Summary

• Implement load balancing between multiple LLM deployments (OpenAI, Google Palm/Gemini, Cohere)
• Routing strategies: least-tokens (default), simple-shuffle, latency-based
• Timeout/retry with exponential backoff via axios interceptors
• Streaming support for all providers
• Token usage tracking with cost calculation
• Sentry and Posthog logging callbacks
• JSONNet configuration support
• Mock servers for testing

Features Implemented

1. Load Balancing: Picks deployment below rate-limit with least tokens used
2. Reliability: Timeouts, retries, exponential backoff
3. Streaming: Full streaming support
4. Token Usage: Tracks prompt/completion/total tokens and cost
5. Logging: Sentry + Posthog callbacks

Tests

• 8 E2E tests passing covering all features
• Mock servers for OpenAI, Gemini, Cohere

Usage Example

import { Router } from "@arakoodev/edgechains.js/ai";

const router = new Router({
  modelList: [
    { modelName: "gpt-3.5-turbo", provider: "openai", apiKey: "sk-xxx", rpm: 3000, tpm: 90000 },
    { modelName: "gpt-3.5-turbo", provider: "openai", apiKey: "sk-yyy", rpm: 3000, tpm: 90000 },
  ],
  routingStrategy: "least-tokens",
  numRetries: 3,
  timeout: 30000,
});

const response = await router.completion({
  model: "gpt-3.5-turbo",
  messages: [{ role: "user", content: "Hello!" }],
});

/claim #286 closes #286