Training multimodal models in deep learning often requires a large amount of high-quality annotated data from diverse modalities such as images, text, and audio. However, acquiring such data in large quantities can be challenging and costly. Active learning has emerged as a powerful paradigm to address this issue by selectively annotating the most informative samples, thereby reducing annotation costs and improving model performance. However, existing active learning methods encounter limitations in terms of inefficient data scanning and costly maintenance when dealing with large-scale updates. To overcome these challenges, this study proposes a novel approach called So-CBI (semi-ordered class boundary index) that efficiently retrieves samples for multimodal model training. So-CBI incorporates inter-class boundary perception and a semi-ordered indexing structure to minimize maintenance costs and enhance retrieval efficiency. Experimental evaluations on various datasets demonstrate the effectiveness of So-CBI in the context of active learning.