Shopify’s Generative Discovery Model
AI Summary
Shopify outlines its production-grade generative discovery model, which unifies search and recommendation across billions of products. By adapting a pretrained open-weight language model, encoding product catalogs into semantic IDs via RQ-KMeans, and jointly training on search and recommendation logs, Shopify improved retrieval relevance and candidate generation. The system addresses the latency constraints of live requests using Liquid AI’s LFM2-2.6B model optimized via the RIPL inference stack and TensorRT-LLM.
Key Takeaways
- The catalog uses RQ-KMeans with three codebooks of 8,192 entries each to map product embeddings into 3-token semantic IDs.
- Search and recommendation are jointly trained; removing either stream degrades overall metrics, proving mutual reinforcement.
- Reinforcement learning utilizing an LLM relevance judge yielded a +1.5 lift in OneDCG@10 for search.
Why it matters
Scaling product discovery across a multi-billion item catalog within tight real-time latency budgets is a fundamental challenge in e-commerce architecture. This approach demonstrates how generative AI, semantic ID tokenization, and multi-task learning can replace disjointed legacy search and recommendation systems with a unified foundation model.
Conversation 0
Be the first to share your perspective.