<p>Partially relevant video retrieval (PRVR) aims to identify untrimmed videos associated with textual queries. Current models face challenges in handling redundant multi-scale video clips and often fail to capture multi-granularity semantics, which are critical for aligning variable-length clips with text queries. To address these issues, an efficient Adaptive Window and Length-sensitive clustering Network (AWLC-Net) is proposed, which incorporates two key optimization strategies. First, to reduce redundant clips for the clip-scale video feature construction, the adaptive window-based gaussian attention module in AWLC-Net dynamically constrains attention to localized temporal windows while preserving multi-scale modeling capabilities. Second, by leveraging a fast k-medoids algorithm, length-sensitive clustering is designed to comprehensively consider semantic and length information, thereby selecting diverse clips and capturing rich semantics. Experimental results on two benchmark PRVR datasets demonstrate that our method significantly improves accuracy and efficiency while effectively addressing video clip redundancy and multi-granularity semantics.</p>