Finding the perfect DIY project or that dream recipe on Pinterest just got easier. Behind the scenes, Pinterest is leveraging the power of large language models (LLMs) to supercharge its search relevance. Ever wonder how Pinterest manages to surface exactly what you're looking for amidst billions of pins? The secret lies in a sophisticated blend of AI and human ingenuity. Traditionally, search engines relied on basic keyword matching. However, Pinterest’s visual nature demands a deeper understanding of context and intent. That’s where LLMs come in. These powerful AI models, trained on vast amounts of text and code, can discern the subtle nuances of language and connect seemingly disparate ideas. In Pinterest's case, LLMs are used to analyze not only search queries, but also rich text associated with each pin, including titles, descriptions, and even captions generated from the images themselves. They also tap into user-curated boards and linked webpage content to form a comprehensive understanding of what each pin represents. This multifaceted approach allows Pinterest’s search engine to go beyond simple keyword matching and grasp the true intent behind a user's search. For example, a search for “rustic wedding centerpiece” might also surface pins related to “DIY mason jar decorations” or “outdoor table settings,” even if those keywords aren't explicitly present in the pin’s description. But processing power from LLMs comes at a cost: speed. To ensure a snappy search experience, Pinterest employs a clever trick called “knowledge distillation.” This involves training a smaller, faster model to mimic the behavior of the larger, more complex LLM. This smaller “student” model learns from the LLM’s vast knowledge base but operates much more efficiently, delivering real-time search results without the lag. Through this innovative combination of cutting-edge AI and strategic optimization, Pinterest is continuously refining its search capabilities, making it easier than ever to discover the inspiration you need. This approach has led to significant improvements in search relevance and user engagement worldwide. While the current system relies on distilling knowledge from LLMs into smaller models, Pinterest is already looking towards the future, exploring the integration of real-time servable LLMs and advanced visual-language models to take search relevance to the next level.