Details
OmniSearchSage is a multi-task, multi-entity embedding model: it learns a single unified representation space for three types of items — search queries (text), Pins (images + text), and products (catalog items) — so that similar items end up close together regardless of their format. The model was described at The Web Conference (WWW '24) in April 2024. Pinterest reports improvements of more than 8% in search relevance, more than 7% in user engagement, and more than 5% in ad click-through rate following deployment. Additionally, Pinterest uses a large language model distillation pipeline — where a large model (LLaMA-3 8B) teaches a smaller, faster model to score relevance — to further improve search quality with less computational cost.
Have evidence about Pinterest's AI practices? Submit a report.
Report a Sighting →