What We Learned from Using LLMs in Pinterest
summary
Pinterest integrates LLMs into search relevance, fine-tuning a cross-encoder teacher on human labels and using it to label daily logs for semi-supervised distillation into a faster by-encoder student; this pipeline handles over 6 billion searches per month across 45+ languages. VLM-generated image captions plus user action signals (board titles, high-engagement queries) become text features. An 8B-class teacher yields ~12% and ~20% lifts versus stronger baselines, offline-caches pin embeddings, gets ~85% query cache hits, and ships substantial relevance gains.
AI-written summary · about
chapters (8)
- 0:00Introduction to Pinterest and its search functionality.
- 1:52Overview of the Pinterest search backend architecture.
- 2:29The search relevance model.
- 2:55Key learnings from using LLMs for search relevance.
- 5:04The value of VLM-generated captions and user actions as content annotations.
- 7:16Productionizing LLMs with knowledge distillation.
- 12:14The utility of relevance-tuned LLM embeddings as general-purpose semantic representations.
- 13:55Q&A session.
related talks
01h 30mUsing LLMs to improve your Search Engine
24mConversational Search with LLMs
30mWhere there's a will, there's a way: The mad genius of LLMs as a classifiers
15mSearch Engineering Reimagined: Leveraging LLM in ML Workflows
30mLLM-Based Recommendation Systems
49mLLM—Search Engine Killer?
35mReading the Mind of an LLM
28mThe Data That Shapes Foundational LLMs