Scaling Conditional Learned Retrieval for Pinterest Home Feed

Back to Devoured - August 31, 2026

DEVOURED

Scaling Conditional Learned Retrieval for Pinterest Home Feed

Data aibackend Medium

Pinterest scaled its home feed retrieval by conditioning candidate embeddings on specific user contexts like interests and board data.

What: Pinterest updated its retrieval stack to use conditioned sequence modeling, semantic IDs, and GPU-accelerated serving to represent diverse user intents without increasing compute costs.

Why it matters: Personalized feeds often rely on static embeddings that fail to capture dynamic user context; conditioning embeddings on explicit session intent allows for more precise content surfacing.

Decoder

Candidate Generation: The first stage of a recommendation system that filters millions of items down to a few thousand for ranking.
Semantic ID: A compact, learned representation of an item's content used for efficient similarity search.

Original article

Pinterest scaled candidate generation by conditioning retrieval embeddings on explicit contexts such as interests, boards, and pins. The system combines conditioned sequence modeling, semantic IDs, unified models, slot routing, request deduplication, and GPU serving so one home feed can represent multiple user intents without exploding serving cost.

Explore all 65 articles in today's digest