Shopify's SID-LLM Unifies Search and Recommendation via Generative Semantic IDs
_reachsumit · x · 2026-11-10
Shopify has published an arXiv paper on SID-LLM, which casts both product search and recommendation as generating semantic IDs — short content-derived token sequences representing products — over a catalog of billions of items under strict latency constraints.
Key points:
- Joint pre-training on catalog and behavioral data gives search and recommendation a shared representation and training objective.
- Reinforcement learning further improves judged search relevance.
- Distillation from task-specific experts beats joint supervised fine-tuning on recommendation without a statistically detected loss in search quality.
- For serving, multi-token prediction plus a lightweight ranker improves retrieval quality over the autoregressive baseline while cutting latency.
- As a standalone retriever it matches Shopify's existing multi-model search stack at production latency; integrated into production search and recommendation it delivers significant gains in offline evals and online A/B tests.
More from Apps
- Mainframe, a design tool for designers who ship, opens waitlist — vinvan · 2026-11-10
- Turn any YouTube tutorial into an AI-coached learning playbook with one prompt — jacob_posel · 2026-11-10
- Psylo 1.4.0 ships with iPhone Duo support, multi-window tabs, stronger anti-fingerprinting — gnukeith · 2026-11-10
- Grok Bot tests in-chat rental car search with live desk hours, address and phone — XFreeze · 2026-11-10
- Hera shows how to mass-produce brand ad creatives with AI in 4 steps — ycombinator · 2026-11-10
- Muse AI companion launches on Apple Watch and WearOS, Alexandr Wang announces — alexandr_wang · 2026-11-10