Think of COSMO as Amazon’s ‘common sense brain’ for shopping: it teaches computers the everyday knowledge humans use when browsing products (like knowing that hiking boots go with rainy weather, or that a phone case should fit a specific phone model) and then uses that knowledge to make search, recommendations, and product understanding much smarter.
Traditional product search and recommendation only see what’s literally written in titles and descriptions; they miss implicit relationships, usage context, and everyday common sense (e.g., that a yoga mat is for exercise, or that a certain cable is compatible with a device). COSMO generates and serves large-scale, machine-usable common sense knowledge so that e-commerce systems can better understand products, queries, and customer intents, improving relevance and conversion.
Proprietary graph/knowledge base built from Amazon’s massive catalog, customer interaction signals, and historical data, combined with large-scale ML pipelines and serving infrastructure that are deeply integrated into Amazon’s e-commerce stack. This creates a hard-to-replicate feedback loop of data, models, and deployment scale.
Hybrid
Knowledge Graph
High (Custom Models/Infra)
Maintaining freshness, consistency, and quality of a massive, evolving common sense knowledge graph across millions of products and billions of interactions, while keeping inference and query latency low for real-time e-commerce workloads.
Early Majority
Unlike generic semantic search or off-the-shelf recommendation engines, COSMO is focused specifically on generating and serving structured common sense knowledge tailored to e-commerce: product usage, compatibility, context, and relationships. It acts as a reusable knowledge layer that can be tapped by many internal systems (search, recommendations, ads, QA) rather than being a single-purpose model.