AlleCompanion: Production-Scale Complementary Product Recommendation at Allegro
September 7, 2026
Allegro's AlleCompanion framework uses a category-constrained Two Tower architecture to distinguish true semantic compatibility from noisy co-purchase data. It incorporates a Category Adapter to constrain candidate embeddings within logical complementary boundaries for large-scale retrieval.
HOW THIS AFFECTS YOU
●
builderYou can implement category-constrained embedding spaces to reduce noise in product recommendation systems.
●
founderThis demonstrates how to move beyond simple co-purchase heuristics to build high-value semantic upselling engines.