Search Swarm: Multi-agent Large Language Models Framework for E-commerce Product Search

Abstract

Search engines are vital for online e-commerce butoften struggle with long, detailed queries. We introduceSearch Swarm, a novel multi-agent system designedto improve search engine navigation on platformslike Amazon by accurately locating relevantproducts based on user instructions. Search Swarmemploys multiple large language model (LLM)agents, each with a specific role: query planner,searcher, critic, and attribute selector. These agentscollaborate to generate search queries, evaluate results,and identify the best product options tailoredto users’ needs. Our framework outperforms existingmethods like ReAct and Reflexion in theWebShop environment, achieving a reward score of62.64, compared to scores of 54.1, 59.8, 61.5, and58.2 for other approaches. Furthermore, in a comparisonwith a basic rule-based method on Amazon,Search Swarm achieved a score 38.71 pointshigher and a 41% greater success rate, demonstratingits superior ability to provide relevant productmatches over traditional search engines.