4BAAE,OAAO;wBACpB,IAAI,KAAK,CAAC,KAAK,KAAK,EAAE,EAAE,CAAC;4BACvB,IAAI,CAAC,WAAW,EAAE,CAAC;gCACjB,WAAW,GAAG,IAAI,CAAC;gCACnB,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;4BACzD,CAAC;4BACD,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;wBAC7E,CAAC;wBACD,IAAI,KAAK,CAAC,MAAM,EAAE,CAAC;4BACjB,OAAO,GAAG,IAAI,CAAC;4BACf,WAAW,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;4BACnD,IAAI,WAAW;gCAAE,UAAU,CAAC,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;4BACtE,UAAU,CAAC,OAAO,CAAC;gCACjB,IAAI,EAAE,QAAQ;gCACd,KAAK,EAAE,WAAW;gCAClB,YAAY,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,EAAE,SAAS,EAAE;6BAClD,CAAC,CAAC;4BACH,UAAU,CAAC,KAAK,EAAE,CAAC;wBACrB,CAAC;oBACH,CAAC,CAAC,CAAC;oBAEH,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;wBAC7B,IAAI,OAAO;4BAAE,OAAO;wBACpB,OAAO,GAAG,IAAI,CAAC;wBACf,WAAW,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;wBACnD,UAAU,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;oBAC1B,CAAC,CAAC,CAAC;gBACL,CAAC;gBACD,MAAM,EAAE,GAAG,EAAE;oBACX,OAAO,GAAG,IAAI,CAAC;oBACf,MAAM,EAAE,IAAI,EAAE,CAAC;gBACjB,CAAC;aACF,CAAC,CAAC;YACH,OAAO,EAAE,MAAM,EAAE,CAAC;QACpB,CAAC;KACF,CAAC;AACJ,CAAC;AAED,8EAA8E;AAC9E,kBAAkB;AAClB,8EAA8E;AAE9E,SAAS,oBAAoB;IAC3B,OAAO;QACL,oBAAoB,EAAE,IAAI;QAC1B,QAAQ,EAAE,QAAQ;QAClB,OAAO,EAAE,kBAAkB;QAC3B,oBAAoB,EAAE,QAAQ,EAAE,2CAA2C;QAC3E,qBAAqB,EAAE,IAAI,EAAE,sDAAsD;QAEnF,KAAK,CAAC,OAAO,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE;YACnC,IAAI,WAAW,EAAE,OAAO,EAAE,CAAC;gBACzB,MAAM,IAAI,UAAU,CAAC,qBAAqB,EAAE,EAAE,EAAE,sBAAsB,CAAC,CAAC;YAC1E,CAAC;YACD,MAAM,EAAE,UAAU,EAAE,GAAG,MAAM,KAAK,CAAC,MAAM,CAAC,CAAC;YAC3C,OAAO,EAAE,UAAU,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;QACtC,CAAC;KACF,CAAC;AACJ,CAAC","sourcesContent":["import type {\n  EmbeddingModelV3,\n  LanguageModelV3,\n  LanguageModelV3CallOptions,\n  LanguageModelV3Content,\n  LanguageModelV3GenerateResult,\n  LanguageModelV3StreamPart,\n  LanguageModelV3StreamResult,\n} from '@ai-sdk/provider';\nimport { embed, getActiveModel, sendMessage, setModel, streamMessage } from '../index';\nimport { ModelError, type SetModelOptions } from '../types';\nimport { convertCallOptions, extractOutput, EMPTY_USAGE, newPartId } from './convert';\n\n// ---------------------------------------------------------------------------\n// Vercel AI SDK provider for expo-ai-kit — `import { expoAiKit } from 'expo-ai-kit/ai'`.\n//\n// Implements the LanguageModelV3 spec (AI SDK 6; also accepted by AI SDK 7),\n// wrapping the same core sendMessage/streamMessage/embed calls as the rest of\n// the library — the single-flight INFERENCE_BUSY guard, stateless-model\n// semantics, and typed ModelError contract all apply unchanged.\n//\n// Zero-dependency by construction: everything imported from '@ai-sdk/provider'\n// is a *type* (erased at compile time). The package is only needed by consumers\n// who use this entry point, as an optional peer for its typings.\n// ---------------------------------------------------------------------------\n\nconst PROVIDER = 'expo-ai-kit';\n\n/**\n * Sentinel model id meaning \"whatever model is currently active\" — the OS\n * built-in by default, or whatever the app last activated via setModel().\n * The provider never switches models for this id.\n */\nexport const AUTO_MODEL_ID = 'auto';\n\n/** Model id for the built-in embedding model (Apple NLContextualEmbedding, iOS 17+). */\nexport const EMBEDDING_MODEL_ID = 'apple-nl-contextual';\n\n/**\n * Per-instance model settings, applied when this instance activates its model\n * (same shape as {@link setModel}'s options). If the model is already active,\n * it is NOT reloaded — on-device model loads are expensive, so sampling config\n * effectively belongs to whoever activated the model first.\n */\nexport type ExpoAiKitModelSettings = SetModelOptions;\n\nexport interface ExpoAiKitProvider {\n  /** Shorthand for {@link ExpoAiKitProvider.languageModel}. */\n  (modelId?: string, settings?: ExpoAiKitModelSettings): LanguageModelV3;\n  /**\n   * A LanguageModelV3 over the on-device model. Pass any model id accepted by\n   * setModel() ('apple-fm', 'mlkit', 'gemma-e2b', a registerModel() id, …) to\n   * have the provider activate it before generating, or omit / pass 'auto' to\n   * use whatever model is already active.\n   */\n  languageModel(modelId?: string, settings?: ExpoAiKitModelSettings): LanguageModelV3;\n  /** An EmbeddingModelV3 over embed() — iOS-only (Apple NLContextualEmbedding). */\n  embeddingModel(modelId?: string): EmbeddingModelV3;\n  /** @deprecated Spec alias for {@link ExpoAiKitProvider.embeddingModel}. */\n  textEmbeddingModel(modelId?: string): EmbeddingModelV3;\n  /** expo-ai-kit has no image models; always throws MODEL_NOT_FOUND. */\n  imageModel(modelId: string): never;\n}\n\n/**\n * Create an expo-ai-kit provider for the Vercel AI SDK.\n *\n * ```ts\n * import { generateText, streamText } from 'ai';\n * import { expoAiKit } from 'expo-ai-kit/ai';\n *\n * const { text } = await generateText({\n *   model: expoAiKit(),               // the active on-device model\n *   prompt: 'Capital of France?',\n * });\n *\n * const result = streamText({\n *   model: expoAiKit('gemma-e2b'),    // activates gemma-e2b first (must be downloaded)\n *   prompt: 'Write a short story',\n * });\n * ```\n *\n * On-device caveats (all documented in the guide):\n * - One generation at a time — concurrent calls reject with INFERENCE_BUSY.\n * - Per-call sampling (temperature, topK, …) is reported as an unsupported-\n *   setting warning; sampling is fixed when the model is activated.\n * - Tool calling and JSON output ride the same prompt protocol as\n *   generateText()/generateObject() — single-shot here, since the AI SDK owns\n *   the loop. Streaming buffers when tools/JSON are requested (the envelope\n *   must be parsed whole, not surfaced as text deltas).\n * - embeddingModel() is iOS-only for now; Android throws DEVICE_NOT_SUPPORTED.\n */\nexport function createExpoAiKit(): ExpoAiKitProvider {\n  const languageModel = (\n    modelId: string = AUTO_MODEL_ID,\n    settings?: ExpoAiKitModelSettings\n  ): LanguageModelV3 => createLanguageModel(modelId, settings);\n\n  const embeddingModel = (modelId: string = EMBEDDING_MODEL_ID): EmbeddingModelV3 => {\n    if (modelId !== EMBEDDING_MODEL_ID) {\n      throw new ModelError(\n        'MODEL_NOT_FOUND',\n        modelId,\n        `expo-ai-kit has one embedding model: \"${EMBEDDING_MODEL_ID}\" (Apple NLContextualEmbedding).`\n      );\n    }\n    return createEmbeddingModel();\n  };\n\n  const provider = languageModel as ExpoAiKitProvider;\n  provider.languageModel = languageModel;\n  provider.embeddingModel = embeddingModel;\n  provider.textEmbeddingModel = embeddingModel;\n  provider.imageModel = (modelId: string): never => {\n    throw new ModelError('MODEL_NOT_FOUND', modelId, 'expo-ai-kit does not provide image models.');\n  };\n  return provider;\n}\n\n/** The default provider instance. */\nexport const expoAiKit: ExpoAiKitProvider = createExpoAiKit();\n\n// ---------------------------------------------------------------------------\n// Language model\n// ---------------------------------------------------------------------------\n\nfunction createLanguageModel(\n  modelId: string,\n  settings?: ExpoAiKitModelSettings\n): LanguageModelV3 {\n  /**\n   * Make sure this instance's model is the active one. 'auto' never switches.\n   * Deliberately skipped when already active: reloading a multi-GB model per\n   * call would be catastrophic, so settings apply on activation only.\n   */\n  const ensureModel = async (): Promise<string> => {\n    if (modelId === AUTO_MODEL_ID) {\n      try {\n        return getActiveModel();\n      } catch {\n        return AUTO_MODEL_ID; // native module unavailable (e.g. web) — let the call itself fail\n      }\n    }\n    let active: string | undefined;\n    try {\n      active = getActiveModel();\n    } catch {\n      active = undefined;\n    }\n    if (active !== modelId) {\n      await setModel(modelId, settings);\n    }\n    return modelId;\n  };\n\n  return {\n    specificationVersion: 'v3',\n    provider: PROVIDER,\n    modelId,\n    supportedUrls: {}, // no URLs are consumed natively — text-only models\n\n    async doGenerate(options: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> {\n      const call = convertCallOptions(options);\n      const activeModelId = await ensureModel();\n\n      const { text } = await sendMessage(call.messages, { signal: options.abortSignal });\n      const output = extractOutput(text, call.toolNames, call.jsonMode);\n\n      const content: LanguageModelV3Content[] =\n        output.kind === 'tool-call'\n          ? [\n              {\n                type: 'tool-call',\n                toolCallId: newPartId('call'),\n                toolName: output.toolName,\n                input: output.input,\n              },\n            ]\n          : [{ type: 'text', text: output.text }];\n\n      return {\n        content,\n        finishReason: {\n          unified: output.kind === 'tool-call' ? 'tool-calls' : 'stop',\n          raw: undefined,\n        },\n        usage: EMPTY_USAGE,\n        warnings: call.warnings,\n        response: { modelId: activeModelId, timestamp: new Date() },\n      };\n    },\n\n    async doStream(options: LanguageModelV3CallOptions): Promise<LanguageModelV3StreamResult> {\n      if (typeof ReadableStream === 'undefined') {\n        throw new Error(\n          'expo-ai-kit/ai: streaming needs a ReadableStream polyfill in React Native ' +\n            '(e.g. web-streams-polyfill) — the AI SDK requires the same polyfills. See the docs.'\n        );\n      }\n\n      const call = convertCallOptions(options);\n      const activeModelId = await ensureModel();\n      const abortSignal = options.abortSignal;\n\n      // Tool calling / JSON mode: the envelope must be parsed as a whole — a\n      // half-streamed `{\"tool\": …` would surface as garbage text deltas — so\n      // those runs buffer and emit the parsed result in one go.\n      if (call.toolNames.length > 0 || call.jsonMode) {\n        const stream = new ReadableStream<LanguageModelV3StreamPart>({\n          start: async (controller) => {\n            controller.enqueue({ type: 'stream-start', warnings: call.warnings });\n            controller.enqueue({\n              type: 'response-metadata',\n              modelId: activeModelId,\n              timestamp: new Date(),\n            });\n            try {\n              const { text } = await sendMessage(call.messages, { signal: abortSignal });\n              const output = extractOutput(text, call.toolNames, call.jsonMode);\n              if (output.kind === 'tool-call') {\n                const id = newPartId('call');\n                controller.enqueue({ type: 'tool-input-start', id, toolName: output.toolName });\n                controller.enqueue({ type: 'tool-input-delta', id, delta: output.input });\n                controller.enqueue({ type: 'tool-input-end', id });\n                controller.enqueue({\n                  type: 'tool-call',\n                  toolCallId: id,\n                  toolName: output.toolName,\n                  input: output.input,\n                });\n                controller.enqueue({\n                  type: 'finish',\n                  usage: EMPTY_USAGE,\n                  finishReason: { unified: 'tool-calls', raw: undefined },\n                });\n              } else {\n                const id = newPartId('text');\n                controller.enqueue({ type: 'text-start', id });\n                controller.enqueue({ type: 'text-delta', id, delta: output.text });\n                controller.enqueue({ type: 'text-end', id });\n                controller.enqueue({\n                  type: 'finish',\n                  usage: EMPTY_USAGE,\n                  finishReason: { unified: 'stop', raw: undefined },\n                });\n              }\n              controller.close();\n            } catch (error) {\n              controller.error(error);\n            }\n          },\n        });\n        return { stream };\n      }\n\n      // Plain text: real token streaming over streamMessage().\n      let handle: ReturnType<typeof streamMessage> | undefined;\n      let settled = false;\n      const stream = new ReadableStream<LanguageModelV3StreamPart>({\n        start: (controller) => {\n          controller.enqueue({ type: 'stream-start', warnings: call.warnings });\n          controller.enqueue({\n            type: 'response-metadata',\n            modelId: activeModelId,\n            timestamp: new Date(),\n          });\n\n          const textId = newPartId('text');\n          let textStarted = false;\n\n          const onAbort = () => {\n            if (settled) return;\n            settled = true;\n            handle?.stop(); // stops native generation; promise resolves with partial text\n            controller.error(new ModelError('INFERENCE_CANCELLED', '', 'Aborted by caller'));\n          };\n          abortSignal?.addEventListener('abort', onAbort, { once: true });\n\n          handle = streamMessage(call.messages, (event) => {\n            if (settled) return;\n            if (event.token !== '') {\n              if (!textStarted) {\n                textStarted = true;\n                controller.enqueue({ type: 'text-start', id: textId });\n              }\n              controller.enqueue({ type: 'text-delta', id: textId, delta: event.token });\n            }\n            if (event.isDone) {\n              settled = true;\n              abortSignal?.removeEventListener('abort', onAbort);\n              if (textStarted) controller.enqueue({ type: 'text-end', id: textId });\n              controller.enqueue({\n                type: 'finish',\n                usage: EMPTY_USAGE,\n                finishReason: { unified: 'stop', raw: undefined },\n              });\n              controller.close();\n            }\n          });\n\n          handle.promise.catch((error) => {\n            if (settled) return;\n            settled = true;\n            abortSignal?.removeEventListener('abort', onAbort);\n            controller.error(error);\n          });\n        },\n        cancel: () => {\n          settled = true;\n          handle?.stop();\n        },\n      });\n      return { stream };\n    },\n  };\n}\n\n// ---------------------------------------------------------------------------\n// Embedding model\n// ---------------------------------------------------------------------------\n\nfunction createEmbeddingModel(): EmbeddingModelV3 {\n  return {\n    specificationVersion: 'v3',\n    provider: PROVIDER,\n    modelId: EMBEDDING_MODEL_ID,\n    maxEmbeddingsPerCall: Infinity, // embed() batches internally; no API limit\n    supportsParallelCalls: true, // embeddings bypass the single-flight inference guard\n\n    async doEmbed({ values, abortSignal }) {\n      if (abortSignal?.aborted) {\n        throw new ModelError('INFERENCE_CANCELLED', '', 'Aborted before start');\n      }\n      const { embeddings } = await embed(values);\n      return { embeddings, warnings: [] };\n    },\n  };\n}\n"]}

package/build/index.d.ts

@@ -1,6 +1,7 @@
-import { LLMMessage, LLMSendOptions, LLMResponse, LLMStreamOptions, LLMStreamCallback, LLMStreamHandle, BuiltInModel, DownloadableModel, SetModelOptions, JSONSchema, GenerateObjectOptions, GenerateObjectResult, GenerateTextOptions, GenerateTextResult } from './types';
+import { LLMMessage, LLMSendOptions, LLMResponse, LLMStreamOptions, LLMStreamCallback, LLMStreamHandle, BuiltInModel, DownloadableModel, SetModelOptions, JSONSchema, GenerateObjectOptions, GenerateObjectResult, GenerateTextOptions, GenerateTextResult, EmbedResult } from './types';
 export * from './types';
 export * from './models';
+export * from './rag';
 /**
  * Check if on-device AI is available on the current device.
  * Returns false on unsupported platforms (web, etc.).
@@ -174,6 +175,47 @@ export declare function generateObject<T = unknown>(messages: LLMMessage[], sche
  * ```
  */
 export declare function generateText(messages: LLMMessage[], options?: GenerateTextOptions): Promise<GenerateTextResult>;
+/**
+ * Turn text into embedding vectors for semantic search / on-device RAG.
+ *
+ * Returns one vector per input string (in order), which you can compare with
+ * {@link cosineSimilarity} or store in a {@link createVectorStore} to retrieve
+ * the most relevant chunks before a {@link sendMessage} / {@link generateText}
+ * call. Pair with {@link chunkText} to split documents first.
+ *
+ * **iOS-only for now**, backed by Apple's `NLContextualEmbedding` — a zero-
+ * download, OS-maintained model (no app-size cost, works even where Apple
+ * Intelligence isn't enabled, iOS 17+). On Android/web it throws
+ * `DEVICE_NOT_SUPPORTED`; the RAG toolkit (`chunkText`, `cosineSimilarity`,
+ * `createVectorStore`) still works there with any vector source you bring.
+ *
+ * Embeddings don't use the generation KV-cache, so `embed()` is **not** subject
+ * to the single-flight `INFERENCE_BUSY` guard — it can run alongside other work.
+ *
+ * @param texts - Non-empty array of strings to embed.
+ * @returns `{ embeddings, dimensions }` — `embeddings[i]` is the vector for `texts[i]`.
+ * @throws {ModelError} DEVICE_NOT_SUPPORTED off iOS, or if no embedding model is
+ *   available on the device.
+ *
+ * @example
+ * ```ts
+ * import { embed, chunkText, createVectorStore } from 'expo-ai-kit';
+ *
+ * const chunks = chunkText(document);
+ * const { embeddings } = await embed(chunks);
+ *
+ * const store = createVectorStore<{ text: string }>();
+ * store.addMany(chunks.map((text, i) => ({ id: `c${i}`, vector: embeddings[i], metadata: { text } })));
+ *
+ * const { embeddings: [q] } = await embed([question]);
+ * const context = store.search(q, { topK: 4 }).map((h) => h.metadata!.text).join('\n\n');
+ * const { text } = await sendMessage([
+ *   { role: 'system', content: `Answer using only this context:\n${context}` },
+ *   { role: 'user', content: question },
+ * ]);
+ * ```
+ */
+export declare function embed(texts: string[]): Promise<EmbedResult>;
 /**
  * Get all built-in models available on the current platform.
  *

package/build/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,UAAU,EACV,cAAc,EACd,WAAW,EACX,gBAAgB,EAEhB,iBAAiB,EACjB,eAAe,EACf,YAAY,EACZ,iBAAiB,EAIjB,eAAe,EACf,UAAU,EACV,qBAAqB,EACrB,oBAAoB,EACpB,mBAAmB,EACnB,kBAAkB,EAInB,MAAM,SAAS,CAAC;AAiBjB,cAAc,SAAS,CAAC;AACxB,cAAc,UAAU,CAAC;AA0HzB;;;GAGG;AACH,wBAAsB,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC,CAKpD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAsB,WAAW,CAC/B,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,CAAC,EAAE,cAAc,GACvB,OAAO,CAAC,WAAW,CAAC,CAmEtB;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,wBAAgB,aAAa,CAC3B,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,EAAE,iBAAiB,EAC1B,OAAO,CAAC,EAAE,gBAAgB,GACzB,eAAe,CAwFjB;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwCG;AACH,wBAAsB,cAAc,CAAC,CAAC,GAAG,OAAO,EAC9C,QAAQ,EAAE,UAAU,EAAE,EACtB,MAAM,EAAE,UAAU,EAClB,OAAO,CAAC,EAAE,qBAAqB,GAC9B,OAAO,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAoElC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,wBAAsB,YAAY,CAChC,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,CAAC,EAAE,mBAAmB,GAC5B,OAAO,CAAC,kBAAkB,CAAC,CAmJ7B;AAMD;;;;;;;GAOG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC,CAKhE;AAED;;;;;;;GAOG;AACH,wBAAsB,qBAAqB,IAAI,OAAO,CAAC,iBAAiB,EAAE,CAAC,CAkC1E;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAsB,mBAAmB,IAAI,OAAO,CAAC,iBAAiB,GAAG,IAAI,CAAC,CAM7E;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,aAAa,CACjC,OAAO,EAAE,MAAM,EACf,OAAO,CAAC,EAAE;IAAE,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAA;CAAE,GACpD,OAAO,CAAC,IAAI,CAAC,CA+Cf;AAED;;;;;;;GAOG;AACH,wBAAsB,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAKnE;AAED;;;;;;;GAOG;AACH,wBAAsB,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAOhE;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,wBAAsB,QAAQ,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC,CAQxF;AAED;;;;GAIG;AACH,wBAAgB,cAAc,IAAI,MAAM,CAEvC;AAED;;;;;GAKG;AACH,wBAAsB,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC,CAEjD"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,UAAU,EACV,cAAc,EACd,WAAW,EACX,gBAAgB,EAEhB,iBAAiB,EACjB,eAAe,EACf,YAAY,EACZ,iBAAiB,EAIjB,eAAe,EACf,UAAU,EACV,qBAAqB,EACrB,oBAAoB,EACpB,mBAAmB,EACnB,kBAAkB,EAIlB,WAAW,EACZ,MAAM,SAAS,CAAC;AAiBjB,cAAc,SAAS,CAAC;AACxB,cAAc,UAAU,CAAC;AACzB,cAAc,OAAO,CAAC;AA0HtB;;;GAGG;AACH,wBAAsB,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC,CAKpD;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAsB,WAAW,CAC/B,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,CAAC,EAAE,cAAc,GACvB,OAAO,CAAC,WAAW,CAAC,CAmEtB;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,wBAAgB,aAAa,CAC3B,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,EAAE,iBAAiB,EAC1B,OAAO,CAAC,EAAE,gBAAgB,GACzB,eAAe,CAwFjB;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwCG;AACH,wBAAsB,cAAc,CAAC,CAAC,GAAG,OAAO,EAC9C,QAAQ,EAAE,UAAU,EAAE,EACtB,MAAM,EAAE,UAAU,EAClB,OAAO,CAAC,EAAE,qBAAqB,GAC9B,OAAO,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,CAoElC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,wBAAsB,YAAY,CAChC,QAAQ,EAAE,UAAU,EAAE,EACtB,OAAO,CAAC,EAAE,mBAAmB,GAC5B,OAAO,CAAC,kBAAkB,CAAC,CAmJ7B;AAMD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuCG;AACH,wBAAsB,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,WAAW,CAAC,CAkBjE;AAMD;;;;;;;GAOG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC,CAKhE;AAED;;;;;;;GAOG;AACH,wBAAsB,qBAAqB,IAAI,OAAO,CAAC,iBAAiB,EAAE,CAAC,CAkC1E;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAsB,mBAAmB,IAAI,OAAO,CAAC,iBAAiB,GAAG,IAAI,CAAC,CAM7E;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,aAAa,CACjC,OAAO,EAAE,MAAM,EACf,OAAO,CAAC,EAAE;IAAE,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAA;CAAE,GACpD,OAAO,CAAC,IAAI,CAAC,CA+Cf;AAED;;;;;;;GAOG;AACH,wBAAsB,cAAc,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAKnE;AAED;;;;;;;GAOG;AACH,wBAAsB,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAOhE;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,wBAAsB,QAAQ,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC,CAQxF;AAED;;;;GAIG;AACH,wBAAgB,cAAc,IAAI,MAAM,CAEvC;AAED;;;;;GAKG;AACH,wBAAsB,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC,CAEjD"}

package/build/index.js

@@ -6,6 +6,7 @@ import { buildToolInstruction, parseToolCall, buildUnknownToolRepair, buildToolA
 import { getAllModels, getRegistryEntry } from './models';
 export * from './types';
 export * from './models';
+export * from './rag';
 const DEFAULT_SYSTEM_PROMPT = 'You are a helpful, friendly assistant. Answer the user directly and concisely.';
 const DEFAULT_OBJECT_SYSTEM_PROMPT = 'You output structured data as JSON. Follow the provided JSON Schema exactly.';
 let streamIdCounter = 0;
@@ -598,6 +599,64 @@ export async function generateText(messages, options) {
     };
 }
 // ============================================================================
+// Embeddings API
+// ============================================================================
+/**
+ * Turn text into embedding vectors for semantic search / on-device RAG.
+ *
+ * Returns one vector per input string (in order), which you can compare with
+ * {@link cosineSimilarity} or store in a {@link createVectorStore} to retrieve
+ * the most relevant chunks before a {@link sendMessage} / {@link generateText}
+ * call. Pair with {@link chunkText} to split documents first.
+ *
+ * **iOS-only for now**, backed by Apple's `NLContextualEmbedding` — a zero-
+ * download, OS-maintained model (no app-size cost, works even where Apple
+ * Intelligence isn't enabled, iOS 17+). On Android/web it throws
+ * `DEVICE_NOT_SUPPORTED`; the RAG toolkit (`chunkText`, `cosineSimilarity`,
+ * `createVectorStore`) still works there with any vector source you bring.
+ *
+ * Embeddings don't use the generation KV-cache, so `embed()` is **not** subject
+ * to the single-flight `INFERENCE_BUSY` guard — it can run alongside other work.
+ *
+ * @param texts - Non-empty array of strings to embed.
+ * @returns `{ embeddings, dimensions }` — `embeddings[i]` is the vector for `texts[i]`.
+ * @throws {ModelError} DEVICE_NOT_SUPPORTED off iOS, or if no embedding model is
+ *   available on the device.
+ *
+ * @example
+ * ```ts
+ * import { embed, chunkText, createVectorStore } from 'expo-ai-kit';
+ *
+ * const chunks = chunkText(document);
+ * const { embeddings } = await embed(chunks);
+ *
+ * const store = createVectorStore<{ text: string }>();
+ * store.addMany(chunks.map((text, i) => ({ id: `c${i}`, vector: embeddings[i], metadata: { text } })));
+ *
+ * const { embeddings: [q] } = await embed([question]);
+ * const context = store.search(q, { topK: 4 }).map((h) => h.metadata!.text).join('\n\n');
+ * const { text } = await sendMessage([
+ *   { role: 'system', content: `Answer using only this context:\n${context}` },
+ *   { role: 'user', content: question },
+ * ]);
+ * ```
+ */
+export async function embed(texts) {
+    if (Platform.OS !== 'ios') {
+        throw new ModelError('DEVICE_NOT_SUPPORTED', '', Platform.OS === 'android'
+            ? 'embed() is iOS-only for now (Apple NLContextualEmbedding); Android support via MediaPipe is planned. ' +
+                'The RAG toolkit (chunkText, cosineSimilarity, createVectorStore) works on Android with any vector source.'
+            : 'embed() is only available on iOS');
+    }
+    if (!Array.isArray(texts) || texts.length === 0) {
+        throw new Error('texts array cannot be empty');
+    }
+    if (!texts.every((t) => typeof t === 'string')) {
+        throw new Error('texts must be an array of strings');
+    }
+    return wrapNative(() => ExpoAiKitModule.embed(texts));
+}
+// ============================================================================
 // Model Management API
 // ============================================================================
 /**