Recipe: Robust Error Handling

Problem: AI APIs fail often (Rate limits, Overloaded, Content Filters). Solution: Implement Retries, Fallbacks, and graceful UI degradation.

1. Exponential Backoff (Server-Side)

If OpenAI says "429 Too Many Requests", wait and try again.

async function callWithRetry(fn, retries = 3) {
  for (let i = 0; i < retries; i++) {
    try {
      return await fn();
    } catch (error) {
      if (error.status === 429 && i < retries - 1) {
        const delay = Math.pow(2, i) * 1000; // 1s, 2s, 4s
        await new Promise(r => setTimeout(r, delay));
        continue;
      }
      throw error;
    }
  }
}

2. Model Fallback

If GPT-4o is down, fall back to a cheaper/more available model.

try {
  return await callOpenAI('gpt-4o');
} catch (e) {
  console.warn('GPT-4o failed, falling back to 3.5');
  return await callOpenAI('gpt-3.5-turbo');
}

3. UI Error Boundaries (Client-Side)

Use the onError callback in Vercel AI SDK.

const { messages, reload } = useChat({
  onError: (error) => {
    toast.error("AI is overloaded. Retrying...");
    // Auto-retry after 2 seconds
    setTimeout(() => reload(), 2000);
  }
});

Common Error Codes

Code	Meaning	Action
401	Invalid Key	Check .env files.
429	Rate Limited	Retry with backoff. Upgrade tier.
500	Server Error	OpenAI is down. Check status page.
503	Overloaded	Retry immediately (often transient).