Beni

Solution

Beni replaced a split architecture built around Milvus and a separate catalog database with a unified MongoDB deployment that holds vector embeddings, product attributes, and user engagement signals in one place. Listings from resale marketplaces such as eBay, Poshmark, and Depop are continuously ingested into MongoDB Atlas, enriched with generated embeddings, and combined with popularity and behavioral data so the company can run hybrid search across more than 300 million one-of-one items. MongoDB Search and MongoDB Vector Search let Beni mix reverse-image and text retrieval with ranking logic based on relevance and shopper activity, while Atlas sharding distributes the growing catalog across smaller, more cost-effective instances. Keeping the ETL pipeline inside one data platform also lets the team add new attributes such as style or boldness without rebuilding an external vector store from scratch.

Data flow

Listings from marketplaces such as eBay, Poshmark, and Depop are ingested into MongoDB Atlas, enriched with embeddings and attributes, combined with user engagement data, and queried through MongoDB Search and MongoDB Vector Search so Beni can return personalized resale results in its extension, app, and website.