Faire faces unique challenges in recommending products within its vast marketplace, which has millions of products and numerous brands. The company initially used DeepFM, an embedding-based method, but found it insufficient for generalizing to new items or for browsing behavior where retailers seek personalized recommendations without specific intent. To improve this, Faire explored relationships between retailers and products, recognizing that retailers with similar tastes often purchase related items. By modeling these interactions as a graph, they can discover new, relevant products even when past engagement data is limited. The core of their solution is a bipartite engagement graph linking retailers and products. Each retailer node is enriched with metadata like store type and country, while product nodes incorporate features derived from names and descriptions. This setup allows the system to create high-quality embeddings for both retailers and products. The architecture employs separate embedding towers for retailers and products, utilizing pre-trained text embeddings for efficiency. A graph attention layer (GATConv) processes these embeddings, allowing the model to weigh connections based on the strength of interactions, ensuring that more frequent purchases have a greater influence on recommendations. To manage scalability, especially given the high number of connections, Faire uses one-hop neighbor sampling to limit the number of products each retailer considers during computation. This method ensures that the model remains efficient while still providing relevant recommendations based on the most influential connections. Overall, the implementation of graph neural networks enhances the personalization of recommendations, helping retailers discover products that align with their preferences and expanding beyond their previous purchases.