DOI: https://doi.org/10.1145/3726302.3731948
Modern e-commerce platforms require search engines that go beyond simple keyword matching to accurately capture customer intent. Traditional keyword-based retrieval struggles with complex, multi-attribute queries, potentially leading to suboptimal results and poor customer experience. To address these challenges, we introduce a late interaction-based semantic search engine designed for IKEA product search. This approach significantly improves retrieval quality while maintaining low latency, ensuring a more effective and seamless search experience for customers.Our approach departs from single-vector embeddings by leveraging token-level late interaction scoring, enabling fine-grained alignment between search queries and product descriptions. To enhance search effectiveness, we introduce three key contributions: (1) large-scale synthetic query generation using LLMs to diversify training data, (2) strong negative sampling to improve contrastive learning, and (3) adaptive thresholding to dynamically refine ranking cutoffs and prevent over-retrieval biases.In live A/B testing on IKEA.com for long tail queries in the U.S. market, our system outperforms IKEA Boolean search with a 3.1% increase in click-through rate, a 1.96% boost in conversions, 1.78% increase in search interaction rate, and a 2.18% rise in add-to-cart actions. These results validate the effectiveness of efficient token-level retrieval and adaptive ranking in large-scale commercial search.