The Retry Budget: Why My n8n LLM Workflows Retry Exactly Once, Then Page a Human

Описание и применение

В статье описывается, как установить бюджет повторных попыток для LLM‑воркфлоу в n8n, чтобы экономить токены и повышать надёжность, классифицируя ошибки на три группы и применяя разные стратегии повторов.

Задача

When an LLM completion fails validation, it is unclear whether to retry automatically or send it to human review, risking unnecessary cost and inefficiency.

Результат

Implemented a retry budget that reduces unnecessary prompt costs, improves reliability, and provides clear escalation to human review.

Рабочий процесс

• Classify error type and count attempts in a Code node
• If transient and attempts ≤3, set action to retry_backoff
• If repairable and first attempt, set action to retry_with_feedback
• If deterministic or repeated repairable failure, set action to review
• Route based on action field to retry or review queue

Доказательства

Every LLM failure in my workflows falls into one of three buckets, and each bucket gets a different response:

Источник

I logged 200 repairable failures over a month. First feedback-retry success rate: 71%. Second retry on the remaining 29%: 9% of those.

Источник