Retrieval-Augmented Generation (RAG) is the standard approach for augmenting LLM knowledge, yet high operational complexity and personnel dependency remain persistent challenges. This study empirically compares RAG and MCP (Model Context Protocol)-based pipelines on 105 Docker documentation queries using Qwen 2.5 7B and Llama 3.1 8B models. Tuned RAG achieves peak accuracy (75.0%), but the 45.6 percentage point gap from naive RAG (29.4%) reveals extreme configuration sensitivity. MCP parallel achieves 64.0% without parameter tuning, while reducing tuning parameters from 9 to 2, failure points from 3 to 1, and eliminating external dependencies beyond the LLM. On 14B models, MCP (73.1%) outscored tuned RAG (66.7%), suggesting a possible tipping point as model capabilities improve. We present quantitative evidence for the performance–operational cost trade-off between RAG and MCP, demonstrating that MCP-inspired architectures constitute a viable alternative where operational simplicity is prioritized.