The Schema Guard: How I Catch LLM Output Changes Before They Break My n8n Pipelines
Описание и применение
Защита схемы предотвращает ошибки в рабочих процессах n8n, вызванные изменениями формата вывода LLM.
Задача
An LLM output drift caused a workflow to misroute emails to the low‑confidence bucket without any error signal.
Результат
The guard prevented 300 emails from being misrouted, restoring correct workflow behavior without manual intervention.
Рабочий процесс
- Parse LLM JSON
- Run Guard node to validate schema
- If guard passes, continue to downstream nodes; otherwise quarantine
Доказательства
My inbox triage workflow ran fine for six weeks. Then, on a Tuesday, it quietly started dumping every single email into the "low confidence" bucket -- three hundred of them, all waiting for human review.
The LLM had simply decided that morning to return confidence on a 0-100 scale instead of 0-1.
Right after the node that parses the model's JSON -- and before any node that reads a field from it -- sits a Guard node.
- Дата записи: 2026-10-15T02:40:58.217Z