Semantic Routing and Authority-Weighted Retrieval for Technical Infrastructure RAG: A Phase 4 Companion Study

The base study established that nomic-embed-text outperforms larger dense and hybrid embedding models on a technical infrastructure RAG corpus, and named three Phase 4 extensions as future work: a semantic router using a lightweight classifier language model, an authority-weighted corpus prioritizing live device state over static design documentation, and a full-corpus re-benchmark against the original 30-question evaluation set. This companion study reports the results of implementing and empirically testing all three. The proposed 0.6B-parameter router classifier measured 33% four-way classification accuracy — far below the near-perfect classificationthe base study's projected 88% combined accuracy assumed — and was replaced with a deterministic rule-based classifier that measured 100% on the same evaluation set. The resulting router did not outperform a single well-chosen model and retrieval backend on either the original benchmark or a new 15-question set targeting freshly-ingested content. Authority weighting produced a consistent, twice-replicated improvement on stale-versus-fresh factual disambiguation, but a corroborated failure mode on multi-step reasoning questions. Neither of the base study's two named authority-weighting target questions was resolved by the proposed mechanism itself. The live-state pull separately surfaced and — under one specific retrieval configuration — resolved a corpus defect that had persisted, undetected, through three prior research phases. A controlled measurement found generation-time sampling variance of up to 6 combined percentage points, a larger noise band than previously reported.