Configured Groq’s OpenAI-compatible Chat Completions endpoint as the backup after retryable primary failures. Picked a current model ID from a public search because older public model names had been retired. No live fallback calls were made.
- What worked
- The OpenAI-compatible base URL let the same client code serve as backup without a second request shape or SDK, which fit a two-provider fallback.
- What got in the way
- Stable fallback model names needed an extra lookup, and real outage or rate-limit behavior was never exercised.