Auto Prompt Router to LLM Ever found yourself wondering which AI model to use for your specific task? Should you use GPT-4 for that complex coding problem, or would Claude be better for creative writing? This package takes the guesswork out of model selection.

Auto Prompt Router intelligently analyzes your prompts and automatically selects the best language model based on what you're trying to accomplish, your performance needs, and your budget.

Why This Exists With so many great AI models available (GPT-4, Claude, Gemini, Llama, etc.), choosing the right one for each task has become a real challenge. Each model has its strengths - some excel at coding, others at creative tasks. Some are lightning fast but less accurate, others are incredibly smart but slower and more expensive.

This router solves that problem by:

Understanding what type of task you're doing (coding, writing, analysis, etc.) Knowing the strengths and weaknesses of 80+ different models Considering your specific needs (speed, accuracy, cost) Making the optimal choice for you automatically Installation npm install auto-llm-selector Quick Example simply run npx tsx sample.ts after cloning and check the output

import { AutoPromptRouter, type RouterConfig, type PromptProperties, } from 'auto-llm-selector';

// Set up the router const router = new AutoPromptRouter({ OPEN_ROUTER_API_KEY: 'your-api-key', // Get one from openrouter.ai });

await router.initialize();

// Ask for help with a coding problem const result = await router.getModelRecommendation( 'Help me fix this Python bug - my function keeps returning None', { accuracy: 0.9, // I need this to be right cost: 0.5, // Moderate budget speed: 0.7, // Fairly quick response needed tokenLimit: 4000, // Not a huge response needed reasoning: true, // This requires some thinking } );

console.log(Best model: ${result.model}); console.log(Why: ${result.reason}); // Might output: "Selected GPT-4 for its excellent coding capabilities and reasoning skills" Real-World Examples When You Need Help Coding const codingHelp = await router.getModelRecommendation( 'Write a function to validate email addresses with regex', { accuracy: 0.95, // Code needs to be correct cost: 0.4, // Keep costs reasonable speed: 0.8, // Want a quick answer tokenLimit: 3000, reasoning: true, // Logic is important } ); // Likely picks: GPT-4 or Claude Sonnet (great at coding) Creative Writing Tasks const storyWriting = await router.getModelRecommendation( 'Write a short story about a robot learning to paint', { accuracy: 0.7, // Creativity over perfect grammar cost: 0.3, // Budget-conscious speed: 0.4, // Quality over speed tokenLimit: 8000, // Longer creative content reasoning: false, // Pure creativity } ); // Likely picks: Claude (excellent for creative tasks) or GPT-4 Quick Questions const quickChat = await router.getModelRecommendation( "What's the weather like in Tokyo right now?", { accuracy: 0.6, // Simple question cost: 0.1, // Keep it cheap speed: 0.9, // Want instant response tokenLimit: 500, reasoning: false, } ); // Likely picks: GPT-3.5-turbo or Claude Haiku (fast and cheap) How It Actually Works The magic happens in a few steps:

Understanding Your Prompt: The system reads your prompt and figures out what category it falls into - is this a coding question? Creative writing? Data analysis? It uses both AI embeddings and keyword matching to be really accurate.

Knowing the Models: We've tested and profiled over many different AI models. Each one gets scored on things like:

How good is it at coding vs creative tasks? How fast does it respond? How much does it cost? Can it handle complex reasoning? Making the Choice: An AI looks at your prompt, your requirements, and all the model profiles, then picks the best match. It even explains why it made that choice.

Learning Over Time: The system gets smarter as it sees more examples of what works well. // TODO: NOT YET FULLY IMPLEMENTED

What You Get Back Every recommendation includes:

The best model for your specific prompt and needs A clear explanation of why this model was chosen Confidence scores so you know how sure the system is Category classification showing how your prompt was understood Available Model Categories The router recognizes these types of tasks:

Coding & Development - Programming, debugging, code reviews, technical docs Creative Writing - Stories, poems, creative content, marketing copy Data Analysis - Research, comparisons, insights, business analysis Complex Reasoning - Logic puzzles, math problems, strategic thinking Conversation - Chat, Q&A, customer support, casual discussion General Knowledge - Facts, explanations, how-to guides Exports & Imports import { // Main class - does all the intelligent routing AutoPromptRouter,

// Core types for configuration and responses type RouterConfig, // Settings for API key, model selection, logging, analytics type PromptProperties, // Your requirements: accuracy, cost, speed, etc. type ModelSelection, // The recommendation result with model and reasoning type PromptCategory, // Classification result with type and confidence type AnalyticsConfig, // Privacy-first analytics configuration options

// Advanced types (optional, for custom integrations) type ModelProfile, // Complete model capability and characteristic data type ModelCapabilities, // Scores for coding, creative, analytical, etc. type ModelCharacteristics, // Speed, cost, accuracy tiers and provider info type ModelInfo, // Raw model data from OpenRouter API

// Enums PromptType, // Available categories: coding, creative, analytical, etc. } from 'auto-llm-selector'; Most users only need: AutoPromptRouter, RouterConfig, PromptProperties, and ModelSelection.

Getting Started Get an API Key: Sign up at OpenRouter.ai - they provide access to all the major AI models through one API Install the package: npm install auto-llm-selector Try the examples above to see how it works Check out the full API docs in the /docs folder Configuration Options const config: RouterConfig = { OPEN_ROUTER_API_KEY: 'your-key', // Required selectorModel: 'anthropic/claude-3-sonnet', // Optional: which model makes the selection enableLogging: true, // Optional: see detailed logs

// Optional: Privacy-first analytics (opt-in) analytics: { enabled: true, // Must be explicitly enabled collectPromptMetrics: true, // Prompt classification & model selection (COLLECTING RAW PROMPTS FOR TESTING PURPOSES, WILL BE REVERTED TO HASH SHORTLY) collectModelPerformance: true, // Model response times & success rates collectSemanticFeatures: true, // Classification confidence & embedding metrics collectSystemInfo: true, // Anonymized platform info (Node version, OS) debugMode: false, // Verbose analytics logging }, }; Requirements Node.js 16 or higher An OpenRouter API key (free tier available) Internet connection (for model data and AI classification) Development

Get started

npm install npm run build

Development mode

npm run dev

Code quality

npm run lint npm run typecheck npm run format Environment Variables Create a .env file and add the following:

OPEN_ROUTER_API_KEY=your-key NODE_ENV=development Troubleshooting TensorFlow Dependencies If you encounter TensorFlow-related errors:

npm uninstall @tensorflow/tfjs-node @tensorflow-models/universal-sentence-encoder npm install @tensorflow/tfjs-node @tensorflow-models/universal-sentence-encoder Performance & Analytics Classification Speed: ~100-300ms for hybrid semantic + keyword analysis Selection Speed: ~500-1500ms depending on selected LLM model Model Coverage: 80+ models from OpenAI, Anthropic, Google, Meta, and others Cache Efficiency: Model profiles cached for optimal performance Privacy-First Analytics (Optional) When enabled, the router collects anonymized usage data to improve model recommendations:

🔒 Privacy-First: Prompts are hashed (never stored as plain text) 📊 Zero Performance Impact: Async batch processing in background 🎯 Opt-In Only: Analytics disabled by default, must be explicitly enabled 📈 Improves Recommendations: Helps train better model selection algorithms What's collected:

Prompt categories and classification confidence (hashed prompts only) Model selection results and performance metrics System information (Node version, platform - anonymized) User fingerprint (based on system characteristics, no personal data) What's NOT collected:

Plain text prompts or responses IP addresses or personal identifiers API keys or sensitive configuration data Supported Models The router includes pre-curated profiles for popular models from:

OpenAI: GPT-4, GPT-4-Turbo, GPT-4o, GPT-3.5-Turbo Anthropic: Claude-3 (Opus, Sonnet, Haiku) Google: Gemini Pro, Gemini Flash Meta: Llama 3.1 (405B, 70B, 8B) Open Source: Mixtral, Wizard, and many more Apart from these everytime we initialize the library new models are fetched from OpenRouter API and cached for optimal performance.

Contributing Found a bug? Have an idea? Pull requests welcome! This is an open source project and we'd love your help making it better.

Keywords Auto Prompt Router To LLM,