Is Vercel AI SDK Down?

If Prismix shows Vercel AI SDK as "Operational" but you're having issues:

• Streaming response cuts off mid-generation — when deployed to Vercel Edge Runtime, streaming chunks can be lost if the edge function times out or the response is buffered by an intermediate proxy. Ensure your route uses export const runtime = 'edge' and that your Vercel plan supports the required execution duration. For long generations, consider switching to Node.js runtime with streaming enabled.
• Tool calling returning empty parameters — the AI SDK's tool() helper requires a Zod schema for parameters. If the schema is missing or the model doesn't support function calling, tool arguments arrive as empty objects. Verify the underlying provider model (e.g., gpt-4o, claude-3-5-sonnet) supports tool use and that your Zod schema is passed correctly.
• Provider SDK throwing 429 when the model is rate-limited — the Vercel AI SDK wraps provider errors but doesn't automatically retry 429s. If the underlying OpenAI or Anthropic API is rate-limited, the error surfaces immediately. Add exponential backoff manually or enable automatic retries via the provider options if supported by your SDK version.
• Next.js App Router server action not awaiting stream fully — calling streamText() inside a Next.js Server Action without properly returning the stream to the client results in a truncated response. Use the createStreamableValue helper or the Route Handler pattern instead of Server Actions for streaming responses.
• AI Gateway cache returning stale response — Vercel AI Gateway caches model responses by prompt hash. If your prompt includes dynamic content (timestamps, user IDs) but caching is enabled, you may receive a stale cached response. Disable caching for dynamic prompts by setting cache: false in the gateway options or add a unique cache-busting key per request.
• useChat hook state not syncing after page refresh — the useChat hook stores messages in component state by default. After a page refresh, history is lost unless you persist it via initialMessages loaded from a database or local storage. Implement a persistence layer and pass saved messages as initialMessages to the hook on mount.