<![CDATA[The rapid growth of e-commerce catalogs complicates product discovery, particularly for items with complex technical specifications like luxury watches. Conventional keyword searches and TF-IDF methods often fail to capture underlying semantic relationships. The primary objective of this study is to propose a robust e-commerce recommendation system utilizing Content-Based Filtering enhanced with Sentence-BERT (SBERT) semantic embeddings. This study employs an experimental comparative research design. The methodology involves aggregating product attributes into descriptive sentences and pre-processing them to minimize representation bias. These sentences are transformed into high-dimensional embeddings using the lightweight all-MiniLM-L6-v2 SBERT model, with similarities calculated via the Cosine Similarity algorithm. The system's performance is comparatively evaluated against a baseline TF-IDF method. Main outcomes and experimental results across 20 testing scenarios demonstrate that SBERT significantly outperformed the baseline, achieving an average Precision@5 of 93.00%, a Recall@5 of 2.71%, and a highly efficient latency of 0.38 ms. In conclusion, SBERT provides a superior, scalable solution for recommending complex products by accurately capturing the semantic similarity of technical specifications and textual representations of visual characteristics. The approach inherently resolves the item cold-start problem, and its successful integration into a web application confirms its feasibility for real-time similarity computation in modern e-commerce platforms.]]>
