Large-scale document tagging systems face critical deployment challenges: uncontrolled label generation leads to computational waste and poor user experience. While existing hierarchical multi-label methods achieve reasonable accuracy, they lack precise control over output characteristics, such as label count, hierarchy depth, and computational efficiency required by real-world applications. This paper introduces Level-Guided Token Filtering (LGTF), a computationally efficient framework that enables precise control over label count, hierarchy depth, and generation speed. Unlike traditional attention masking approaches, LGTF directly filters irrelevant tokens at the vocabulary level prior to probability computation, reducing computational complexity while maintaining output accuracy. Experimental results demonstrate that LGTF achieves state-of-the-art (SOTA) performance on the Hierarchical Multi-Label Document Tagging (HMDT) task, significantly outperforming existing approaches in both accuracy and output control. This work highlights the potential of LGTF to enhance the effectiveness of controlled generation tasks in large-scale tagging systems.