An inference application retries a failed request that triggers an external business action. Which approach best fits?

1. Review output-token limits and validate structured-response completeness.
2. Choose the smaller model if it satisfies all approved requirements.
3. Version and align the embedding pipeline for both retrieval sides.
4. Make the action idempotent and distinguish inference retries from action execution.