Vector Data Management Systems (VDMS) play a vital role in large-scale information retrieval by providing tunable parameters to optimize similarity search. However, performance tuning remains challenging due to inherent structural and operational constraints. Despite algorithmic advances, existing methods still fall short of real-world performance requirements. This paper proposes a novel tuning framework for vector databases that enhances management efficiency through a Bayesian-optimized window adaptive discarding mechanism. We present an approach where Bayesian optimization guides the window discarding strategy, enabling adaptive retention of high-value configurations and intelligent elimination of obsolete data under varying workloads. The proposed index recommendation system leverages probabilistic modeling to dynamically adjust window size and discarding criteria based on multi-objective feedback. Experimental results demonstrate that this Bayesian-driven window adaptive method significantly improves optimization performance and practical utility compared to prior approaches.