PinLanding: Turn Billions of Products into Instant Shopping Collections with Multimodal AI

Faye Zhang, Jasmine Wan, Qianyu Cheng, Matthew Hichar, Eric Wan, Jinfeng Rao13 Jan 2026

Problem

Online retailers and social platforms hold catalogs of billions of items. Shopping collections (topic pages such as "yellow party dress for summer") were built from user search history and manual curation, so they cover the head of popular queries and miss the long tail. Long conversational queries ("long red satin dress with lace trim under $200") have no page at all. At Pinterest's catalog scale, running a multimodal model on every item to fill the gaps is too expensive and too fragile.

Example from the article: a vision-language model returns "boho", "bohemian" and "boho-chic" for one concept, plus ultra-specific tags like "black insoles", so most attributes apply to very few products and no collection can be built from them.

Solution in brief

Generate collections from the content, not from query logs. A vision-language model (VLM) proposes attributes, a curation step shrinks them to a small vocabulary, and a CLIP-style dual encoder assigns that vocabulary to every product. A collection is a tuple of attributes, filled by scoring products against it.

Questions this case answers
  • How to cover the long tail of search intents with topic pages when the existing pages come only from query logs and manual curation
  • How to label a catalog of millions of items with a controlled vocabulary when a generative model is too costly and too noisy to run on every item
  • How to generate thousands of searchable shopping collections from a product catalog instead of mining them from search logs?
  • How to turn noisy free-form VLM attributes into a compact, consistent attribute vocabulary?

Tools and methods

  • VLM — proposes key-value attributes from product image and metadata.
  • Frequency filter, embedding merge, LLM-as-judge — curate attributes into a vocabulary and rate topics.
  • CLIP-style dual encoder — cheap attribute assignment to every product.
  • Ray — streaming batch inference on a CPU+GPU cluster.
  • Apache Spark — scoring and joining topics with products.
Continue reading — it's free

Google or a one-time email code. No password.