Describe an exterior lighting fault in plain language. The system grounds every diagnosis in retrieved evidence and declines when that evidence isn't enough - it won't guess.
Agentic RAG over 89 real NHTSA complaint reports and a 69-entry synthetic DTC catalogue - hybrid BM25 + vector retrieval, cross-encoder reranked, with citation enforcement checked in code rather than asked of the model. Evaluated on 48 hand-written scenarios (40 with a known answer, 8 with none, to make refusal measurable):