• Faye Zhang, Staff Software Engineer; Jasmine Wan, Machine Learning Engineer I; Qianyu Cheng, Machine Learning Engineer II; Matthew Hichar, Machine Learning Engineer II; Eric Wan, Sr. • Software Engineer; Jinfeng Rao, Sr. • Staff Machine Learning Engineer Online retailers and social platforms now operate catalogs with billions of items. • Pinterest is one example, but the underlying challenge of how to organize products into precise, navigable shopping collections at web scale is shared across large e‑commerce and social discovery systems. • Historically, collections have been derived from user search history and manual curation. • In the age of multimodal large language models (LLMs), it is now possible to invert this process and generate collections directly from the content itself while still grounding them in how people search.
Article Summaries:
- Pinterest has unveiled PinLanding, a production‑ready pipeline that uses multimodal large language models to automatically generate shopping collections from product content. The system first analyzes user search history, autocomplete, filter usage and browsing paths to map shopping intent and identify high‑volume “head” queries versus long‑tail conversational requests. It then fine‑tunes a vision‑language model to produce structured attribute tuples (e.g., color, style, price) for each item, normalizing and de‑duplicating the output. These attributes feed a vocabulary validated by an LLM‑as‑judge, enabling the construction of attribute‑based feeds and continuous evaluation to improve AI‑native search coverage.
Sources: