Black-Box Unsupervised Domain Adaptation (BBDA) aims to transfer the knowledge learned in a black-box source model to the target domain without requiring source data or model parameters. By obviating source data or model parameters, BBDA avoids privacy or data transmission issues, thereby expanding the application of domain adaptation in more pragmatic scenarios. Existing BBDA methods suffer from a lack of deliberate learning of target domain. Furthermore, these methods rely solely on the knowledge from the source model, thus being affected by source bias. To address these issues, we propose a novel adaptive ternary division paradigm that incorporates the generic knowledge from Vision Language Models (VLM) to achieve target adaptation. Experiments on various benchmark datasets demonstrate that our method achieves new state-of-the-art performance, validating the effectiveness of the proposed method.