H1: hybrid retrieval system for product search in e-commerce

This paper presents the effectiveness of using the H1 system for retrieving products from various vendors in the marketplace. The H1 system is a hybrid model that combines the benefits of lexical-based and semantic-based retrieval techniques, similar to other state-of-the-art product retrieval systems. The novelty of this approach lies in its combination of token-level retrievals. The advantage of the H1 system over other existing solutions is its ability to handle complex search queries containing brands with multi-word brands. For example, search queries like "new balance sneakers" and "Gloria Jeans children's clothing" will be split into separate tokens "new balance" and "Gloria Jeans", respectively, which helps reduce the retrieval model's size and improves its autonomous performance. The H1 system achieved mAP@12 score of 56.1% and R@1K score of 86.6% on the public WANDS dataset, outperforming other state-of-the-art models. These results demonstrate the effectiveness of the approach and its potential for improving product search experiences for online shoppers.