Don't bolt agentic enhancements onto RAG uniformly — query decomposition and reflection help in some domains and hurt in others

arXiv:2606.05658·high signal

Agent-Orchestrated Adaptive RAG (arXiv:2606.05658, submitted June 4, 2026) finds that 'agentic enhancements are not universally beneficial': dynamic query decomposition gained +0.17 MRR on a structured DevOps benchmark but degraded ranking precision on a multi-hop benchmark, and the self-reflective loop improved citation accuracy only at substantial latency cost. The takeaway is adaptive, cost-aware orchestration over uniformly aggressive reasoning pipelines. Actionable: gate decomposition/reflection per query type and domain, and measure the latency tax before turning them on everywhere.