milvus-io/milvus

A purpose-built vector database designed to handle billions of vectors with low-latency approximate nearest neighbor search. It supports multiple index types including HNSW, IVF, and DiskANN, so you can tune the cost-versus-recall tradeoff to match your workload rather than accepting someone else's defaults. Deployment options range from a lightweight standalone mode for early prototyping to a fully distributed Kubernetes setup when you need horizontal scale. The architecture separates storage from compute, which makes it genuinely cloud-native rather than just containerized. Worth a serious look if you are building retrieval-augmented generation pipelines, semantic search, recommendation engines, or multimodal similarity features where embedding volumes will grow fast. The SDK coverage across Python, Go, Java, and Node is solid, and the documentation is unusually thorough for an open-source infrastructure project. The honest reservation is operational complexity at scale. Running the distributed version introduces a meaningful DevOps surface area that small teams should budget for before committing. -> Best for: technical founders building AI-native products who expect high embedding volume and want to avoid per-query managed-service costs.