## Adaptive Control in RAG Systems Retrieval-Augmented Generation (RAG) systems integrate the retrieval of external information with text generation, but require careful control to meet specific service-level objectives (SLOs). A recent study focuses on this aspect, modeling per-query control as a discrete choice between different actions: retrieval depth, generation mode (guarded vs. auto), or query refusal. ## Objectives and Results The research uses an offline dataset constructed from SQuAD 2.0, evaluating accuracy, token cost, hallucination/refusal indicators, and an SLO-weighted reward. Two simple policy-learning objectives were evaluated: supervised classification of the per-state best action and a reward-weighted variant. The results show that a strong fixed baseline (low k, guarded prompting) performs competitively. Learned policies mainly provide additional cost savings under a quality-focused SLO, but can exhibit refusal collapse under a cheap SLO when refusal is heavily rewarded. ## Implications The study provides a reproducible case study of SLO-aware control for RAG pipelines, with an emphasis on failure modes and reporting conventions. The goal is not to propose a new retriever or language model, but rather to provide practical guidance for implementing more efficient and reliable RAG systems.