Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
import { AppleLLM } from '@/lib/appleLLM';
import { LocalReasoningService } from '../LocalReasoningService';
jest.mock('@/lib/appleLLM', () => ({
AppleLLM: { generateText: jest.fn(), countTokens: jest.fn(async () => null) },
}));
const mockGenerateText = AppleLLM.generateText as jest.Mock;
const mockCountTokens = AppleLLM.countTokens as jest.Mock;
beforeEach(() => {
jest.clearAllMocks();
});
it('passes the caller signal to Apple Intelligence so a timeout stops the run', async () => {
mockGenerateText.mockResolvedValue({ text: ' Hello. ', model: 'apple-foundation-models' });
const { signal } = new AbortController();
await expect(
LocalReasoningService.processText({ text: 'um hello', systemPrompt: 'Clean up.', signal }),
).resolves.toEqual({ text: 'Hello.', model: 'apple-fm' });
expect(mockGenerateText).toHaveBeenCalledWith(expect.objectContaining({ prompt: 'um hello' }), {
signal,
});
});
it('reports an aborted run as a cancelled local request', async () => {
mockGenerateText.mockRejectedValue(Object.assign(new Error('cancelled'), { name: 'AbortError' }));
await expect(
LocalReasoningService.processText({ text: 'um hello', systemPrompt: 'Clean up.' }),
).rejects.toMatchObject({ name: 'LocalReasoningError', code: 'APPLE_LLM_CANCELLED' });
});
// The On-Device cleanup timeout assumes a generation rate; dev builds log the measured one.
describe('dev log', () => {
let log: jest.SpyInstance;
beforeEach(() => {
log = jest.spyOn(console, 'log').mockImplementation(() => undefined);
jest.spyOn(Date, 'now').mockReturnValueOnce(0).mockReturnValueOnce(2_000);
});
afterEach(() => {
jest.restoreAllMocks();
});
it('reports output tokens per second', async () => {
mockGenerateText.mockResolvedValue({ text: 'Hello.', model: 'apple-foundation-models' });
mockCountTokens.mockResolvedValueOnce(50);
await LocalReasoningService.processText({ text: 'um hello', systemPrompt: 'Clean up.' });
expect(mockCountTokens).toHaveBeenCalledWith({ prompt: 'Hello.' });
expect(log).toHaveBeenCalledWith(
'[reasoning] provider=local model=apple-fm elapsedMs=2000 outputTokens=50 tokensPerSecond=25.0',
);
});
it('marks the count as estimated where iOS cannot count tokens', async () => {
mockGenerateText.mockResolvedValue({ text: 'Hello there.', model: 'apple-foundation-models' });
await LocalReasoningService.processText({ text: 'um hello', systemPrompt: 'Clean up.' });
expect(log).toHaveBeenCalledWith(
'[reasoning] provider=local model=apple-fm elapsedMs=2000 outputTokens=4 (estimated) tokensPerSecond=2.0',
);
});
});