Hint-Augmented Re-ranking: Efficient Product Search using LLM-Based Query Decomposition

2025-11-17

AI Summary

In e-commerce search, queries containing superlative terms (e.g., “best”, “most popular”) imply implicit multi-attribute comparisons, demanding both deep semantic understanding and low-latency inference. To address this, we propose a prompt-enhanced two-stage re-ranking framework. First, a large language model (LLM) parses the user’s latent intent and decomposes the original query into structured attribute-value prompts. Second, these prompts are injected into a lightweight ranking model, enabling effective transfer of high-fidelity semantic knowledge while preserving real-time performance. This design circumvents the prohibitive latency of direct LLM-based re-ranking. Experiments demonstrate substantial improvements: +10.9 MAP and +5.9 MRR over strong baselines, with no degradation in deployment efficiency. Our key contributions are (i) an intent-driven structured prompt generation mechanism and (ii) a co-optimization strategy that tightly couples prompt engineering with lightweight model adaptation.

Technology Category

• Search and Optimization: Distributed Search
• Machine Learning: Learning Preferences or Rankings
• Natural Language Processing: Prompt Engineering / Prompting

Application Category

• Search and Retrieval-Augmented AI: Web search models and ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Search queries with superlatives (e.g., best, most popular) require comparing candidates across multiple dimensions, demanding linguistic understanding and domain knowledge. We show that LLMs can uncover latent intent behind these expressions in e-commerce queries through a framework that extracts structured interpretations or hints. Our approach decomposes queries into attribute-value hints generated concurrently with retrieval, enabling efficient integration into the ranking pipeline. Our method improves search performanc eby 10.9 points in MAP and ranking by 5.9 points in MRR over baselines. Since direct LLM-based reranking faces prohibitive latency, we develop an efficient approach transferring superlative interpretations to lightweight models. Our findings provide insights into how superlative semantics can be represented and transferred between models, advancing linguistic interpretation in retrieval systems while addressing practical deployment constraints.

Problem

• Interpreting superlative queries requiring multi-dimensional comparison in e-commerce search
• Addressing prohibitive latency of direct LLM-based reranking for practical deployment
• Transferring linguistic interpretations from LLMs to lightweight models efficiently

Innovation

• LLM-based query decomposition for e-commerce search
• Extracting attribute-value hints from superlative queries
• Transferring interpretations to lightweight models for efficiency