QUPID: Quantified Understanding for Enhanced Performance, Insights, and Decisions in Korean Search Engines
Ohjoon Kwon, Changsu Lee, Jihye Back, Lim Sun Suk, Inho Kang, Donghyeon Jeon
TL;DR
This work introduces QUPID, a heterogeneous ensemble that pairs a generative small language model with an embedding-based SLM to perform query-document relevance labeling in Korean information retrieval. By combining explicit generation-based judgments with implicit semantic similarity, QUPID achieves higher labeling accuracy than representative LLMs while delivering dramatically lower inference latency (about $60\times$ faster) and enabling scalable deployment in production search pipelines. The approach relies on a curated dataset of roughly $1$ million Q-D pairs, including real-world and synthetic hard negatives, and trains each component to output probabilistic relevance labels that are fused via a lightweight ensemble. Empirical results show substantial gains in Cohen's Kappa ($\kappa = 0.646$) and $n\mathrm{DCG}@5$ (approximately $+1.9\%$), underscoring the practical impact of architectural diversity for efficient, high-quality information retrieval.
Abstract
Large language models (LLMs) have been widely used for relevance assessment in information retrieval. However, our study demonstrates that combining two distinct small language models (SLMs) with different architectures can outperform LLMs in this task. Our approach -- QUPID -- integrates a generative SLM with an embedding-based SLM, achieving higher relevance judgment accuracy while reducing computational costs compared to state-of-the-art LLM solutions. This computational efficiency makes QUPID highly scalable for real-world search systems processing millions of queries daily. In experiments across diverse document types, our method demonstrated consistent performance improvements (Cohen's Kappa of 0.646 versus 0.387 for leading LLMs) while offering 60x faster inference times. Furthermore, when integrated into production search pipelines, QUPID improved nDCG@5 scores by 1.9%. These findings underscore how architectural diversity in model combinations can significantly enhance both search relevance and operational efficiency in information retrieval systems.