* Load model data from OPFS as a ReadableStream. * * @param key - Model identifier or nested path * @returns Readable stream of the model dat

Load data from storage as a ReadableStream. Priority: local filesystem > OPFS > memory cache.

* Load a voice activity detection model * @param modelPath Path to the model file * @param configJson Optional configuration JSON * @returns

* Load a TTS voice model via sherpa-onnx. * Model files must already be written to sherpa-onnx virtual FS.

Get the local file path for a catalog entry, or null if not downloaded (reads from cached state). */

Fallback to stderr

* Map architecture names to standard names. */

Mean pooling: average all token embeddings (excluding padding)

* Merge user-provided options over component defaults. * * For numeric fields, zero/negative values mean "use default" (except guidance_scale * whe

* Get the number of documents currently in the pipeline's vector store. * * @param pipelineHandle Pipeline handle * @return Document co

* Create a network error. * * @param message The error message * @param code The specific error code (defaults to NETWORK_E

* Create a network unavailable error. * * @param cause The underlying throwable cause * @return An SDKError with NETWORK_UN

Normalize vector to unit length (L2 normalization)

* Subscribe to events of a specific type. * @returns Unsubscribe function

* Generic typed event subscription * Example: events.on('generation', handler)

* Called when a model assignment fails. * * @param componentType The component type * @param modelId The model ID

* Called when an audio chunk is ready during streaming synthesis. * * @param audioChunk The audio chunk bytes */

* Called when an audio chunk is available during streaming. * * @param chunk The audio chunk */

* Called before an HTTP request is sent. * Can be used to add headers, modify the URL, etc. * * @param url The request URL

* Called when detection completes. * * @param result The detection result */

* Called when detection starts. */

* Called when a download fails. * * @param downloadId The download ID * @param modelId The model ID * @param error

* Called when an error occurs. * * @param errorCode The error code * @param errorMessage The error message */

* Subscribe to executor events * Returns unsubscribe function

* Called when generation completes. * * @param result The generation result */

* Called when generation starts. * * @param prompt The input prompt */

* Called when a platform LLM request completes. * * @param response The LLM response * @param error Error code (0 for succe

* Subscribe to model events

* Called when response generation is complete. * * @param result The response result */

* Called when a response token is generated during streaming. * * @param token The generated token */

* Register a callback for speech activity events. * Called when speech starts, ends, or is ongoing.

* Called when audio synthesis is complete. * * @param result The synthesis result */

* Subscribe to voice events

* Open a notes app (Samsung Notes or Google Keep) and optionally create a new note. * Returns true if successfully launched. */

* Open X (formerly Twitter) using explicit component intent. * getLaunchIntentForPackage fails for X because it has 20+ LAUNCHER activity-aliases

* Parse native error data object into SDKError.

* Parse LLM output for tool calls. * Uses C++ when available, falls back to TypeScript regex parsing.

Returns the next token without consuming it, or null if no more tokens */

* Prepare the VLM context for generation: reset state, configure sampler, * build prompt, load image (if provided), tokenize, and evaluate. * After

Print model status

Bjoern Hoehrmann LUT

* Process accumulated audio through the voice pipeline */

* Process current audio through the pipeline: STT -> LLM -> TTS

* Process audio samples through VAD. * Returns whether speech was detected in this frame. * * The Silero VAD expects 512-sample windows at 16

* Process an image with streaming output. * * Lock is only held for state transitions, NOT during the blocking native * JNI call, so th

* Save a boolean value */

* Emit a device event. * Maps to rac_analytics_device_t struct in C++. */

* Register analytics events callback with telemetry manager. * Events from C++ will be routed to the telemetry manager for batching and HTTP tran

* Convert Float32 PCM audio data to WAV format. * * TTS backends typically output raw Float32 PCM samples in range [-1.0, 1.0]. * This

* Get Supabase anon key for development mode. * @return Supabase anon key or null if not configured */

* Get Supabase URL for development mode. * @return Supabase URL or null if not configured */

* Set model assignment callbacks. * The callback object must implement: * - httpGet(endpoint: String, requiresAuth: Boolean): String (return