RAG that reads your old postmortems nails root causes 87% of the time
35–55% of incident time is spent hunting runbooks and old tickets, not fixing anything. This team built a five-layer RAG system grounded in ServiceNow, Confluence, PagerDuty and GitHub knowledge, and it hit 87.3% root-cause accuracy across 2,400 scenarios while cutting mean diagnosis time 59%. The lesson: most incident knowledge lives in unstructured text, not metrics, so standalone LLMs improvise badly without retrieval.