Enterprise operations teams confront a structural failure mode in incident response: the same error signature carries different root causes in different environments, yet the knowledge required to disambiguate them is scattered across systems optimized for tracking rather than retrieval. This paper introduces RAGOps, an architecture that formalizes operational knowledge retrieval as a metadata-constrained, typed, non-stationary ranking problem: an environment predicate is applied before similarity, the corpus is a typed union of five knowledge forms (runbooks, incident cases, curated terminal runs, validated operational rules, and deduplicated error signatures), and ranking evolves with feedback without model retraining. The architecture combines curation-first ingestion with a secrets-rejecting gate ahead of embedding, fully local inference, single-round-trip multi-corpus first-stage retrieval, and agent integration via the Model Context Protocol under a read-auto/write-approved permission asymmetry, in two interoperable deployment profiles (PostgreSQL/pgvector server; sqlite-vec embedded). On an open, deterministic synthetic benchmark (342 items, 160 graded queries), the environment predicate is the dominant quality lever regardless of retrieval method - raising Recall@10 from 0.36 to 0.85 for BM25 and 0.19 to 0.82 for dense retrieval while eliminating wrong-environment results that otherwise contaminate 66% of top-5 slots. Filtered BM25 is the strongest single configuration on lexically-anchored operational text (MRR@10 0.93), and a MiniLM-family cross-encoder fails to exploit superior lexical candidates when given them - evidence that measured domain-token embedding drift propagates to rerankers. Deployment feasibility is reported from two anonymized enterprise contexts. The benchmark, runner, and results are released for exact reproduction at https://github.com/sunilrreddy/ragops-bench. An extended technical report is included as a supplementary file.
Revuri et al. (Sat,) studied this question.