Milvus

When your app converts text or images into numerical vectors (via OpenAI, Cohere, or any embedding model), Milvus finds the closest matches across millions or billions of vectors in milliseconds. Apache 2.0, Go/C++. Cloud-native architecture with storage and compute separated. Supports multiple index types (HNSW, IVF, DiskANN), hybrid search (vectors + scalar filters), and multi-tenancy. The SDK supports Python, Java, Go, Node.js, and REST.

Self-hosting is free. Full feature set, no restrictions. Runs on Kubernetes with Helm charts or standalone via Docker. Minimum production setup: 3 nodes for high availability. Zilliz Cloud (managed Milvus) has a free tier with 2 collections and 1M vectors. Paid starts at ~$65/mo for a dedicated cluster. Serverless option available for variable workloads.

Solo: Milvus Lite (embedded mode) or Zilliz free tier for prototyping. Small teams: Zilliz free tier or self-host with Docker. Medium to large: self-host on Kubernetes or Zilliz Cloud depending on ops capacity.

The catch: self-hosted Milvus on Kubernetes is operationally heavy: etcd, MinIO, Pulsar (or Kafka) as dependencies. That's a lot of infrastructure for a vector search feature. If your dataset is under 1M vectors, Chroma or pgvector (Postgres extension) gets you there with dramatically less infrastructure.