How Amazon Uses LLMs to Recommend Products

Amazon's search team built COSMO, a commonsense knowledge graph powered by LLMs, to bridge the semantic gap between customer search queries and product listings. Traditional keyword-matching systems fail when intent requires human reasoning (e.g., 'shoes for pregnant women' → slip-resistant shoes). COSMO uses OPT-175B to generate millions of candidate commonsense explanations from query-purchase and co-purchase behavior pairs, then filters them through rule-based, similarity, and human-annotated classifiers (DeBERTa-large) to produce a graph of 6.3M nodes and 29M edges. A smaller instruction-tuned model (COSMO-LM, based on LLaMA 7B/13B) handles new queries in production at lower cost. Deployed via a two-tier caching architecture on SageMaker, COSMO powers search relevance, session-based recommendations, and navigation. A/B tests on ~10% of U.S. traffic showed a 0.7% relative sales increase and 8% navigation engagement lift, translating to hundreds of millions in additional annual revenue.