ruvector 0.2.21 → 0.2.23

package/dist/core/onnx-llm.d.ts

@@ -1,206 +0,0 @@
-/**
- * ONNX LLM Text Generation for RuVector
- *
- * Provides real local LLM inference using ONNX Runtime via transformers.js
- * Supports small models that run efficiently on CPU:
- * - SmolLM 135M - Smallest, fast (~135MB)
- * - SmolLM 360M - Better quality (~360MB)
- * - TinyLlama 1.1B - Best small model quality (~1GB quantized)
- * - Qwen2.5 0.5B - Good balance (~500MB)
- *
- * Features:
- * - Automatic model downloading and caching
- * - Quantized INT4/INT8 models for efficiency
- * - Streaming generation support
- * - Temperature, top-k, top-p sampling
- * - KV cache for efficient multi-turn conversations
- */
-export interface OnnxLLMConfig {
-    /** Model ID (default: 'Xenova/smollm-135m-instruct') */
-    modelId?: string;
-    /** Cache directory for models */
-    cacheDir?: string;
-    /** Use quantized model (default: true) */
-    quantized?: boolean;
-    /** Device: 'cpu' | 'webgpu' (default: 'cpu') */
-    device?: 'cpu' | 'webgpu';
-    /** Maximum context length */
-    maxLength?: number;
-}
-export interface GenerationConfig {
-    /** Maximum new tokens to generate (default: 128) */
-    maxNewTokens?: number;
-    /** Temperature for sampling (default: 0.7) */
-    temperature?: number;
-    /** Top-p nucleus sampling (default: 0.9) */
-    topP?: number;
-    /** Top-k sampling (default: 50) */
-    topK?: number;
-    /** Repetition penalty (default: 1.1) */
-    repetitionPenalty?: number;
-    /** Stop sequences */
-    stopSequences?: string[];
-    /** System prompt for chat models */
-    systemPrompt?: string;
-    /** Enable streaming (callback for each token) */
-    onToken?: (token: string) => void;
-}
-export interface GenerationResult {
-    /** Generated text */
-    text: string;
-    /** Number of tokens generated */
-    tokensGenerated: number;
-    /** Time taken in milliseconds */
-    timeMs: number;
-    /** Tokens per second */
-    tokensPerSecond: number;
-    /** Model used */
-    model: string;
-    /** Whether model was loaded from cache */
-    cached: boolean;
-}
-export declare const AVAILABLE_MODELS: {
-    readonly 'trm-tinystories': {
-        readonly id: "Xenova/TinyStories-33M";
-        readonly name: "TinyStories 33M (TRM)";
-        readonly size: "~65MB";
-        readonly description: "Ultra-tiny model for stories and basic generation";
-        readonly contextLength: 512;
-    };
-    readonly 'trm-gpt2-tiny': {
-        readonly id: "Xenova/gpt2";
-        readonly name: "GPT-2 124M (TRM)";
-        readonly size: "~250MB";
-        readonly description: "Classic GPT-2 tiny for general text";
-        readonly contextLength: 1024;
-    };
-    readonly 'trm-distilgpt2': {
-        readonly id: "Xenova/distilgpt2";
-        readonly name: "DistilGPT-2 (TRM)";
-        readonly size: "~82MB";
-        readonly description: "Distilled GPT-2, fastest general model";
-        readonly contextLength: 1024;
-    };
-    readonly 'smollm-135m': {
-        readonly id: "HuggingFaceTB/SmolLM-135M-Instruct";
-        readonly name: "SmolLM 135M";
-        readonly size: "~135MB";
-        readonly description: "Smallest instruct model, very fast";
-        readonly contextLength: 2048;
-    };
-    readonly 'smollm-360m': {
-        readonly id: "HuggingFaceTB/SmolLM-360M-Instruct";
-        readonly name: "SmolLM 360M";
-        readonly size: "~360MB";
-        readonly description: "Small model, fast, better quality";
-        readonly contextLength: 2048;
-    };
-    readonly 'smollm2-135m': {
-        readonly id: "HuggingFaceTB/SmolLM2-135M-Instruct";
-        readonly name: "SmolLM2 135M";
-        readonly size: "~135MB";
-        readonly description: "Latest SmolLM v2, improved capabilities";
-        readonly contextLength: 2048;
-    };
-    readonly 'smollm2-360m': {
-        readonly id: "HuggingFaceTB/SmolLM2-360M-Instruct";
-        readonly name: "SmolLM2 360M";
-        readonly size: "~360MB";
-        readonly description: "Latest SmolLM v2, better quality";
-        readonly contextLength: 2048;
-    };
-    readonly 'qwen2.5-0.5b': {
-        readonly id: "Qwen/Qwen2.5-0.5B-Instruct";
-        readonly name: "Qwen2.5 0.5B";
-        readonly size: "~300MB quantized";
-        readonly description: "Good balance of speed and quality, multilingual";
-        readonly contextLength: 4096;
-    };
-    readonly tinyllama: {
-        readonly id: "TinyLlama/TinyLlama-1.1B-Chat-v1.0";
-        readonly name: "TinyLlama 1.1B";
-        readonly size: "~600MB quantized";
-        readonly description: "Best small model quality, slower";
-        readonly contextLength: 2048;
-    };
-    readonly 'codegemma-2b': {
-        readonly id: "google/codegemma-2b";
-        readonly name: "CodeGemma 2B";
-        readonly size: "~1GB quantized";
-        readonly description: "Code generation specialist";
-        readonly contextLength: 8192;
-    };
-    readonly 'deepseek-coder-1.3b': {
-        readonly id: "deepseek-ai/deepseek-coder-1.3b-instruct";
-        readonly name: "DeepSeek Coder 1.3B";
-        readonly size: "~700MB quantized";
-        readonly description: "Excellent for code tasks";
-        readonly contextLength: 4096;
-    };
-    readonly 'phi-2': {
-        readonly id: "microsoft/phi-2";
-        readonly name: "Phi-2 2.7B";
-        readonly size: "~1.5GB quantized";
-        readonly description: "High quality small model";
-        readonly contextLength: 2048;
-    };
-    readonly 'phi-3-mini': {
-        readonly id: "microsoft/Phi-3-mini-4k-instruct";
-        readonly name: "Phi-3 Mini";
-        readonly size: "~2GB quantized";
-        readonly description: "Best quality tiny model";
-        readonly contextLength: 4096;
-    };
-};
-export type ModelKey = keyof typeof AVAILABLE_MODELS;
-/**
- * Check if transformers.js is available
- */
-export declare function isTransformersAvailable(): Promise<boolean>;
-/**
- * Initialize the ONNX LLM with specified model
- */
-export declare function initOnnxLLM(config?: OnnxLLMConfig): Promise<boolean>;
-/**
- * Generate text using ONNX LLM
- */
-export declare function generate(prompt: string, config?: GenerationConfig): Promise<GenerationResult>;
-/**
- * Generate with streaming (token by token)
- */
-export declare function generateStream(prompt: string, config?: GenerationConfig): Promise<AsyncGenerator<string, GenerationResult, undefined>>;
-/**
- * Chat completion with conversation history
- */
-export declare function chat(messages: Array<{
-    role: 'system' | 'user' | 'assistant';
-    content: string;
-}>, config?: GenerationConfig): Promise<GenerationResult>;
-/**
- * Get model information
- */
-export declare function getModelInfo(): {
-    model: string | null;
-    ready: boolean;
-    availableModels: typeof AVAILABLE_MODELS;
-};
-/**
- * Unload the current model to free memory
- */
-export declare function unload(): Promise<void>;
-export declare class OnnxLLM {
-    private config;
-    private initialized;
-    constructor(config?: OnnxLLMConfig);
-    init(): Promise<boolean>;
-    generate(prompt: string, config?: GenerationConfig): Promise<GenerationResult>;
-    chat(messages: Array<{
-        role: 'system' | 'user' | 'assistant';
-        content: string;
-    }>, config?: GenerationConfig): Promise<GenerationResult>;
-    unload(): Promise<void>;
-    get ready(): boolean;
-    get model(): string | null;
-}
-export default OnnxLLM;
-//# sourceMappingURL=onnx-llm.d.ts.map

package/dist/core/onnx-llm.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"onnx-llm.d.ts","sourceRoot":"","sources":["../../src/core/onnx-llm.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAaH,MAAM,WAAW,aAAa;IAC5B,wDAAwD;IACxD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,iCAAiC;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,0CAA0C;IAC1C,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,gDAAgD;IAChD,MAAM,CAAC,EAAE,KAAK,GAAG,QAAQ,CAAC;IAC1B,6BAA6B;IAC7B,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,gBAAgB;IAC/B,oDAAoD;IACpD,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,8CAA8C;IAC9C,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,4CAA4C;IAC5C,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,mCAAmC;IACnC,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,wCAAwC;IACxC,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,qBAAqB;IACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,oCAAoC;IACpC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,iDAAiD;IACjD,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;CACnC;AAED,MAAM,WAAW,gBAAgB;IAC/B,qBAAqB;IACrB,IAAI,EAAE,MAAM,CAAC;IACb,iCAAiC;IACjC,eAAe,EAAE,MAAM,CAAC;IACxB,iCAAiC;IACjC,MAAM,EAAE,MAAM,CAAC;IACf,wBAAwB;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,iBAAiB;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,0CAA0C;IAC1C,MAAM,EAAE,OAAO,CAAC;CACjB;AAMD,eAAO,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAmHnB,CAAC;AAEX,MAAM,MAAM,QAAQ,GAAG,MAAM,OAAO,gBAAgB,CAAC;AAYrD;;GAEG;AACH,wBAAsB,uBAAuB,IAAI,OAAO,CAAC,OAAO,CAAC,CAOhE;AAED;;GAEG;AACH,wBAAsB,WAAW,CAAC,MAAM,GAAE,aAAkB,GAAG,OAAO,CAAC,OAAO,CAAC,CAqD9E;AAED;;GAEG;AACH,wBAAsB,QAAQ,CAC5B,MAAM,EAAE,MAAM,EACd,MAAM,GAAE,gBAAqB,GAC5B,OAAO,CAAC,gBAAgB,CAAC,CA0C3B;AAED;;GAEG;AACH,wBAAsB,cAAc,CAClC,MAAM,EAAE,MAAM,EACd,MAAM,GAAE,gBAAqB,GAC5B,OAAO,CAAC,cAAc,CAAC,MAAM,EAAE,gBAAgB,EAAE,SAAS,CAAC,CAAC,CA0D9D;AAED;;GAEG;AACH,wBAAsB,IAAI,CACxB,QAAQ,EAAE,KAAK,CAAC;IAAE,IAAI,EAAE,QAAQ,GAAG,MAAM,GAAG,WAAW,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAAC,EAC3E,MAAM,GAAE,gBAAqB,GAC5B,OAAO,CAAC,gBAAgB,CAAC,CAsB3B;AAED;;GAEG;AACH,wBAAgB,YAAY,IAAI;IAC9B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACrB,KAAK,EAAE,OAAO,CAAC;IACf,eAAe,EAAE,OAAO,gBAAgB,CAAC;CAC1C,CAMA;AAED;;GAEG;AACH,wBAAsB,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC,CAQ5C;AAMD,qBAAa,OAAO;IAClB,OAAO,CAAC,MAAM,CAAgB;IAC9B,OAAO,CAAC,WAAW,CAAS;gBAEhB,MAAM,GAAE,aAAkB;IAIhC,IAAI,IAAI,OAAO,CAAC,OAAO,CAAC;IAMxB,QAAQ,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAK9E,IAAI,CACR,QAAQ,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,QAAQ,GAAG,MAAM,GAAG,WAAW,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,EAC3E,MAAM,CAAC,EAAE,gBAAgB,GACxB,OAAO,CAAC,gBAAgB,CAAC;IAKtB,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAK7B,IAAI,KAAK,IAAI,OAAO,CAEnB;IAED,IAAI,KAAK,IAAI,MAAM,GAAG,IAAI,CAEzB;CACF;AAED,eAAe,OAAO,CAAC"}

package/dist/core/onnx-llm.js

@@ -1,430 +0,0 @@
-"use strict";
-/**
- * ONNX LLM Text Generation for RuVector
- *
- * Provides real local LLM inference using ONNX Runtime via transformers.js
- * Supports small models that run efficiently on CPU:
- * - SmolLM 135M - Smallest, fast (~135MB)
- * - SmolLM 360M - Better quality (~360MB)
- * - TinyLlama 1.1B - Best small model quality (~1GB quantized)
- * - Qwen2.5 0.5B - Good balance (~500MB)
- *
- * Features:
- * - Automatic model downloading and caching
- * - Quantized INT4/INT8 models for efficiency
- * - Streaming generation support
- * - Temperature, top-k, top-p sampling
- * - KV cache for efficient multi-turn conversations
- */
-var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
-    if (k2 === undefined) k2 = k;
-    var desc = Object.getOwnPropertyDescriptor(m, k);
-    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
-      desc = { enumerable: true, get: function() { return m[k]; } };
-    }
-    Object.defineProperty(o, k2, desc);
-}) : (function(o, m, k, k2) {
-    if (k2 === undefined) k2 = k;
-    o[k2] = m[k];
-}));
-var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
-    Object.defineProperty(o, "default", { enumerable: true, value: v });
-}) : function(o, v) {
-    o["default"] = v;
-});
-var __importStar = (this && this.__importStar) || (function () {
-    var ownKeys = function(o) {
-        ownKeys = Object.getOwnPropertyNames || function (o) {
-            var ar = [];
-            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
-            return ar;
-        };
-        return ownKeys(o);
-    };
-    return function (mod) {
-        if (mod && mod.__esModule) return mod;
-        var result = {};
-        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
-        __setModuleDefault(result, mod);
-        return result;
-    };
-})();
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.OnnxLLM = exports.AVAILABLE_MODELS = void 0;
-exports.isTransformersAvailable = isTransformersAvailable;
-exports.initOnnxLLM = initOnnxLLM;
-exports.generate = generate;
-exports.generateStream = generateStream;
-exports.chat = chat;
-exports.getModelInfo = getModelInfo;
-exports.unload = unload;
-const path = __importStar(require("path"));
-const fs = __importStar(require("fs"));
-// Force native dynamic import (avoids TypeScript transpiling to require)
-// eslint-disable-next-line @typescript-eslint/no-implied-eval
-const dynamicImport = new Function('specifier', 'return import(specifier)');
-// ============================================================================
-// Available Models
-// ============================================================================
-exports.AVAILABLE_MODELS = {
-    // =========================================================================
-    // TRM - Tiny Random Models (smallest, fastest)
-    // =========================================================================
-    'trm-tinystories': {
-        id: 'Xenova/TinyStories-33M',
-        name: 'TinyStories 33M (TRM)',
-        size: '~65MB',
-        description: 'Ultra-tiny model for stories and basic generation',
-        contextLength: 512,
-    },
-    'trm-gpt2-tiny': {
-        id: 'Xenova/gpt2',
-        name: 'GPT-2 124M (TRM)',
-        size: '~250MB',
-        description: 'Classic GPT-2 tiny for general text',
-        contextLength: 1024,
-    },
-    'trm-distilgpt2': {
-        id: 'Xenova/distilgpt2',
-        name: 'DistilGPT-2 (TRM)',
-        size: '~82MB',
-        description: 'Distilled GPT-2, fastest general model',
-        contextLength: 1024,
-    },
-    // =========================================================================
-    // SmolLM - Smallest production-ready models
-    // =========================================================================
-    'smollm-135m': {
-        id: 'HuggingFaceTB/SmolLM-135M-Instruct',
-        name: 'SmolLM 135M',
-        size: '~135MB',
-        description: 'Smallest instruct model, very fast',
-        contextLength: 2048,
-    },
-    'smollm-360m': {
-        id: 'HuggingFaceTB/SmolLM-360M-Instruct',
-        name: 'SmolLM 360M',
-        size: '~360MB',
-        description: 'Small model, fast, better quality',
-        contextLength: 2048,
-    },
-    'smollm2-135m': {
-        id: 'HuggingFaceTB/SmolLM2-135M-Instruct',
-        name: 'SmolLM2 135M',
-        size: '~135MB',
-        description: 'Latest SmolLM v2, improved capabilities',
-        contextLength: 2048,
-    },
-    'smollm2-360m': {
-        id: 'HuggingFaceTB/SmolLM2-360M-Instruct',
-        name: 'SmolLM2 360M',
-        size: '~360MB',
-        description: 'Latest SmolLM v2, better quality',
-        contextLength: 2048,
-    },
-    // =========================================================================
-    // Qwen - Chinese/English bilingual models
-    // =========================================================================
-    'qwen2.5-0.5b': {
-        id: 'Qwen/Qwen2.5-0.5B-Instruct',
-        name: 'Qwen2.5 0.5B',
-        size: '~300MB quantized',
-        description: 'Good balance of speed and quality, multilingual',
-        contextLength: 4096,
-    },
-    // =========================================================================
-    // TinyLlama - Llama architecture in tiny form
-    // =========================================================================
-    'tinyllama': {
-        id: 'TinyLlama/TinyLlama-1.1B-Chat-v1.0',
-        name: 'TinyLlama 1.1B',
-        size: '~600MB quantized',
-        description: 'Best small model quality, slower',
-        contextLength: 2048,
-    },
-    // =========================================================================
-    // Code-specialized models
-    // =========================================================================
-    'codegemma-2b': {
-        id: 'google/codegemma-2b',
-        name: 'CodeGemma 2B',
-        size: '~1GB quantized',
-        description: 'Code generation specialist',
-        contextLength: 8192,
-    },
-    'deepseek-coder-1.3b': {
-        id: 'deepseek-ai/deepseek-coder-1.3b-instruct',
-        name: 'DeepSeek Coder 1.3B',
-        size: '~700MB quantized',
-        description: 'Excellent for code tasks',
-        contextLength: 4096,
-    },
-    // =========================================================================
-    // Phi models - Microsoft's tiny powerhouses
-    // =========================================================================
-    'phi-2': {
-        id: 'microsoft/phi-2',
-        name: 'Phi-2 2.7B',
-        size: '~1.5GB quantized',
-        description: 'High quality small model',
-        contextLength: 2048,
-    },
-    'phi-3-mini': {
-        id: 'microsoft/Phi-3-mini-4k-instruct',
-        name: 'Phi-3 Mini',
-        size: '~2GB quantized',
-        description: 'Best quality tiny model',
-        contextLength: 4096,
-    },
-};
-// ============================================================================
-// ONNX LLM Generator
-// ============================================================================
-let pipeline = null;
-let transformers = null;
-let loadedModel = null;
-let loadPromise = null;
-let loadError = null;
-/**
- * Check if transformers.js is available
- */
-async function isTransformersAvailable() {
-    try {
-        await dynamicImport('@xenova/transformers');
-        return true;
-    }
-    catch {
-        return false;
-    }
-}
-/**
- * Initialize the ONNX LLM with specified model
- */
-async function initOnnxLLM(config = {}) {
-    if (pipeline && loadedModel === config.modelId) {
-        return true;
-    }
-    if (loadError)
-        throw loadError;
-    if (loadPromise) {
-        await loadPromise;
-        return pipeline !== null;
-    }
-    const modelId = config.modelId || 'HuggingFaceTB/SmolLM-135M-Instruct';
-    loadPromise = (async () => {
-        try {
-            console.error(`Loading ONNX LLM: ${modelId}...`);
-            // Import transformers.js
-            transformers = await dynamicImport('@xenova/transformers');
-            const { pipeline: createPipeline, env } = transformers;
-            // Configure cache directory
-            if (config.cacheDir) {
-                env.cacheDir = config.cacheDir;
-            }
-            else {
-                env.cacheDir = path.join(process.env.HOME || '/tmp', '.ruvector', 'models', 'onnx-llm');
-            }
-            // Ensure cache directory exists
-            if (!fs.existsSync(env.cacheDir)) {
-                fs.mkdirSync(env.cacheDir, { recursive: true });
-            }
-            // Disable remote model fetching warnings
-            env.allowRemoteModels = true;
-            env.allowLocalModels = true;
-            // Create text generation pipeline
-            console.error(`Downloading model (first run may take a while)...`);
-            pipeline = await createPipeline('text-generation', modelId, {
-                quantized: config.quantized !== false,
-                device: config.device || 'cpu',
-            });
-            loadedModel = modelId;
-            console.error(`ONNX LLM ready: ${modelId}`);
-        }
-        catch (e) {
-            loadError = new Error(`Failed to initialize ONNX LLM: ${e.message}`);
-            throw loadError;
-        }
-    })();
-    await loadPromise;
-    return pipeline !== null;
-}
-/**
- * Generate text using ONNX LLM
- */
-async function generate(prompt, config = {}) {
-    if (!pipeline) {
-        await initOnnxLLM();
-    }
-    if (!pipeline) {
-        throw new Error('ONNX LLM not initialized');
-    }
-    const start = performance.now();
-    // Build the input text (apply chat template if needed)
-    let inputText = prompt;
-    if (config.systemPrompt) {
-        // Apply simple chat format
-        inputText = `<|system|>\n${config.systemPrompt}<|end|>\n<|user|>\n${prompt}<|end|>\n<|assistant|>\n`;
-    }
-    // Generate
-    const outputs = await pipeline(inputText, {
-        max_new_tokens: config.maxNewTokens || 128,
-        temperature: config.temperature || 0.7,
-        top_p: config.topP || 0.9,
-        top_k: config.topK || 50,
-        repetition_penalty: config.repetitionPenalty || 1.1,
-        do_sample: (config.temperature || 0.7) > 0,
-        return_full_text: false,
-    });
-    const timeMs = performance.now() - start;
-    const generatedText = outputs[0]?.generated_text || '';
-    // Estimate tokens (rough approximation)
-    const tokensGenerated = Math.ceil(generatedText.split(/\s+/).length * 1.3);
-    return {
-        text: generatedText.trim(),
-        tokensGenerated,
-        timeMs,
-        tokensPerSecond: tokensGenerated / (timeMs / 1000),
-        model: loadedModel || 'unknown',
-        cached: true,
-    };
-}
-/**
- * Generate with streaming (token by token)
- */
-async function generateStream(prompt, config = {}) {
-    if (!pipeline) {
-        await initOnnxLLM();
-    }
-    if (!pipeline) {
-        throw new Error('ONNX LLM not initialized');
-    }
-    const start = performance.now();
-    let fullText = '';
-    let tokenCount = 0;
-    // Build input text
-    let inputText = prompt;
-    if (config.systemPrompt) {
-        inputText = `<|system|>\n${config.systemPrompt}<|end|>\n<|user|>\n${prompt}<|end|>\n<|assistant|>\n`;
-    }
-    // Create streamer
-    const { TextStreamer } = transformers;
-    const streamer = new TextStreamer(pipeline.tokenizer, {
-        skip_prompt: true,
-        callback_function: (text) => {
-            fullText += text;
-            tokenCount++;
-            if (config.onToken) {
-                config.onToken(text);
-            }
-        },
-    });
-    // Generate with streamer
-    await pipeline(inputText, {
-        max_new_tokens: config.maxNewTokens || 128,
-        temperature: config.temperature || 0.7,
-        top_p: config.topP || 0.9,
-        top_k: config.topK || 50,
-        repetition_penalty: config.repetitionPenalty || 1.1,
-        do_sample: (config.temperature || 0.7) > 0,
-        streamer,
-    });
-    const timeMs = performance.now() - start;
-    // Return generator that yields the collected text
-    async function* generator() {
-        yield fullText;
-        return {
-            text: fullText.trim(),
-            tokensGenerated: tokenCount,
-            timeMs,
-            tokensPerSecond: tokenCount / (timeMs / 1000),
-            model: loadedModel || 'unknown',
-            cached: true,
-        };
-    }
-    return generator();
-}
-/**
- * Chat completion with conversation history
- */
-async function chat(messages, config = {}) {
-    if (!pipeline) {
-        await initOnnxLLM();
-    }
-    if (!pipeline) {
-        throw new Error('ONNX LLM not initialized');
-    }
-    // Build conversation text from messages
-    let conversationText = '';
-    for (const msg of messages) {
-        if (msg.role === 'system') {
-            conversationText += `<|system|>\n${msg.content}<|end|>\n`;
-        }
-        else if (msg.role === 'user') {
-            conversationText += `<|user|>\n${msg.content}<|end|>\n`;
-        }
-        else if (msg.role === 'assistant') {
-            conversationText += `<|assistant|>\n${msg.content}<|end|>\n`;
-        }
-    }
-    conversationText += '<|assistant|>\n';
-    return generate(conversationText, { ...config, systemPrompt: undefined });
-}
-/**
- * Get model information
- */
-function getModelInfo() {
-    return {
-        model: loadedModel,
-        ready: pipeline !== null,
-        availableModels: exports.AVAILABLE_MODELS,
-    };
-}
-/**
- * Unload the current model to free memory
- */
-async function unload() {
-    if (pipeline) {
-        // Note: transformers.js doesn't have explicit dispose, but we can null the reference
-        pipeline = null;
-        loadedModel = null;
-        loadPromise = null;
-        loadError = null;
-    }
-}
-// ============================================================================
-// Class wrapper for OOP usage
-// ============================================================================
-class OnnxLLM {
-    constructor(config = {}) {
-        this.initialized = false;
-        this.config = config;
-    }
-    async init() {
-        if (this.initialized)
-            return true;
-        this.initialized = await initOnnxLLM(this.config);
-        return this.initialized;
-    }
-    async generate(prompt, config) {
-        if (!this.initialized)
-            await this.init();
-        return generate(prompt, config);
-    }
-    async chat(messages, config) {
-        if (!this.initialized)
-            await this.init();
-        return chat(messages, config);
-    }
-    async unload() {
-        await unload();
-        this.initialized = false;
-    }
-    get ready() {
-        return this.initialized;
-    }
-    get model() {
-        return loadedModel;
-    }
-}
-exports.OnnxLLM = OnnxLLM;
-exports.default = OnnxLLM;