A vector database that actually wants to scale

What it does

Milvus stores and searches vector embeddings—those high-dimensional representations of text, images, and other unstructured data that power modern AI applications. It handles approximate nearest neighbor (ANN) search at scale, supports metadata filtering and hybrid search, and runs either as a single-node standalone or a distributed cluster. A lightweight Python wrapper, Milvus Lite, exists for local quickstarts.

The interesting bit

Most vector databases bolt distribution on later; Milvus was born distributed. Its compute-storage separation lets you scale query nodes independently from data nodes—handy when your RAG app suddenly gets read-heavy. The project also supports GPU indexing via NVIDIA CAGRA and keeps one foot in sparse vectors (BM25, SPLADE, BGE-M3) for full-text search, not just dense semantic similarity.

Key highlights

• Separated compute and storage architecture with horizontal scaling on Kubernetes
• Supports HNSW, IVF, FLAT, SCANN, DiskANN, plus quantization and mmap variants
• GPU acceleration available, including NVIDIA CAGRA
• Hot/cold tiering: hot data in memory/SSD, cold data on cheaper storage
• Multi-tenancy isolation at database, collection, partition, or partition-key level
• Mandatory auth, TLS, and RBAC for enterprise deployments
• Backed by Zilliz; also offered as managed Zilliz Cloud with serverless, dedicated, and BYOC options

Caveats

• The README is heavy on managed-service upsells and quickstart snippets; architecture details require clicking through to external docs
• 44K stars suggest popularity, but the project is a substantial distributed system—expect operational complexity at scale, not a drop-in SQLite replacement

Verdict

Worth evaluating if you’re building RAG, recommendation, or image-search systems that need to grow past a single node. Skip it if you just need a quick in-process vector index for a prototype; Milvus Lite or simpler alternatives will save you hours.