Shopify deploys shared generative discovery model for search and recommendation

Shopify has deployed a shared generative discovery model in production that jointly optimizes search and recommendation, delivering a 7.95% lift in orders and 9.85% lift in conversion rate on the Shop home feed.
The model represents billions of products using three-token semantic IDs generated via RQ-KMeans, allowing related items to share training signals and improving discovery recall for new products by 10-15% relative.
Serving latency is optimized using Liquid AI's LFM2-2.6B base model and multi-token prediction, reducing constraint overhead from over 50 ms to about 1 ms per request while maintaining quality.
Future development focuses on integrating intent reasoning, multimodal visual representations, and online continuous training into a single model that spans candidate generation and ranking.

Technical Implementation

Product embeddings are converted into semantic IDs using RQ-KMeans with three codebooks of 8,192 entries each, selected over RQ-VAE for better downstream performance.
Training combines search and recommendation data, with experiments showing that removing either stream degrades performance on the other, confirming mutual benefit from joint training.
Offline experiments show that reinforcement learning using an LLM relevance judge improves search retrieval quality by 1.5 points in OneDCG@10.

The production stack uses the RIPL inference engine with a customized TensorRT-LLM configuration for the LFM2-2.6B model.
Constrained decoding is accelerated by representing the catalog prefix tree as a sparse GPU index, enabling vectorized validation of valid product paths.
Multi-token prediction predicts all three codewords in a single pass, allowing larger candidate sets to be generated within the same latency budget as autoregressive generation.