@beltoinc/slyos-sdk 1.5.4 → 1.5.6

package/src/index.ts

@@ -0,0 +1,2147 @@
+import axios from 'axios';
+import { pipeline, env } from '@huggingface/transformers';
+
+// @ts-ignore - Force CPU in Node.js
+if (env.backends?.onnx?.wasm) {
+  env.backends.onnx.wasm.proxy = false;
+}
+
+// ─── Types ──────────────────────────────────────────────────────────
+
+interface SlyOSConfig {
+  apiKey: string;
+  apiUrl?: string;
+  onProgress?: ProgressCallback;
+  onEvent?: EventCallback;
+}
+
+interface GenerateOptions {
+  temperature?: number;
+  maxTokens?: number;
+  topP?: number;
+}
+
+interface TranscribeOptions {
+  language?: string;
+  returnTimestamps?: boolean;
+}
+
+type ModelCategory = 'llm' | 'stt';
+type QuantizationLevel = 'q4' | 'q8' | 'fp16' | 'fp32';
+
+interface ModelInfo {
+  hfModel: string;
+  task: string;
+  category: ModelCategory;
+  sizesMB: Record<QuantizationLevel, number>;
+  minRAM_MB: Record<QuantizationLevel, number>;
+}
+
+interface DeviceProfile {
+  cpuCores: number;
+  memoryMB: number;
+  estimatedStorageMB: number;
+  platform: 'web' | 'nodejs';
+  os: string;
+  recommendedQuant: QuantizationLevel;
+  maxContextWindow: number;
+  // Enhanced device intelligence fields
+  deviceFingerprint?: string;
+  gpuRenderer?: string;
+  gpuVramMb?: number;
+  screenWidth?: number;
+  screenHeight?: number;
+  pixelRatio?: number;
+  browserName?: string;
+  browserVersion?: string;
+  networkType?: string;
+  latencyToApiMs?: number;
+  timezone?: string;
+  wasmAvailable?: boolean;
+  webgpuAvailable?: boolean;
+}
+
+interface ProgressEvent {
+  stage: 'initializing' | 'profiling' | 'downloading' | 'loading' | 'ready' | 'generating' | 'transcribing' | 'error';
+  progress: number; // 0-100
+  message: string;
+  detail?: any;
+}
+
+interface SlyEvent {
+  type: 'auth' | 'device_registered' | 'device_profiled' | 'model_download_start' | 'model_download_progress' | 'model_loaded' | 'inference_start' | 'inference_complete' | 'error' | 'fallback_success' | 'fallback_error' | 'telemetry_flushed' | 'token';
+  data?: any;
+  timestamp: number;
+}
+
+type ProgressCallback = (event: ProgressEvent) => void;
+type EventCallback = (event: SlyEvent) => void;
+
+// ─── OpenAI Compatibility Types ──────────────────────────────────────
+
+interface OpenAIMessage {
+  role: 'system' | 'user' | 'assistant';
+  content: string;
+}
+
+interface OpenAIChatCompletionRequest {
+  messages: OpenAIMessage[];
+  temperature?: number;
+  top_p?: number;
+  max_tokens?: number;
+  frequency_penalty?: number;
+  presence_penalty?: number;
+  stop?: string | string[];
+}
+
+interface OpenAIChoice {
+  index: number;
+  message: OpenAIMessage;
+  finish_reason: string;
+}
+
+interface OpenAIUsage {
+  prompt_tokens: number;
+  completion_tokens: number;
+  total_tokens: number;
+}
+
+interface OpenAIChatCompletionResponse {
+  id: string;
+  object: 'chat.completion';
+  created: number;
+  model: string;
+  choices: OpenAIChoice[];
+  usage: OpenAIUsage;
+}
+
+// ─── AWS Bedrock Compatibility Types ─────────────────────────────────
+
+interface BedrockTextGenerationConfig {
+  maxTokenCount?: number;
+  temperature?: number;
+  topP?: number;
+  topK?: number;
+  stopSequences?: string[];
+}
+
+interface BedrockInvokeRequest {
+  inputText: string;
+  textGenerationConfig?: BedrockTextGenerationConfig;
+}
+
+interface BedrockResult {
+  outputText: string;
+  tokenCount: number;
+}
+
+interface BedrockInvokeResponse {
+  results: BedrockResult[];
+  input_text_token_count?: number;
+}
+
+// ─── Fallback Configuration ─────────────────────────────────────────
+
+type FallbackProvider = 'openai' | 'bedrock';
+
+interface FallbackConfig {
+  provider: FallbackProvider;
+  apiKey: string;
+  model: string;
+  region?: string; // for Bedrock
+}
+
+interface SlyOSConfigWithFallback extends SlyOSConfig {
+  fallback?: FallbackConfig;
+}
+
+// ─── OpenAI Compatible Client ───────────────────────────────────────
+
+interface OpenAICompatibleClient {
+  chat: {
+    completions: {
+      create(request: OpenAIChatCompletionRequest & { model: string }): Promise<OpenAIChatCompletionResponse>;
+    };
+  };
+}
+
+// ─── RAG Types ──────────────────────────────────────────────────
+
+interface RAGOptions {
+  knowledgeBaseId: string;
+  query: string;
+  topK?: number;
+  modelId: string;
+  temperature?: number;
+  maxTokens?: number;
+  // NEW: streaming callback
+  onToken?: (token: string, partial: string) => void;
+}
+
+interface RAGChunk {
+  id: string;
+  documentId: string;
+  documentName: string;
+  content: string;
+  similarityScore: number;
+  metadata?: Record<string, any>;
+}
+
+interface RAGResponse {
+  query: string;
+  retrievedChunks: RAGChunk[];
+  generatedResponse: string;
+  context: string;
+  latencyMs: number;
+  tierUsed: 1 | 2 | 3;
+  // NEW: detailed timing metrics
+  timing: {
+    retrievalMs: number;      // Time spent retrieving/embedding chunks
+    contextBuildMs: number;   // Time spent building context
+    firstTokenMs: number;     // Time to first token (from generation start)
+    generationMs: number;     // Total generation time
+    totalMs: number;          // End-to-end latency
+    tokensGenerated: number;  // Number of tokens in response
+    tokensPerSecond: number;  // Generation throughput
+  };
+  // NEW: dynamic config used
+  config: {
+    maxContextChars: number;
+    maxGenTokens: number;
+    chunkSize: number;
+    topK: number;
+    contextWindowUsed: number;
+    deviceTier: 'low' | 'mid' | 'high';
+  };
+}
+
+interface OfflineIndex {
+  metadata: {
+    kb_id: string;
+    kb_name: string;
+    chunk_size: number;
+    embedding_dim: number;
+    total_chunks: number;
+    synced_at: string;
+    expires_at: string;
+    sync_token: string;
+  };
+  chunks: Array<{
+    id: string;
+    document_id: string;
+    document_name: string;
+    content: string;
+    chunk_index: number;
+    embedding: number[] | null;
+    metadata: Record<string, any>;
+  }>;
+}
+
+// ─── Model Registry ─────────────────────────────────────────────────
+
+const modelMap: Record<string, ModelInfo> = {
+  // LLM models (1B+)
+  'quantum-1.7b': {
+    hfModel: 'onnx-community/SmolLM2-1.7B-Instruct',
+    task: 'text-generation',
+    category: 'llm',
+    sizesMB: { q4: 900, q8: 1700, fp16: 3400, fp32: 6800 },
+    minRAM_MB: { q4: 2048, q8: 3072, fp16: 5120, fp32: 8192 },
+  },
+  'quantum-3b': {
+    hfModel: 'onnx-community/Qwen2.5-3B-Instruct',
+    task: 'text-generation',
+    category: 'llm',
+    sizesMB: { q4: 1600, q8: 3200, fp16: 6400, fp32: 12800 },
+    minRAM_MB: { q4: 3072, q8: 5120, fp16: 8192, fp32: 16384 },
+  },
+  'quantum-code-3b': {
+    hfModel: 'onnx-community/Qwen2.5-Coder-3B-Instruct',
+    task: 'text-generation',
+    category: 'llm',
+    sizesMB: { q4: 1600, q8: 3200, fp16: 6400, fp32: 12800 },
+    minRAM_MB: { q4: 3072, q8: 5120, fp16: 8192, fp32: 16384 },
+  },
+  'quantum-8b': {
+    hfModel: 'onnx-community/Qwen2.5-7B-Instruct',
+    task: 'text-generation',
+    category: 'llm',
+    sizesMB: { q4: 4200, q8: 8400, fp16: 16800, fp32: 33600 },
+    minRAM_MB: { q4: 6144, q8: 10240, fp16: 20480, fp32: 40960 },
+  },
+  // STT models
+  'voicecore-base': {
+    hfModel: 'onnx-community/whisper-base',
+    task: 'automatic-speech-recognition',
+    category: 'stt',
+    sizesMB: { q4: 40, q8: 75, fp16: 150, fp32: 300 },
+    minRAM_MB: { q4: 512, q8: 512, fp16: 1024, fp32: 2048 },
+  },
+  'voicecore-small': {
+    hfModel: 'onnx-community/whisper-small',
+    task: 'automatic-speech-recognition',
+    category: 'stt',
+    sizesMB: { q4: 100, q8: 200, fp16: 400, fp32: 800 },
+    minRAM_MB: { q4: 1024, q8: 1024, fp16: 2048, fp32: 4096 },
+  },
+};
+
+// ─── Context Window Sizing ──────────────────────────────────────────
+
+function recommendContextWindow(memoryMB: number, quant: QuantizationLevel): number {
+  // More RAM + smaller quant = larger context window
+  const base = quant === 'q4' ? 1024 : quant === 'q8' ? 2048 : quant === 'fp16' ? 4096 : 8192;
+
+  if (memoryMB >= 16384) return Math.min(base * 4, 32768);
+  if (memoryMB >= 8192) return Math.min(base * 2, 16384);
+  if (memoryMB >= 4096) return base;
+  return Math.max(512, Math.floor(base / 2));
+}
+
+function selectQuantization(memoryMB: number, modelId: string): QuantizationLevel {
+  const info = modelMap[modelId];
+  if (!info) return 'q4';
+
+  // ONNX/WASM has protobuf size limits — fp16 files >2GB crash on many systems.
+  // For LLMs, cap at q4 via WASM. FP16/Q8 need native backends (llama.cpp).
+  // STT models are small enough for q8/fp16.
+  if (info.category === 'llm') {
+    return 'q4'; // safest for ONNX/WASM across all platforms
+  }
+
+  // STT models: try from best quality down
+  const quants: QuantizationLevel[] = ['fp16', 'q8', 'q4'];
+  for (const q of quants) {
+    if (memoryMB >= info.minRAM_MB[q]) return q;
+  }
+  return 'q4'; // fallback
+}
+
+// ─── Context Window Detection ──────────────────────────────────────
+
+async function detectContextWindowFromHF(hfModelId: string): Promise<number> {
+  try {
+    const configUrl = `https://huggingface.co/${hfModelId}/raw/main/config.json`;
+    const response = await axios.get(configUrl, { timeout: 5000 });
+    const config = response.data;
+
+    // Try multiple context window field names
+    const contextWindow =
+      config.max_position_embeddings ||
+      config.n_positions ||
+      config.max_seq_len ||
+      config.model_max_length ||
+      2048;
+
+    return contextWindow;
+  } catch {
+    // Default if config cannot be fetched
+    return 2048;
+  }
+}
+
+// ─── SDK Version ────────────────────────────────────────────────────
+const SDK_VERSION = '1.4.1';
+
+// ─── Persistent Device Identity ─────────────────────────────────────
+
+async function hashString(str: string): Promise<string> {
+  const isNode = typeof window === 'undefined';
+  if (isNode) {
+    const crypto = await import('crypto');
+    return crypto.createHash('sha256').update(str).digest('hex').substring(0, 32);
+  } else {
+    const encoder = new TextEncoder();
+    const data = encoder.encode(str);
+    const hashBuffer = await crypto.subtle.digest('SHA-256', data);
+    return Array.from(new Uint8Array(hashBuffer))
+      .map(b => b.toString(16).padStart(2, '0'))
+      .join('')
+      .substring(0, 32);
+  }
+}
+
+async function getOrCreateDeviceId(): Promise<string> {
+  const isNode = typeof window === 'undefined';
+
+  if (isNode) {
+    // Node.js: persist in ~/.slyos/device-id
+    try {
+      const fs = await import('fs');
+      const path = await import('path');
+      const os = await import('os');
+      const slyosDir = path.join(os.homedir(), '.slyos');
+      const idFile = path.join(slyosDir, 'device-id');
+
+      try {
+        const existing = fs.readFileSync(idFile, 'utf-8').trim();
+        if (existing) return existing;
+      } catch {}
+
+      const deviceId = `device-${Date.now()}-${Math.random().toString(36).substr(2, 12)}`;
+      fs.mkdirSync(slyosDir, { recursive: true });
+      fs.writeFileSync(idFile, deviceId);
+      return deviceId;
+    } catch {
+      return `device-${Date.now()}-${Math.random().toString(36).substr(2, 12)}`;
+    }
+  } else {
+    // Browser: persist in localStorage
+    const key = 'slyos_device_id';
+    try {
+      const existing = localStorage.getItem(key);
+      if (existing) return existing;
+    } catch {}
+
+    const deviceId = `device-${Date.now()}-${Math.random().toString(36).substr(2, 12)}`;
+    try { localStorage.setItem(key, deviceId); } catch {}
+    return deviceId;
+  }
+}
+
+async function generateDeviceFingerprint(): Promise<string> {
+  const isNode = typeof window === 'undefined';
+  let components: string[] = [];
+
+  if (isNode) {
+    try {
+      const os = await import('os');
+      const cpus = os.cpus();
+      components.push(cpus[0]?.model || 'unknown-cpu');
+      components.push(String(os.totalmem()));
+      components.push(os.platform());
+      components.push(os.arch());
+      components.push(String(cpus.length));
+    } catch {}
+  } else {
+    components.push(String(navigator.hardwareConcurrency || 0));
+    components.push(String((navigator as any).deviceMemory || 0));
+    components.push(navigator.platform || 'unknown');
+    // WebGL renderer for GPU fingerprint
+    try {
+      const canvas = document.createElement('canvas');
+      const gl = canvas.getContext('webgl') || canvas.getContext('experimental-webgl') as WebGLRenderingContext | null;
+      if (gl) {
+        const ext = gl.getExtension('WEBGL_debug_renderer_info');
+        if (ext) {
+          components.push(gl.getParameter(ext.UNMASKED_RENDERER_WEBGL) || 'unknown-gpu');
+        }
+      }
+    } catch {}
+    components.push(String(screen.width || 0));
+    components.push(String(screen.height || 0));
+  }
+
+  return await hashString(components.join('|'));
+}
+
+// ─── Enhanced Device Profiling ──────────────────────────────────────
+
+function detectGPU(): { renderer: string | null; vramMb: number } {
+  if (typeof window === 'undefined') return { renderer: null, vramMb: 0 };
+  try {
+    const canvas = document.createElement('canvas');
+    const gl = canvas.getContext('webgl') || canvas.getContext('experimental-webgl') as WebGLRenderingContext | null;
+    if (!gl) return { renderer: null, vramMb: 0 };
+    const ext = gl.getExtension('WEBGL_debug_renderer_info');
+    const renderer = ext ? gl.getParameter(ext.UNMASKED_RENDERER_WEBGL) : null;
+    // Rough VRAM estimate from renderer string
+    let vramMb = 0;
+    if (renderer) {
+      const match = renderer.match(/(\d+)\s*MB/i);
+      if (match) vramMb = parseInt(match[1]);
+      else if (/RTX\s*40/i.test(renderer)) vramMb = 8192;
+      else if (/RTX\s*30/i.test(renderer)) vramMb = 6144;
+      else if (/GTX/i.test(renderer)) vramMb = 4096;
+      else if (/Apple M[2-4]/i.test(renderer)) vramMb = 8192;
+      else if (/Apple M1/i.test(renderer)) vramMb = 4096;
+      else if (/Intel/i.test(renderer)) vramMb = 1024;
+    }
+    return { renderer, vramMb };
+  } catch {
+    return { renderer: null, vramMb: 0 };
+  }
+}
+
+function detectBrowser(): { name: string; version: string } {
+  if (typeof window === 'undefined' || typeof navigator === 'undefined') return { name: 'node', version: process.version || 'unknown' };
+  const ua = navigator.userAgent;
+  if (/Edg\//i.test(ua)) { const m = ua.match(/Edg\/([\d.]+)/); return { name: 'Edge', version: m?.[1] || '' }; }
+  if (/Chrome\//i.test(ua)) { const m = ua.match(/Chrome\/([\d.]+)/); return { name: 'Chrome', version: m?.[1] || '' }; }
+  if (/Firefox\//i.test(ua)) { const m = ua.match(/Firefox\/([\d.]+)/); return { name: 'Firefox', version: m?.[1] || '' }; }
+  if (/Safari\//i.test(ua)) { const m = ua.match(/Version\/([\d.]+)/); return { name: 'Safari', version: m?.[1] || '' }; }
+  return { name: 'unknown', version: '' };
+}
+
+function detectNetworkType(): string {
+  if (typeof navigator === 'undefined') return 'unknown';
+  const conn = (navigator as any).connection || (navigator as any).mozConnection || (navigator as any).webkitConnection;
+  if (!conn) return 'unknown';
+  return conn.effectiveType || conn.type || 'unknown';
+}
+
+async function measureApiLatency(apiUrl: string): Promise<number> {
+  try {
+    const start = Date.now();
+    await axios.head(`${apiUrl}/api/health`, { timeout: 5000 });
+    return Date.now() - start;
+  } catch {
+    try {
+      const start = Date.now();
+      await axios.get(`${apiUrl}/api/health`, { timeout: 5000 });
+      return Date.now() - start;
+    } catch {
+      return -1;
+    }
+  }
+}
+
+// ─── Device Profiling ───────────────────────────────────────────────
+
+async function profileDevice(): Promise<DeviceProfile> {
+  const isNode = typeof window === 'undefined';
+  let cpuCores = 4;
+  let memoryMB = 4096;
+  let estimatedStorageMB = 10000;
+  let platform: 'web' | 'nodejs' = isNode ? 'nodejs' : 'web';
+  let os = 'unknown';
+
+  if (isNode) {
+    // Node.js environment
+    try {
+      const osModule = await import('os');
+      cpuCores = osModule.cpus().length;
+      memoryMB = Math.round(osModule.totalmem() / (1024 * 1024));
+      os = `${osModule.platform()} ${osModule.release()}`;
+
+      // Estimate free disk via df-like check
+      try {
+        const { execSync } = await import('child_process');
+        const dfOutput = execSync('df -m . 2>/dev/null || echo "0 0 0 0"', { encoding: 'utf-8' });
+        const lines = dfOutput.trim().split('\n');
+        if (lines.length > 1) {
+          const parts = lines[1].split(/\s+/);
+          estimatedStorageMB = parseInt(parts[3]) || 10000; // Available column
+        }
+      } catch {
+        estimatedStorageMB = 10000;
+      }
+    } catch {
+      // Fallback
+    }
+  } else {
+    // Browser environment
+    cpuCores = navigator.hardwareConcurrency || 4;
+    memoryMB = ((navigator as any).deviceMemory || 4) * 1024; // deviceMemory is in GB
+    os = navigator.userAgent;
+
+    // Storage Manager API (Chrome 61+)
+    try {
+      if (navigator.storage && navigator.storage.estimate) {
+        const estimate = await navigator.storage.estimate();
+        estimatedStorageMB = Math.round((estimate.quota || 0) / (1024 * 1024));
+      }
+    } catch {
+      estimatedStorageMB = 5000;
+    }
+  }
+
+  const recommendedQuant = selectQuantization(memoryMB, 'quantum-1.7b'); // default baseline
+  const maxContextWindow = recommendContextWindow(memoryMB, recommendedQuant);
+
+  // Enhanced profiling
+  const gpu = detectGPU();
+  const browser = detectBrowser();
+  const networkType = detectNetworkType();
+  const timezone = Intl?.DateTimeFormat?.()?.resolvedOptions?.()?.timeZone || 'unknown';
+
+  let screenWidth = 0, screenHeight = 0, pixelRatio = 0;
+  let wasmAvailable = false, webgpuAvailable = false;
+
+  if (!isNode) {
+    screenWidth = screen?.width || 0;
+    screenHeight = screen?.height || 0;
+    pixelRatio = window?.devicePixelRatio || 1;
+  }
+
+  // Capability detection
+  try { wasmAvailable = typeof WebAssembly !== 'undefined'; } catch {}
+  if (!isNode) {
+    try { webgpuAvailable = !!(navigator as any).gpu; } catch {}
+  }
+
+  return {
+    cpuCores,
+    memoryMB,
+    estimatedStorageMB,
+    platform,
+    os,
+    recommendedQuant,
+    maxContextWindow,
+    gpuRenderer: gpu.renderer || undefined,
+    gpuVramMb: gpu.vramMb || undefined,
+    screenWidth: screenWidth || undefined,
+    screenHeight: screenHeight || undefined,
+    pixelRatio: pixelRatio || undefined,
+    browserName: browser.name,
+    browserVersion: browser.version,
+    networkType,
+    timezone,
+    wasmAvailable,
+    webgpuAvailable,
+  };
+}
+
+// ─── Main SDK Class ─────────────────────────────────────────────────
+
+interface TelemetryEntry {
+  latency_ms: number;
+  tokens_generated: number;
+  success: boolean;
+  model_id: string;
+  timestamp: number;
+}
+
+class SlyOS {
+  private apiKey: string;
+  private apiUrl: string;
+  private deviceId: string;
+  private token: string | null = null;
+  private models: Map<string, any> = new Map();
+  private deviceProfile: DeviceProfile | null = null;
+  private onProgress: ProgressCallback | null;
+  private onEvent: EventCallback | null;
+  private fallbackConfig: FallbackConfig | null;
+  private modelContextWindow: number = 0;
+  // Telemetry batching
+  private telemetryBuffer: TelemetryEntry[] = [];
+  private telemetryFlushTimer: any = null;
+  private static readonly TELEMETRY_BATCH_SIZE = 10;
+  private static readonly TELEMETRY_FLUSH_INTERVAL = 60000; // 60 seconds
+
+  constructor(config: SlyOSConfigWithFallback) {
+    this.apiKey = config.apiKey;
+    this.apiUrl = config.apiUrl || 'https://api.slyos.world';
+    this.deviceId = ''; // Set asynchronously in initialize()
+    this.onProgress = config.onProgress || null;
+    this.onEvent = config.onEvent || null;
+    this.fallbackConfig = config.fallback || null;
+  }
+
+  // ── Progress & Event Helpers ────────────────────────────────────
+
+  private emitProgress(stage: ProgressEvent['stage'], progress: number, message: string, detail?: any) {
+    if (this.onProgress) {
+      this.onProgress({ stage, progress, message, detail });
+    }
+  }
+
+  private emitEvent(type: SlyEvent['type'], data?: any) {
+    if (this.onEvent) {
+      this.onEvent({ type, data, timestamp: Date.now() });
+    }
+  }
+
+  // ── Telemetry Batching ─────────────────────────────────────────
+
+  private recordTelemetry(entry: TelemetryEntry) {
+    this.telemetryBuffer.push(entry);
+    if (this.telemetryBuffer.length >= SlyOS.TELEMETRY_BATCH_SIZE) {
+      this.flushTelemetry();
+    } else if (!this.telemetryFlushTimer) {
+      this.telemetryFlushTimer = setTimeout(() => this.flushTelemetry(), SlyOS.TELEMETRY_FLUSH_INTERVAL);
+    }
+  }
+
+  private async flushTelemetry() {
+    if (this.telemetryFlushTimer) {
+      clearTimeout(this.telemetryFlushTimer);
+      this.telemetryFlushTimer = null;
+    }
+    if (this.telemetryBuffer.length === 0 || !this.token) return;
+
+    const batch = [...this.telemetryBuffer];
+    this.telemetryBuffer = [];
+
+    try {
+      await axios.post(`${this.apiUrl}/api/devices/telemetry`, {
+        device_id: this.deviceId,
+        metrics: batch,
+      }, {
+        headers: { Authorization: `Bearer ${this.token}` },
+        timeout: 10000,
+      });
+      this.emitEvent('telemetry_flushed', { count: batch.length });
+    } catch {
+      // Put back on failure for next attempt
+      this.telemetryBuffer.unshift(...batch);
+      // Cap buffer to prevent memory leak
+      if (this.telemetryBuffer.length > 100) {
+        this.telemetryBuffer = this.telemetryBuffer.slice(-100);
+      }
+    }
+  }
+
+  // ── Device Analysis ─────────────────────────────────────────────
+
+  async analyzeDevice(): Promise<DeviceProfile> {
+    try {
+      this.emitProgress('profiling', 10, 'Analyzing device capabilities...');
+      this.deviceProfile = await profileDevice();
+      this.emitProgress('profiling', 100, `Device: ${this.deviceProfile.cpuCores} cores, ${Math.round(this.deviceProfile.memoryMB / 1024 * 10) / 10}GB RAM`);
+      this.emitEvent('device_profiled', this.deviceProfile);
+      return this.deviceProfile;
+    } catch (err: any) {
+      this.emitEvent('error', { method: 'analyzeDevice', error: err.message });
+      throw new Error(`Device analysis failed: ${err.message}`);
+    }
+  }
+
+  getDeviceProfile(): DeviceProfile | null {
+    return this.deviceProfile;
+  }
+
+  getModelContextWindow(): number {
+    return this.modelContextWindow;
+  }
+
+  getDeviceId(): string {
+    return this.deviceId;
+  }
+
+  getSdkVersion(): string {
+    return SDK_VERSION;
+  }
+
+  // Flush remaining telemetry and clean up timers
+  async destroy(): Promise<void> {
+    await this.flushTelemetry();
+    if (this.telemetryFlushTimer) {
+      clearTimeout(this.telemetryFlushTimer);
+      this.telemetryFlushTimer = null;
+    }
+  }
+
+  // ── Smart Model Recommendation ──────────────────────────────────
+
+  recommendModel(category: ModelCategory = 'llm'): { modelId: string; quant: QuantizationLevel; contextWindow: number; reason: string } | null {
+    if (!this.deviceProfile) {
+      throw new Error('Call analyzeDevice() first to get a recommendation.');
+    }
+
+    const mem = this.deviceProfile.memoryMB;
+    const candidates = Object.entries(modelMap).filter(([_, info]) => info.category === category);
+
+    // Sort by size descending — pick the biggest model that fits
+    for (const [id, info] of candidates.sort((a, b) => b[1].sizesMB.q4 - a[1].sizesMB.q4)) {
+      const quant = selectQuantization(mem, id);
+      if (mem >= info.minRAM_MB[quant]) {
+        const ctx = recommendContextWindow(mem, quant);
+        return {
+          modelId: id,
+          quant,
+          contextWindow: ctx,
+          reason: `Best model for ${Math.round(mem / 1024)}GB RAM at ${quant.toUpperCase()} precision`,
+        };
+      }
+    }
+
+    // Fallback to smallest
+    const smallest = candidates.sort((a, b) => a[1].sizesMB.q4 - b[1].sizesMB.q4)[0];
+    if (smallest) {
+      return {
+        modelId: smallest[0],
+        quant: 'q4',
+        contextWindow: 512,
+        reason: 'Limited device memory — using smallest available model at Q4',
+      };
+    }
+
+    return null;
+  }
+
+  // ── Initialize ──────────────────────────────────────────────────
+
+  async initialize(): Promise<DeviceProfile> {
+    this.emitProgress('initializing', 0, 'Starting SlyOS...');
+
+    // Step 1: Persistent device ID
+    this.deviceId = await getOrCreateDeviceId();
+
+    // Step 2: Profile device (enhanced)
+    this.emitProgress('profiling', 5, 'Detecting device capabilities...');
+    this.deviceProfile = await profileDevice();
+
+    // Step 2b: Generate device fingerprint
+    this.deviceProfile.deviceFingerprint = await generateDeviceFingerprint();
+
+    this.emitProgress('profiling', 20, `Detected: ${this.deviceProfile.cpuCores} CPU cores, ${Math.round(this.deviceProfile.memoryMB / 1024 * 10) / 10}GB RAM${this.deviceProfile.gpuRenderer ? ', GPU: ' + this.deviceProfile.gpuRenderer.substring(0, 30) : ''}`);
+    this.emitEvent('device_profiled', this.deviceProfile);
+
+    // Step 3: Authenticate
+    this.emitProgress('initializing', 40, 'Authenticating with API key...');
+    try {
+      const authRes = await axios.post(`${this.apiUrl}/api/auth/sdk`, {
+        apiKey: this.apiKey,
+      });
+      this.token = authRes.data.token;
+      this.emitProgress('initializing', 60, 'Authenticated successfully');
+      this.emitEvent('auth', { success: true });
+    } catch (err: any) {
+      this.emitProgress('error', 0, `Authentication failed: ${err.message}`);
+      this.emitEvent('error', { stage: 'auth', error: err.message });
+      throw new Error(`SlyOS auth failed: ${err.response?.data?.error || err.message}`);
+    }
+
+    // Step 4: Measure API latency
+    const latency = await measureApiLatency(this.apiUrl);
+    if (latency > 0) this.deviceProfile.latencyToApiMs = latency;
+
+    // Step 5: Register device with full intelligence profile
+    this.emitProgress('initializing', 70, 'Registering device...');
+    try {
+      // Determine supported quantizations based on memory
+      const mem = this.deviceProfile.memoryMB;
+      const supportedQuants: string[] = ['q4'];
+      if (mem >= 4096) supportedQuants.push('q8');
+      if (mem >= 8192) supportedQuants.push('fp16');
+      if (mem >= 16384) supportedQuants.push('fp32');
+
+      // Determine recommended tier
+      let recommendedTier = 1;
+      if (mem >= 8192 && this.deviceProfile.cpuCores >= 4) recommendedTier = 2;
+      if (mem >= 16384 && this.deviceProfile.cpuCores >= 8) recommendedTier = 3;
+
+      await axios.post(`${this.apiUrl}/api/devices/register`, {
+        device_id: this.deviceId,
+        device_fingerprint: this.deviceProfile.deviceFingerprint,
+        platform: this.deviceProfile.platform,
+        os_version: this.deviceProfile.os,
+        total_memory_mb: this.deviceProfile.memoryMB,
+        cpu_cores: this.deviceProfile.cpuCores,
+        // Enhanced fields
+        gpu_renderer: this.deviceProfile.gpuRenderer || null,
+        gpu_vram_mb: this.deviceProfile.gpuVramMb || null,
+        screen_width: this.deviceProfile.screenWidth || null,
+        screen_height: this.deviceProfile.screenHeight || null,
+        pixel_ratio: this.deviceProfile.pixelRatio || null,
+        browser_name: this.deviceProfile.browserName || null,
+        browser_version: this.deviceProfile.browserVersion || null,
+        sdk_version: SDK_VERSION,
+        network_type: this.deviceProfile.networkType || null,
+        latency_to_api_ms: this.deviceProfile.latencyToApiMs || null,
+        timezone: this.deviceProfile.timezone || null,
+        // Capabilities
+        wasm_available: this.deviceProfile.wasmAvailable || false,
+        webgpu_available: this.deviceProfile.webgpuAvailable || false,
+        supported_quants: supportedQuants,
+        recommended_tier: recommendedTier,
+      }, {
+        headers: { Authorization: `Bearer ${this.token}` },
+      });
+      this.emitProgress('initializing', 90, 'Device registered');
+      this.emitEvent('device_registered', { deviceId: this.deviceId, fingerprint: this.deviceProfile.deviceFingerprint });
+    } catch (err: any) {
+      // Non-fatal — device registration shouldn't block usage
+      this.emitProgress('initializing', 90, 'Device registration skipped (non-fatal)');
+    }
+
+    // Step 6: Start telemetry flush timer
+    this.telemetryFlushTimer = setTimeout(() => this.flushTelemetry(), SlyOS.TELEMETRY_FLUSH_INTERVAL);
+
+    this.emitProgress('ready', 100, `SlyOS v${SDK_VERSION} ready — ${this.deviceProfile.recommendedQuant.toUpperCase()}, ${this.deviceProfile.gpuRenderer ? 'GPU detected' : 'CPU only'}`);
+
+    return this.deviceProfile;
+  }
+
+  // ── Model Loading ───────────────────────────────────────────────
+
+  getAvailableModels(): Record<string, { models: { id: string; sizesMB: Record<string, number>; minRAM_MB: Record<string, number> }[] }> {
+    const grouped: Record<string, any[]> = { llm: [], stt: [] };
+    for (const [id, info] of Object.entries(modelMap)) {
+      if (!grouped[info.category]) grouped[info.category] = [];
+      grouped[info.category].push({
+        id,
+        sizesMB: info.sizesMB,
+        minRAM_MB: info.minRAM_MB,
+      });
+    }
+    return Object.fromEntries(
+      Object.entries(grouped).map(([cat, models]) => [cat, { models }])
+    );
+  }
+
+  async searchModels(query: string, options?: { limit?: number; task?: string }): Promise<Array<{
+    id: string;
+    name: string;
+    downloads: number;
+    likes: number;
+    task: string;
+    size_category: string;
+  }>> {
+    try {
+      const limit = options?.limit || 20;
+      const filters = ['onnx']; // Filter for ONNX models only
+      if (options?.task) {
+        filters.push(options.task);
+      }
+
+      const filterString = filters.map(f => `"${f}"`).join(',');
+      const url = `https://huggingface.co/api/models?search=${encodeURIComponent(query)}&filter=${encodeURIComponent(`[${filterString}]`)}&sort=downloads&direction=-1&limit=${limit}`;
+
+      const response = await axios.get(url, { timeout: 10000 });
+      const models = Array.isArray(response.data) ? response.data : [];
+
+      return models.map((model: any) => ({
+        id: model.id,
+        name: model.id.split('/')[1] || model.id,
+        downloads: model.downloads || 0,
+        likes: model.likes || 0,
+        task: model.task || 'unknown',
+        size_category: model.size_category || 'unknown',
+      }));
+    } catch (error: any) {
+      this.emitEvent('error', { stage: 'model_search', error: error.message });
+      throw new Error(`Model search failed: ${error.message}`);
+    }
+  }
+
+  canRunModel(modelId: string, quant?: QuantizationLevel): { canRun: boolean; reason: string; recommendedQuant: QuantizationLevel } {
+    const info = modelMap[modelId];
+    if (!info) return { canRun: false, reason: `Unknown model "${modelId}"`, recommendedQuant: 'q4' };
+    if (!this.deviceProfile) return { canRun: true, reason: 'Device not profiled yet — call initialize() first', recommendedQuant: 'q4' };
+
+    const mem = this.deviceProfile.memoryMB;
+    const bestQuant = selectQuantization(mem, modelId);
+
+    if (quant && mem < info.minRAM_MB[quant]) {
+      return {
+        canRun: false,
+        reason: `Not enough RAM for ${quant.toUpperCase()} (need ${info.minRAM_MB[quant]}MB, have ${mem}MB). Try ${bestQuant.toUpperCase()} instead.`,
+        recommendedQuant: bestQuant,
+      };
+    }
+
+    if (mem < info.minRAM_MB.q4) {
+      return {
+        canRun: false,
+        reason: `Model requires at least ${info.minRAM_MB.q4}MB RAM even at Q4. Device has ${mem}MB.`,
+        recommendedQuant: 'q4',
+      };
+    }
+
+    return { canRun: true, reason: `OK at ${bestQuant.toUpperCase()} precision`, recommendedQuant: bestQuant };
+  }
+
+  async loadModel(modelId: string, options?: { quant?: QuantizationLevel }): Promise<void> {
+    const info = modelMap[modelId];
+    let hfModelId: string;
+    let task: string;
+    let estimatedSize: number;
+
+    // Handle curated models
+    if (info) {
+      hfModelId = info.hfModel;
+      task = info.task;
+
+      // Determine quantization
+      let quant: QuantizationLevel = options?.quant || 'fp32';
+      if (!options?.quant && this.deviceProfile) {
+        quant = selectQuantization(this.deviceProfile.memoryMB, modelId);
+        this.emitProgress('downloading', 0, `Auto-selected ${quant.toUpperCase()} quantization for your device`);
+      }
+
+      // Check feasibility
+      const check = this.canRunModel(modelId, quant);
+      if (!check.canRun) {
+        this.emitProgress('error', 0, check.reason);
+        throw new Error(check.reason);
+      }
+
+      estimatedSize = info.sizesMB[quant];
+      this.emitProgress('downloading', 0, `Downloading ${modelId} (${quant.toUpperCase()}, ~${estimatedSize}MB)...`);
+      this.emitEvent('model_download_start', { modelId, quant, estimatedSizeMB: estimatedSize });
+    } else {
+      // Handle custom HuggingFace models
+      hfModelId = modelId;
+      task = 'text-generation'; // Default task
+      estimatedSize = 2048; // Default estimate
+
+      this.emitProgress('downloading', 0, `Loading custom HuggingFace model: ${modelId}...`);
+      this.emitEvent('model_download_start', { modelId, custom: true, estimatedSizeMB: estimatedSize });
+    }
+
+    // Map quant to dtype for HuggingFace
+    const dtypeMap: Record<QuantizationLevel, string> = {
+      q4: 'q4',
+      q8: 'q8',
+      fp16: 'fp16',
+      fp32: 'fp32',
+    };
+
+    let lastReportedPercent = 0;
+    const startTime = Date.now();
+
+    try {
+      // For custom HF models, detect context window
+      let detectedContextWindow = 2048;
+      if (!info) {
+        detectedContextWindow = await detectContextWindowFromHF(hfModelId);
+      }
+
+      const pipe = await pipeline(task as any, hfModelId, {
+        device: 'cpu',
+        dtype: 'q4' as any, // Default to q4 for stability
+        progress_callback: (progressData: any) => {
+          // HuggingFace transformers sends progress events during download
+          if (progressData && typeof progressData === 'object') {
+            let percent = 0;
+            let msg = 'Downloading...';
+
+            if (progressData.status === 'progress' && progressData.progress !== undefined) {
+              percent = Math.round(progressData.progress);
+              const loaded = progressData.loaded ? `${Math.round(progressData.loaded / 1024 / 1024)}MB` : '';
+              const total = progressData.total ? `${Math.round(progressData.total / 1024 / 1024)}MB` : '';
+              msg = loaded && total ? `Downloading: ${loaded} / ${total}` : `Downloading: ${percent}%`;
+            } else if (progressData.status === 'done') {
+              percent = 100;
+              msg = progressData.file ? `Downloaded ${progressData.file}` : 'Download complete';
+            } else if (progressData.status === 'initiate') {
+              msg = progressData.file ? `Starting download: ${progressData.file}` : 'Initiating download...';
+            }
+
+            // Only emit if progress meaningfully changed (avoid flooding)
+            if (percent !== lastReportedPercent || progressData.status === 'done' || progressData.status === 'initiate') {
+              lastReportedPercent = percent;
+              this.emitProgress('downloading', percent, msg, progressData);
+              this.emitEvent('model_download_progress', { modelId, percent, ...progressData });
+            }
+          }
+        },
+      });
+
+      const loadTime = Date.now() - startTime;
+      let contextWindow: number;
+
+      if (info) {
+        // For curated models, use recommendContextWindow
+        const quant = options?.quant || (this.deviceProfile ? selectQuantization(this.deviceProfile.memoryMB, modelId) : 'q4');
+        contextWindow = this.deviceProfile
+          ? recommendContextWindow(this.deviceProfile.memoryMB, quant)
+          : 2048;
+      } else {
+        // For custom HF models, use detected context window
+        contextWindow = detectedContextWindow;
+      }
+
+      this.modelContextWindow = contextWindow;
+      this.models.set(modelId, { pipe, info, quant: 'q4', contextWindow });
+
+      this.emitProgress('ready', 100, `${modelId} loaded (q4, ${(loadTime / 1000).toFixed(1)}s, ctx: ${contextWindow})`);
+      this.emitEvent('model_loaded', { modelId, quant: 'q4', loadTimeMs: loadTime, contextWindow });
+
+      // Telemetry
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'model_load',
+          model_id: modelId,
+          success: true,
+          metadata: { quant: 'q4', loadTimeMs: loadTime, contextWindow, custom: !info },
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Failed to load ${modelId}: ${error.message}`);
+      this.emitEvent('error', { stage: 'model_load', modelId, error: error.message });
+
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'model_load',
+          model_id: modelId,
+          success: false,
+          error_message: error.message,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+      throw error;
+    }
+  }
+
+  // ── Inference: Generate ─────────────────────────────────────────
+
+  async generate(modelId: string, prompt: string | Array<{role: string; content: string}>, options: GenerateOptions = {}): Promise<string> {
+    if (!this.models.has(modelId)) {
+      await this.loadModel(modelId);
+    }
+
+    const loaded = this.models.get(modelId);
+    if (!loaded) {
+      throw new Error(`Model "${modelId}" failed to load. Check your connection and model ID.`);
+    }
+    const { pipe, info, contextWindow } = loaded;
+    if (info.category !== 'llm') {
+      throw new Error(`Model "${modelId}" is not an LLM. Use transcribe() for STT models.`);
+    }
+
+    const maxTokens = Math.min(options.maxTokens || 100, contextWindow || 2048);
+    const isMessages = Array.isArray(prompt);
+
+    this.emitProgress('generating', 0, `Generating response (max ${maxTokens} tokens)...`);
+    this.emitEvent('inference_start', { modelId, maxTokens });
+    const startTime = Date.now();
+
+    try {
+      const result = await pipe(prompt, {
+        max_new_tokens: maxTokens,
+        temperature: options.temperature || 0.7,
+        top_p: options.topP || 0.9,
+        do_sample: true,
+        repetition_penalty: 1.1,
+      });
+
+      let response: string;
+      if (isMessages) {
+        // When using messages format, the pipeline returns the assistant's reply
+        // in the last message of the generated conversation
+        const generated = result[0].generated_text;
+        if (Array.isArray(generated)) {
+          // Transformers.js returns messages array — extract assistant reply
+          const assistantMsg = generated.filter((m: any) => m.role === 'assistant').pop();
+          response = assistantMsg?.content?.trim() || '';
+        } else {
+          response = typeof generated === 'string' ? generated.trim() : '';
+        }
+      } else {
+        const rawOutput = result[0].generated_text;
+        // HuggingFace transformers returns the prompt + generated text concatenated.
+        // Strip the original prompt so we only return the NEW tokens.
+        response = (typeof rawOutput === 'string' && rawOutput.startsWith(prompt as string))
+          ? rawOutput.slice((prompt as string).length).trim()
+          : (typeof rawOutput === 'string' ? rawOutput.trim() : '');
+      }
+      const latency = Date.now() - startTime;
+      const tokensGenerated = response.split(/\s+/).length;
+      const tokensPerSec = (tokensGenerated / (latency / 1000)).toFixed(1);
+
+      this.emitProgress('ready', 100, `Generated ${tokensGenerated} tokens in ${(latency / 1000).toFixed(1)}s (${tokensPerSec} tok/s)`);
+      this.emitEvent('inference_complete', { modelId, latencyMs: latency, tokensGenerated, tokensPerSec: parseFloat(tokensPerSec) });
+
+      // Batch telemetry (new device intelligence)
+      this.recordTelemetry({
+        latency_ms: latency,
+        tokens_generated: tokensGenerated,
+        success: true,
+        model_id: modelId,
+        timestamp: Date.now(),
+      });
+
+      // Legacy telemetry (backwards compatible)
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'inference',
+          model_id: modelId,
+          latency_ms: latency,
+          tokens_generated: tokensGenerated,
+          success: true,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+
+      return response;
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Generation failed: ${error.message}`);
+      this.emitEvent('error', { stage: 'inference', modelId, error: error.message });
+
+      // Batch telemetry (failure)
+      this.recordTelemetry({
+        latency_ms: 0,
+        tokens_generated: 0,
+        success: false,
+        model_id: modelId,
+        timestamp: Date.now(),
+      });
+
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'inference',
+          model_id: modelId,
+          success: false,
+          error_message: error.message,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+      throw error;
+    }
+  }
+
+  /**
+   * Stream text generation token-by-token.
+   * Calls onToken callback for each generated token.
+   */
+  async generateStream(
+    modelId: string,
+    prompt: string | Array<{role: string; content: string}>,
+    options: GenerateOptions & { onToken?: (token: string, partial: string) => void } = {}
+  ): Promise<{ text: string; firstTokenMs: number; totalMs: number; tokensGenerated: number }> {
+    if (!this.models.has(modelId)) {
+      await this.loadModel(modelId);
+    }
+    const loaded = this.models.get(modelId);
+    if (!loaded) throw new Error(`Model "${modelId}" not loaded`);
+    const { pipe, info, contextWindow } = loaded;
+    if (info.category !== 'llm') throw new Error(`Not an LLM`);
+
+    const maxTokens = Math.min(options.maxTokens || 100, contextWindow || 2048);
+    const isMessages = Array.isArray(prompt);
+    const startTime = Date.now();
+    let firstTokenTime = 0;
+    let accumulated = '';
+    let prevText = '';
+    let callbackCount = 0;
+
+    this.emitProgress('generating', 0, `Streaming (max ${maxTokens} tokens)...`);
+
+    try {
+      const result = await pipe(prompt, {
+        max_new_tokens: maxTokens,
+        temperature: options.temperature || 0.7,
+        top_p: options.topP || 0.9,
+        do_sample: true,
+        repetition_penalty: 1.1,
+        // Transformers.js v3 streamer callback — receives decoded output tokens
+        callback_function: (output: any) => {
+          callbackCount++;
+          if (!firstTokenTime) firstTokenTime = Date.now() - startTime;
+
+          // Transformers.js v3 callback_function may receive:
+          // 1. A string (decoded text so far) in some pipeline configurations
+          // 2. Token IDs array/tensor in others
+          // We handle both cases
+          let tokenText = '';
+          if (typeof output === 'string') {
+            tokenText = output;
+          } else if (output && typeof output === 'object') {
+            // For newer Transformers.js: try to extract text if available
+            if (output.text) tokenText = output.text;
+          }
+
+          if (tokenText && tokenText !== prevText) {
+            const newPart = tokenText.startsWith(prevText) ? tokenText.slice(prevText.length) : tokenText;
+            prevText = tokenText;
+            if (newPart) {
+              accumulated += newPart;
+              options.onToken?.(newPart, accumulated);
+              this.emitEvent('token', { token: newPart, partial: accumulated });
+            }
+          }
+        }
+      });
+
+      let response: string;
+      if (isMessages) {
+        const generated = result[0].generated_text;
+        if (Array.isArray(generated)) {
+          const assistantMsg = generated.filter((m: any) => m.role === 'assistant').pop();
+          response = assistantMsg?.content?.trim() || '';
+        } else {
+          response = typeof generated === 'string' ? generated.trim() : '';
+        }
+      } else {
+        const rawOutput = result[0].generated_text;
+        response = (typeof rawOutput === 'string' && rawOutput.startsWith(prompt as string))
+          ? rawOutput.slice((prompt as string).length).trim()
+          : (typeof rawOutput === 'string' ? rawOutput.trim() : '');
+      }
+
+      if (!firstTokenTime) firstTokenTime = Date.now() - startTime;
+      const totalMs = Date.now() - startTime;
+      const tokensGenerated = response.split(/\s+/).filter(Boolean).length;
+
+      this.emitProgress('ready', 100, `Streamed ${tokensGenerated} tokens in ${(totalMs/1000).toFixed(1)}s`);
+      this.emitEvent('inference_complete', { modelId, latencyMs: totalMs, tokensGenerated });
+
+      // Batch telemetry (device intelligence)
+      this.recordTelemetry({
+        latency_ms: totalMs,
+        tokens_generated: tokensGenerated,
+        success: true,
+        model_id: modelId,
+        timestamp: Date.now(),
+      });
+
+      // Legacy telemetry — updates analytics_daily for dashboard counts
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'inference',
+          model_id: modelId,
+          latency_ms: totalMs,
+          tokens_generated: tokensGenerated,
+          success: true,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+
+      return { text: response, firstTokenMs: firstTokenTime, totalMs, tokensGenerated };
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Stream failed: ${error.message}`);
+      throw error;
+    }
+  }
+
+  // ── Inference: Transcribe ───────────────────────────────────────
+
+  async transcribe(modelId: string, audioInput: any, options: TranscribeOptions = {}): Promise<string> {
+    if (!this.models.has(modelId)) {
+      await this.loadModel(modelId);
+    }
+
+    const loaded = this.models.get(modelId);
+    if (!loaded) {
+      throw new Error(`Model "${modelId}" failed to load. Check your connection and model ID.`);
+    }
+    const { pipe, info } = loaded;
+    if (info.category !== 'stt') {
+      throw new Error(`Model "${modelId}" is not an STT model. Use generate() for LLMs.`);
+    }
+
+    this.emitProgress('transcribing', 0, 'Transcribing audio...');
+    this.emitEvent('inference_start', { modelId, type: 'transcription' });
+    const startTime = Date.now();
+
+    try {
+      const result = await pipe(audioInput, {
+        language: options.language || 'en',
+        return_timestamps: options.returnTimestamps || false,
+      });
+
+      const text = result.text;
+      const latency = Date.now() - startTime;
+
+      this.emitProgress('ready', 100, `Transcribed in ${(latency / 1000).toFixed(1)}s`);
+      this.emitEvent('inference_complete', { modelId, latencyMs: latency, type: 'transcription' });
+
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'inference',
+          model_id: modelId,
+          latency_ms: latency,
+          success: true,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+
+      return text;
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Transcription failed: ${error.message}`);
+      this.emitEvent('error', { stage: 'transcription', modelId, error: error.message });
+
+      if (this.token) {
+        await axios.post(`${this.apiUrl}/api/telemetry`, {
+          device_id: this.deviceId,
+          event_type: 'inference',
+          model_id: modelId,
+          success: false,
+          error_message: error.message,
+        }, {
+          headers: { Authorization: `Bearer ${this.token}` },
+        }).catch(() => {});
+      }
+      throw error;
+    }
+  }
+
+  // ── OpenAI Compatibility ────────────────────────────────────────────
+
+  async chatCompletion(modelId: string, request: OpenAIChatCompletionRequest): Promise<OpenAIChatCompletionResponse> {
+    try {
+      // Pass messages directly to generate() — Transformers.js v3 applies the model's
+      // chat template automatically, which produces much better results than raw text
+      const messages = request.messages.map(msg => ({
+        role: msg.role,
+        content: msg.content,
+      }));
+
+      const response = await this.generate(modelId, messages, {
+        temperature: request.temperature,
+        maxTokens: request.max_tokens,
+        topP: request.top_p,
+      });
+
+      // Estimate token counts (rough approximation: ~4 chars per token)
+      const promptTokens = Math.ceil(prompt.length / 4);
+      const completionTokens = Math.ceil(response.length / 4);
+
+      return {
+        id: `chat-${Date.now()}-${Math.random().toString(36).substr(2, 9)}`,
+        object: 'chat.completion',
+        created: Math.floor(Date.now() / 1000),
+        model: modelId,
+        choices: [
+          {
+            index: 0,
+            message: {
+              role: 'assistant',
+              content: response,
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: promptTokens,
+          completion_tokens: completionTokens,
+          total_tokens: promptTokens + completionTokens,
+        },
+      };
+    } catch (error: any) {
+      // Fallback to cloud provider if configured
+      if (this.fallbackConfig?.provider === 'openai') {
+        return this.fallbackToOpenAI(modelId, request);
+      } else if (this.fallbackConfig?.provider === 'bedrock') {
+        return this.fallbackToBedrock(modelId, request);
+      }
+      throw error;
+    }
+  }
+
+  // ── AWS Bedrock Compatibility ──────────────────────────────────────
+
+  async bedrockInvoke(modelId: string, request: BedrockInvokeRequest): Promise<BedrockInvokeResponse> {
+    try {
+      const response = await this.generate(modelId, request.inputText, {
+        temperature: request.textGenerationConfig?.temperature,
+        maxTokens: request.textGenerationConfig?.maxTokenCount,
+        topP: request.textGenerationConfig?.topP,
+      });
+
+      // Estimate token counts
+      const inputTokens = Math.ceil(request.inputText.length / 4);
+      const outputTokens = Math.ceil(response.length / 4);
+
+      return {
+        results: [
+          {
+            outputText: response,
+            tokenCount: outputTokens,
+          },
+        ],
+        input_text_token_count: inputTokens,
+      };
+    } catch (error: any) {
+      // Fallback to cloud provider if configured
+      if (this.fallbackConfig?.provider === 'bedrock') {
+        return this.fallbackToBedrockCloud(modelId, request);
+      } else if (this.fallbackConfig?.provider === 'openai') {
+        return this.fallbackToOpenAICloud(modelId, request);
+      }
+      throw error;
+    }
+  }
+
+  // ── Fallback: OpenAI Cloud ────────────────────────────────────────
+
+  private async fallbackToOpenAI(modelId: string, request: OpenAIChatCompletionRequest): Promise<OpenAIChatCompletionResponse> {
+    if (!this.fallbackConfig) {
+      throw new Error('OpenAI fallback not configured');
+    }
+
+    const mappedModel = this.mapModelToOpenAI(modelId);
+    const payload = {
+      model: this.fallbackConfig.model || mappedModel,
+      messages: request.messages,
+      temperature: request.temperature,
+      max_tokens: request.max_tokens,
+      top_p: request.top_p,
+      frequency_penalty: request.frequency_penalty,
+      presence_penalty: request.presence_penalty,
+      stop: request.stop,
+    };
+
+    try {
+      const response = await axios.post('https://api.openai.com/v1/chat/completions', payload, {
+        headers: {
+          Authorization: `Bearer ${this.fallbackConfig.apiKey}`,
+          'Content-Type': 'application/json',
+        },
+      });
+
+      this.emitEvent('fallback_success', { provider: 'openai', originalModel: modelId, mappedModel: this.fallbackConfig.model });
+      return response.data;
+    } catch (error: any) {
+      this.emitProgress('error', 0, `OpenAI fallback failed: ${error.message}`);
+      this.emitEvent('fallback_error', { provider: 'openai', error: error.message });
+      throw error;
+    }
+  }
+
+  private async fallbackToBedrock(modelId: string, request: OpenAIChatCompletionRequest): Promise<OpenAIChatCompletionResponse> {
+    if (!this.fallbackConfig) {
+      throw new Error('Bedrock fallback not configured');
+    }
+
+    // Convert OpenAI format to Bedrock's expected format (simplified)
+    const lastMessage = request.messages[request.messages.length - 1];
+    const inputText = lastMessage.content;
+
+    const bedrockResponse = await this.invokeBedrockCloud(inputText, {
+      temperature: request.temperature,
+      maxTokenCount: request.max_tokens,
+      topP: request.top_p,
+    });
+
+    // Convert Bedrock response back to OpenAI format
+    const promptTokens = Math.ceil(inputText.length / 4);
+    const completionTokens = bedrockResponse.results[0].tokenCount;
+
+    this.emitEvent('fallback_success', { provider: 'bedrock', originalModel: modelId, mappedModel: this.fallbackConfig.model });
+
+    return {
+      id: `chat-${Date.now()}-${Math.random().toString(36).substr(2, 9)}`,
+      object: 'chat.completion',
+      created: Math.floor(Date.now() / 1000),
+      model: modelId,
+      choices: [
+        {
+          index: 0,
+          message: {
+            role: 'assistant',
+            content: bedrockResponse.results[0].outputText,
+          },
+          finish_reason: 'stop',
+        },
+      ],
+      usage: {
+        prompt_tokens: promptTokens,
+        completion_tokens: completionTokens,
+        total_tokens: promptTokens + completionTokens,
+      },
+    };
+  }
+
+  private async fallbackToOpenAICloud(modelId: string, request: BedrockInvokeRequest): Promise<BedrockInvokeResponse> {
+    if (!this.fallbackConfig) {
+      throw new Error('OpenAI fallback not configured');
+    }
+
+    const mappedModel = this.mapModelToOpenAI(modelId);
+    const payload = {
+      model: this.fallbackConfig.model || mappedModel,
+      messages: [{ role: 'user', content: request.inputText }],
+      temperature: request.textGenerationConfig?.temperature,
+      max_tokens: request.textGenerationConfig?.maxTokenCount,
+      top_p: request.textGenerationConfig?.topP,
+    };
+
+    try {
+      const response = await axios.post('https://api.openai.com/v1/chat/completions', payload, {
+        headers: {
+          Authorization: `Bearer ${this.fallbackConfig.apiKey}`,
+          'Content-Type': 'application/json',
+        },
+      });
+
+      const outputText = response.data.choices[0].message.content;
+      const inputTokens = Math.ceil(request.inputText.length / 4);
+      const outputTokens = response.data.usage.completion_tokens;
+
+      this.emitEvent('fallback_success', { provider: 'openai', originalModel: modelId, mappedModel: this.fallbackConfig.model });
+
+      return {
+        results: [
+          {
+            outputText,
+            tokenCount: outputTokens,
+          },
+        ],
+        input_text_token_count: inputTokens,
+      };
+    } catch (error: any) {
+      this.emitProgress('error', 0, `OpenAI fallback failed: ${error.message}`);
+      this.emitEvent('fallback_error', { provider: 'openai', error: error.message });
+      throw error;
+    }
+  }
+
+  private async fallbackToBedrockCloud(modelId: string, request: BedrockInvokeRequest): Promise<BedrockInvokeResponse> {
+    if (!this.fallbackConfig) {
+      throw new Error('Bedrock fallback not configured');
+    }
+
+    try {
+      return await this.invokeBedrockCloud(request.inputText, request.textGenerationConfig);
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Bedrock fallback failed: ${error.message}`);
+      this.emitEvent('fallback_error', { provider: 'bedrock', error: error.message });
+      throw error;
+    }
+  }
+
+  private async invokeBedrockCloud(inputText: string, config?: BedrockTextGenerationConfig): Promise<BedrockInvokeResponse> {
+    if (!this.fallbackConfig) {
+      throw new Error('Bedrock fallback not configured');
+    }
+
+    const region = this.fallbackConfig.region || 'us-east-1';
+    const model = this.fallbackConfig.model || 'anthropic.claude-3-sonnet-20240229-v1:0';
+
+    // Bedrock endpoint format: https://bedrock-runtime.{region}.amazonaws.com/model/{modelId}/invoke
+    const endpoint = `https://bedrock-runtime.${region}.amazonaws.com/model/${model}/invoke`;
+
+    const payload = {
+      inputText,
+      textGenerationConfig: {
+        maxTokenCount: config?.maxTokenCount || 256,
+        temperature: config?.temperature || 0.7,
+        topP: config?.topP || 0.9,
+        topK: config?.topK,
+        stopSequences: config?.stopSequences,
+      },
+    };
+
+    try {
+      const response = await axios.post(endpoint, payload, {
+        headers: {
+          Authorization: `Bearer ${this.fallbackConfig.apiKey}`,
+          'Content-Type': 'application/json',
+          'X-Amz-Target': 'AmazonBedrockRuntime.InvokeModel',
+        },
+      });
+
+      this.emitEvent('fallback_success', { provider: 'bedrock', model });
+      return response.data;
+    } catch (error: any) {
+      throw new Error(`Bedrock invocation failed: ${error.message}`);
+    }
+  }
+
+  private mapModelToOpenAI(slyModelId: string): string {
+    const modelMapping: Record<string, string> = {
+      'quantum-1.7b': 'gpt-4o-mini',
+      'quantum-3b': 'gpt-4o',
+      'quantum-code-3b': 'gpt-4o',
+      'quantum-8b': 'gpt-4-turbo',
+    };
+    return modelMapping[slyModelId] || 'gpt-4o-mini';
+  }
+
+  // ═══════════════════════════════════════════════════════════
+  // RAG — Retrieval Augmented Generation
+  // ═══════════════════════════════════════════════════════════
+
+  private localEmbeddingModel: any = null;
+  private offlineIndexes: Map<string, OfflineIndex> = new Map();
+
+  /**
+   * Compute dynamic RAG parameters based on device profile and model.
+   */
+  private computeRAGConfig(modelId: string): {
+    maxContextChars: number;
+    maxGenTokens: number;
+    chunkSize: number;
+    topK: number;
+    contextWindow: number;
+    deviceTier: 'low' | 'mid' | 'high';
+  } {
+    const contextWindow = this.modelContextWindow || 2048;
+    const memoryMB = this.deviceProfile?.memoryMB || 4096;
+    const cpuCores = this.deviceProfile?.cpuCores || 4;
+    const hasGPU = !!(this.deviceProfile?.gpuRenderer || this.deviceProfile?.webgpuAvailable);
+
+    // Determine device tier
+    let deviceTier: 'low' | 'mid' | 'high' = 'low';
+    if (memoryMB >= 8192 && cpuCores >= 8) deviceTier = 'high';
+    else if (memoryMB >= 4096 && cpuCores >= 4) deviceTier = 'mid';
+
+    // Context chars: scale with context window AND device capability
+    let maxContextChars: number;
+    if (contextWindow <= 2048) {
+      maxContextChars = deviceTier === 'high' ? 600 : deviceTier === 'mid' ? 400 : 300;
+    } else if (contextWindow <= 4096) {
+      maxContextChars = deviceTier === 'high' ? 1500 : deviceTier === 'mid' ? 1000 : 600;
+    } else {
+      maxContextChars = deviceTier === 'high' ? 3000 : deviceTier === 'mid' ? 2000 : 1000;
+    }
+
+    // Gen tokens: scale with device tier
+    let maxGenTokens: number;
+    if (contextWindow <= 2048) {
+      maxGenTokens = deviceTier === 'high' ? 200 : deviceTier === 'mid' ? 150 : 100;
+    } else {
+      maxGenTokens = deviceTier === 'high' ? 400 : deviceTier === 'mid' ? 300 : 150;
+    }
+
+    // Chunk size: larger chunks for bigger context windows
+    const chunkSize = contextWindow <= 2048 ? 256 : contextWindow <= 4096 ? 512 : 1024;
+
+    // TopK: more chunks for powerful devices
+    const topK = deviceTier === 'high' ? 5 : deviceTier === 'mid' ? 3 : 1;
+
+    return { maxContextChars, maxGenTokens, chunkSize, topK, contextWindow, deviceTier };
+  }
+
+  /**
+   * Tier 2: Cloud-indexed RAG with local inference.
+   * Retrieves relevant chunks from server, generates response locally.
+   */
+  async ragQuery(options: RAGOptions): Promise<RAGResponse> {
+    const startTime = Date.now();
+
+    try {
+      if (!this.token) throw new Error('Not authenticated. Call init() first.');
+
+      const ragConfig = this.computeRAGConfig(options.modelId);
+
+      // Step 1: Retrieve relevant chunks from backend
+      const retrievalStart = Date.now();
+      const searchResponse = await axios.post(
+        `${this.apiUrl}/api/rag/knowledge-bases/${options.knowledgeBaseId}/query`,
+        {
+          query: options.query,
+          top_k: options.topK || ragConfig.topK,
+          model_id: options.modelId
+        },
+        { headers: { Authorization: `Bearer ${this.token}` } }
+      );
+      const retrievalMs = Date.now() - retrievalStart;
+
+      let { retrieved_chunks, context } = searchResponse.data;
+
+      // Step 2: Build context with dynamic limits
+      const contextBuildStart = Date.now();
+      if (context && context.length > ragConfig.maxContextChars) {
+        context = context.substring(0, ragConfig.maxContextChars);
+      }
+
+      // Build messages array for proper chat template handling
+      // This uses the model's built-in chat template (e.g. <|im_start|> for SmolLM/Qwen)
+      // which produces dramatically better results than raw text prompts
+      const messages: Array<{role: string; content: string}> = [
+        { role: 'system', content: `Answer questions using only the following context. Be concise.\n\n${context}` },
+        { role: 'user', content: options.query },
+      ];
+      const contextBuildMs = Date.now() - contextBuildStart;
+
+      // Step 3: Generate response — stream if callback provided
+      const genStart = Date.now();
+      let response: string;
+      let firstTokenMs = 0;
+
+      if (options.onToken) {
+        const streamResult = await this.generateStream(options.modelId, messages, {
+          temperature: options.temperature,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+          onToken: options.onToken,
+        });
+        response = streamResult.text;
+        firstTokenMs = streamResult.firstTokenMs;
+      } else {
+        response = await this.generate(options.modelId, messages, {
+          temperature: options.temperature,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+        });
+        firstTokenMs = Date.now() - genStart; // approximate
+      }
+      const generationMs = Date.now() - genStart;
+      const totalMs = Date.now() - startTime;
+      const tokensGenerated = response.split(/\s+/).length;
+
+      return {
+        query: options.query,
+        retrievedChunks: (retrieved_chunks || []).map((c: any) => ({
+          id: c.id,
+          documentId: c.document_id,
+          documentName: c.document_name,
+          content: c.content,
+          similarityScore: c.similarity_score,
+          metadata: c.metadata
+        })),
+        generatedResponse: response,
+        context,
+        latencyMs: totalMs,
+        tierUsed: 2,
+        timing: {
+          retrievalMs,
+          contextBuildMs,
+          firstTokenMs,
+          generationMs,
+          totalMs,
+          tokensGenerated,
+          tokensPerSecond: generationMs > 0 ? tokensGenerated / (generationMs / 1000) : 0,
+        },
+        config: {
+          maxContextChars: ragConfig.maxContextChars,
+          maxGenTokens: ragConfig.maxGenTokens,
+          chunkSize: ragConfig.chunkSize,
+          topK: options.topK || ragConfig.topK,
+          contextWindowUsed: ragConfig.contextWindow,
+          deviceTier: ragConfig.deviceTier,
+        },
+      };
+    } catch (error: any) {
+      this.emitEvent('error', { stage: 'rag_query', error: error.message });
+      throw new Error(`RAG query failed: ${error.message}`);
+    }
+  }
+
+  /**
+   * Tier 1: Fully local RAG. Zero network calls.
+   * Documents are chunked/embedded on-device, retrieval and generation all local.
+   */
+  async ragQueryLocal(options: RAGOptions & { documents: Array<{ content: string; name?: string }> }): Promise<RAGResponse> {
+    const startTime = Date.now();
+
+    try {
+      const ragConfig = this.computeRAGConfig(options.modelId);
+
+      // Step 1: Load embedding model if needed
+      if (!this.localEmbeddingModel) {
+        await this.loadEmbeddingModel();
+      }
+
+      // Step 2: Chunk and embed documents (dynamic chunk size)
+      const retrievalStart = Date.now();
+      const allChunks: Array<{ content: string; documentName: string; embedding?: number[] }> = [];
+      for (const doc of options.documents) {
+        const chunks = this.chunkTextLocal(doc.content, ragConfig.chunkSize, Math.floor(ragConfig.chunkSize / 4));
+        for (const chunk of chunks) {
+          const embedding = await this.embedTextLocal(chunk);
+          allChunks.push({ content: chunk, documentName: doc.name || 'Document', embedding });
+        }
+      }
+
+      // Step 3: Embed query and search
+      const queryEmbedding = await this.embedTextLocal(options.query);
+      const scored = allChunks
+        .filter(c => c.embedding)
+        .map(c => ({ ...c, similarityScore: this.cosineSimilarity(queryEmbedding, c.embedding!) }))
+        .sort((a, b) => b.similarityScore - a.similarityScore)
+        .slice(0, options.topK || ragConfig.topK);
+      const retrievalMs = Date.now() - retrievalStart;
+
+      // Step 4: Build context
+      const contextBuildStart = Date.now();
+      const bestChunk = scored[0];
+      let context = bestChunk.content
+        .replace(/[^\x20-\x7E\n]/g, ' ')
+        .replace(/\s{2,}/g, ' ')
+        .replace(/<[^>]+>/g, ' ')
+        .replace(/https?:\/\/\S+/g, '')
+        .replace(/[{}()\[\]]/g, '')
+        .trim();
+      if (context.length > ragConfig.maxContextChars) context = context.substring(0, ragConfig.maxContextChars);
+      const messages: Array<{role: string; content: string}> = [
+        { role: 'system', content: `Answer questions using only the following context. Be concise.\n\n${context}` },
+        { role: 'user', content: options.query },
+      ];
+      const contextBuildMs = Date.now() - contextBuildStart;
+
+      // Step 5: Generate — stream if callback provided
+      const genStart = Date.now();
+      let response: string;
+      let firstTokenMs = 0;
+
+      if (options.onToken) {
+        const streamResult = await this.generateStream(options.modelId, messages, {
+          temperature: options.temperature || 0.6,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+          onToken: options.onToken,
+        });
+        response = streamResult.text;
+        firstTokenMs = streamResult.firstTokenMs;
+      } else {
+        response = await this.generate(options.modelId, messages, {
+          temperature: options.temperature || 0.6,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+        });
+        firstTokenMs = Date.now() - genStart;
+      }
+      const generationMs = Date.now() - genStart;
+      const totalMs = Date.now() - startTime;
+      const tokensGenerated = response.split(/\s+/).length;
+
+      return {
+        query: options.query,
+        retrievedChunks: scored.map((c, i) => ({
+          id: `local-${i}`,
+          documentId: 'local',
+          documentName: c.documentName,
+          content: c.content,
+          similarityScore: c.similarityScore,
+          metadata: {}
+        })),
+        generatedResponse: response,
+        context,
+        latencyMs: totalMs,
+        tierUsed: 1,
+        timing: {
+          retrievalMs,
+          contextBuildMs,
+          firstTokenMs,
+          generationMs,
+          totalMs,
+          tokensGenerated,
+          tokensPerSecond: generationMs > 0 ? tokensGenerated / (generationMs / 1000) : 0,
+        },
+        config: {
+          maxContextChars: ragConfig.maxContextChars,
+          maxGenTokens: ragConfig.maxGenTokens,
+          chunkSize: ragConfig.chunkSize,
+          topK: options.topK || ragConfig.topK,
+          contextWindowUsed: ragConfig.contextWindow,
+          deviceTier: ragConfig.deviceTier,
+        },
+      };
+    } catch (error: any) {
+      this.emitEvent('error', { stage: 'rag_local', error: error.message });
+      throw new Error(`Local RAG failed: ${error.message}`);
+    }
+  }
+
+  /**
+   * Tier 3: Offline RAG using a synced knowledge base.
+   * First call syncKnowledgeBase(), then use this for offline queries.
+   */
+  async ragQueryOffline(options: RAGOptions): Promise<RAGResponse> {
+    const startTime = Date.now();
+
+    const index = this.offlineIndexes.get(options.knowledgeBaseId);
+    if (!index) throw new Error(`KB "${options.knowledgeBaseId}" not synced.`);
+    if (new Date(index.metadata.expires_at) < new Date()) throw new Error('Offline index expired.');
+
+    try {
+      const ragConfig = this.computeRAGConfig(options.modelId);
+
+      // Load embedding model
+      if (!this.localEmbeddingModel) await this.loadEmbeddingModel();
+
+      // Search offline index
+      const retrievalStart = Date.now();
+      const queryEmbedding = await this.embedTextLocal(options.query);
+      const scored = index.chunks
+        .filter(c => c.embedding && c.embedding.length > 0)
+        .map(c => ({ ...c, similarityScore: this.cosineSimilarity(queryEmbedding, c.embedding!) }))
+        .sort((a, b) => b.similarityScore - a.similarityScore)
+        .slice(0, options.topK || ragConfig.topK);
+      const retrievalMs = Date.now() - retrievalStart;
+
+      // Build context
+      const contextBuildStart = Date.now();
+      const bestChunk = scored[0];
+      let context = bestChunk.content
+        .replace(/[^\x20-\x7E\n]/g, ' ')
+        .replace(/\s{2,}/g, ' ')
+        .replace(/<[^>]+>/g, ' ')
+        .replace(/https?:\/\/\S+/g, '')
+        .replace(/[{}()\[\]]/g, '')
+        .trim();
+      if (context.length > ragConfig.maxContextChars) context = context.substring(0, ragConfig.maxContextChars);
+      const messages: Array<{role: string; content: string}> = [
+        { role: 'system', content: `Answer questions using only the following context. Be concise.\n\n${context}` },
+        { role: 'user', content: options.query },
+      ];
+      const contextBuildMs = Date.now() - contextBuildStart;
+
+      // Generate
+      const genStart = Date.now();
+      let response: string;
+      let firstTokenMs = 0;
+
+      if (options.onToken) {
+        const streamResult = await this.generateStream(options.modelId, messages, {
+          temperature: options.temperature || 0.6,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+          onToken: options.onToken,
+        });
+        response = streamResult.text;
+        firstTokenMs = streamResult.firstTokenMs;
+      } else {
+        response = await this.generate(options.modelId, messages, {
+          temperature: options.temperature || 0.6,
+          maxTokens: options.maxTokens || ragConfig.maxGenTokens,
+        });
+        firstTokenMs = Date.now() - genStart;
+      }
+      const generationMs = Date.now() - genStart;
+      const totalMs = Date.now() - startTime;
+      const tokensGenerated = response.split(/\s+/).length;
+
+      return {
+        query: options.query,
+        retrievedChunks: scored.map(c => ({
+          id: c.id,
+          documentId: c.document_id,
+          documentName: c.document_name,
+          content: c.content,
+          similarityScore: c.similarityScore,
+          metadata: c.metadata
+        })),
+        generatedResponse: response,
+        context,
+        latencyMs: totalMs,
+        tierUsed: 3,
+        timing: {
+          retrievalMs,
+          contextBuildMs,
+          firstTokenMs,
+          generationMs,
+          totalMs,
+          tokensGenerated,
+          tokensPerSecond: generationMs > 0 ? tokensGenerated / (generationMs / 1000) : 0,
+        },
+        config: {
+          maxContextChars: ragConfig.maxContextChars,
+          maxGenTokens: ragConfig.maxGenTokens,
+          chunkSize: ragConfig.chunkSize,
+          topK: options.topK || ragConfig.topK,
+          contextWindowUsed: ragConfig.contextWindow,
+          deviceTier: ragConfig.deviceTier,
+        },
+      };
+    } catch (error: any) {
+      this.emitEvent('error', { stage: 'rag_offline', error: error.message });
+      throw new Error(`Offline RAG failed: ${error.message}`);
+    }
+  }
+
+  /**
+   * Sync a knowledge base for offline use (Tier 3).
+   * Downloads chunks + embeddings from server, stores locally.
+   */
+  async syncKnowledgeBase(knowledgeBaseId: string, deviceId?: string): Promise<{ chunkCount: number; sizeMb: number; expiresAt: string }> {
+    try {
+      if (!this.token) throw new Error('Not authenticated. Call init() first.');
+
+      const response = await axios.post(
+        `${this.apiUrl}/api/rag/knowledge-bases/${knowledgeBaseId}/sync`,
+        { device_id: deviceId || this.deviceId || 'sdk-device' },
+        { headers: { Authorization: `Bearer ${this.token}` } }
+      );
+
+      const { sync_package, chunk_count, package_size_mb, expires_at } = response.data;
+      this.offlineIndexes.set(knowledgeBaseId, sync_package);
+
+      return {
+        chunkCount: chunk_count,
+        sizeMb: package_size_mb,
+        expiresAt: expires_at
+      };
+    } catch (error: any) {
+      throw new Error(`Sync failed: ${error.message}`);
+    }
+  }
+
+  // --- RAG Helper Methods ---
+
+  private async loadEmbeddingModel(): Promise<void> {
+    this.emitProgress('downloading', 0, 'Loading embedding model (all-MiniLM-L6-v2)...');
+    try {
+      const { pipeline } = await import('@huggingface/transformers');
+      this.localEmbeddingModel = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+      this.emitProgress('ready', 100, 'Embedding model loaded');
+    } catch (error: any) {
+      this.emitProgress('error', 0, `Embedding model failed: ${error.message}`);
+      throw error;
+    }
+  }
+
+  private async embedTextLocal(text: string): Promise<number[]> {
+    if (!this.localEmbeddingModel) throw new Error('Embedding model not loaded');
+    const result = await this.localEmbeddingModel(text, { pooling: 'mean', normalize: true });
+    // Handle different tensor output formats (v2 vs v3 of transformers)
+    if (result.data) return Array.from(result.data);
+    if (result.tolist) return result.tolist().flat();
+    if (Array.isArray(result)) return result.flat();
+    throw new Error('Unexpected embedding output format');
+  }
+
+  private cosineSimilarity(a: number[], b: number[]): number {
+    let dot = 0, normA = 0, normB = 0;
+    for (let i = 0; i < a.length; i++) {
+      dot += a[i] * b[i];
+      normA += a[i] * a[i];
+      normB += b[i] * b[i];
+    }
+    const denom = Math.sqrt(normA) * Math.sqrt(normB);
+    return denom === 0 ? 0 : dot / denom;
+  }
+
+  private chunkTextLocal(text: string, chunkSize: number = 512, overlap: number = 128): string[] {
+    if (!text || text.length === 0) return [];
+    if (overlap >= chunkSize) overlap = Math.floor(chunkSize * 0.25);
+    const chunks: string[] = [];
+    let start = 0;
+    while (start < text.length) {
+      let end = start + chunkSize;
+      if (end < text.length) {
+        const bp = Math.max(text.lastIndexOf('.', end), text.lastIndexOf('\n', end));
+        if (bp > start + chunkSize / 2) end = bp + 1;
+      }
+      const chunk = text.slice(start, end).trim();
+      if (chunk.length > 20) chunks.push(chunk);
+      start = end - overlap;
+      if (start >= text.length) break;
+    }
+    return chunks;
+  }
+
+  // ── Static OpenAI Compatible Factory ────────────────────────────────
+
+  static openaiCompatible(config: { apiKey: string; apiUrl?: string; fallback?: FallbackConfig }): OpenAICompatibleClient {
+    const instance = new SlyOS({
+      apiKey: config.apiKey,
+      apiUrl: config.apiUrl,
+      fallback: { ...config.fallback, provider: config.fallback?.provider || 'openai' } as FallbackConfig,
+    });
+
+    return {
+      chat: {
+        completions: {
+          async create(request: OpenAIChatCompletionRequest & { model: string }): Promise<OpenAIChatCompletionResponse> {
+            const { model, ...chatRequest } = request;
+            return instance.chatCompletion(model, chatRequest);
+          },
+        },
+      },
+    };
+  }
+}
+
+export default SlyOS;
+export type {
+  SlyOSConfig,
+  SlyOSConfigWithFallback,
+  GenerateOptions,
+  TranscribeOptions,
+  DeviceProfile,
+  ProgressEvent,
+  SlyEvent,
+  QuantizationLevel,
+  ModelCategory,
+  OpenAIMessage,
+  OpenAIChatCompletionRequest,
+  OpenAIChatCompletionResponse,
+  OpenAIChoice,
+  OpenAIUsage,
+  BedrockTextGenerationConfig,
+  BedrockInvokeRequest,
+  BedrockInvokeResponse,
+  BedrockResult,
+  FallbackConfig,
+  FallbackProvider,
+  OpenAICompatibleClient,
+  RAGOptions,
+  RAGChunk,
+  RAGResponse,
+  OfflineIndex,
+};