How MIXI built natural language search over billions of photos with Amazon S3 Vectors

This article describes how MIXI built descriptive search for FamilyAlbum, enabling 30 million users to find photos among 20 billion items using natural language queries powered by Amazon S3 Vectors.

• Descriptive search converts images and text queries into embeddings using CLIP-Japanese-Base model for similarity matching
• MIXI initially used provisioned infrastructure but faced cost constraints at scale
• Amazon S3 Vectors provides serverless, pay-per-use vector storage supporting up to 2 billion vectors per index
• Migration achieved 75% reduction in database management costs with consistent sub-1-second search performance
• Index partitioning by family ID and batch PUT optimization enabled efficient scaling without manual sharding
• Simplified operations by eliminating shard management and providing consistent performance under variable load

S3 Vectors proved ideal for FamilyAlbum's cost-sensitive, large-scale vector search workload, enabling sustainable growth across billions of media items.