Milvus: Building a Large-Scale Vector DB for LINE VOOM's Real-Time Recommendation System

LINE VOOM, LINE's video-centric SNS, operates a 'For you' tab that provides personalized content based on user feedback.

The existing recommendation system processed embedding generation, storage, and candidate extraction through a daily offline batch approach. This caused a 'lack of immediacy' problem, where new content took up to a day to be reflected in the recommendation list.

To solve this problem, they decided to switch the system to a model-based real-time candidate generation and online vector search approach. To do this, they had to solve the following two key challenges:

• Securing an online storage to store post embeddings
• Performing instant real-time similarity search without an intermediate storage step

The LINE VOOM team determined that a dedicated vector DB capable of simultaneously supporting real-time storage and similarity processing was needed, and after reviewing various platforms, decided to adopt Milvus.