Researchers at Shopify have posted a paper on arXiv describing SID-LLM, a model that handles both product search and recommendations by generating semantic IDs. These are short sequences of tokens derived from product content that stand in for each product. The paper says Shopify's systems must retrieve from a changing catalog of billions of products under strict latency limits.
According to the abstract, SID-LLM is jointly pre-trained on catalog and behavioral data, so search and recommendation share one representation and training objective. The authors report that reinforcement learning improved judged search relevance, and that distilling from task-specific expert models improved recommendations without a statistically detected loss in search quality. For serving, they pair multi-token prediction with a lightweight ranker, which they say improved retrieval quality and cut latency compared with their autoregressive baseline.
The authors say SID-LLM on its own reached quality comparable to Shopify's existing multi-model search stack while meeting production latency requirements. Adding it to Shopify's production search and recommendation stacks produced significant gains in offline evaluations and online A/B tests, they report. The abstract does not give figures or say whether the system is fully deployed.