What’s a good fallback strategy when a prompt fails?
On failure, retry with a shorter prompt, then switch models or routes (e.g., retrieval on vs. off). Use deterministic validators to decide when to retry. If still failing, escalate to human with a concise summary of what the model tried. Keep logs so you can fix root causes later.
In practice
Retries should change something meaningful—shorter context or a stricter schema—not just rerun the same prompt.
Where to start
- [01]Define clear retry conditions.
- [02]Implement a second, simpler prompt path.
- [03]Add model routing for stubborn cases.
- [04]Capture a failure report for human review.
fallback promptsprompt retrieserror handling LLM