Milvus is an open-source vector database built for storing and searching embeddings — the high-dimensional vectors that represent text, images, and other data by meaning. As AI applications lean on semantic search and retrieval-augmented generation, purpose-built vector stores like Milvus have become core infrastructure. Its whole value is doing similarity search over huge collections of vectors fast, which is exactly what general-purpose databases struggle with.
Why a dedicated vector database
You can bolt vector search onto a general database with an extension, and for modest scale that is often enough. A dedicated vector database like Milvus earns its place when you have many vectors — millions to billions — and need fast similarity search over them. It is engineered specifically for the approximate-nearest-neighbour problem at scale, with the indexing, sharding, and scaling to match. Below that scale, the dedicated system may be more infrastructure than the problem warrants.
Indexing trades recall for speed
The heart of a vector database is its index. Searching billions of vectors for the nearest neighbours exactly is far too slow, so vector databases use approximate-nearest-neighbour (ANN) indexes that trade a little accuracy for enormous speed. Milvus supports several index types, each a different point on the trade-off between search speed, memory use, and recall (the fraction of true nearest neighbours found). Choosing and tuning the index for your accuracy and latency needs is the central engineering task — and understanding that it is an approximate search is essential to using it correctly.
When its scale is justified
Milvus is built for scale — distributed, horizontally scalable, handling very large vector collections. That capability is the reason to choose it and also the reason not to, if you do not need it. For a small collection of vectors, a lighter embedded vector store or a general database extension is simpler to run. Reach for Milvus when your vector count and query load genuinely require a distributed, purpose-built system; its scale is a benefit only once you are operating at scale.
Unlike an exact database lookup, ANN vector search returns the probable nearest neighbours, not guaranteed ones, and the index's recall setting controls how complete the results are. Tune recall against latency for your use case, and design the application knowing a relevant result can occasionally be missed. Treating approximate search as if it were exact leads to subtle quality bugs that are hard to trace back to the index.
Milvus is a purpose-built vector database for fast approximate similarity search over large embedding collections, justified when your scale — millions of vectors and up — outgrows what a general database extension handles. Its power lives in ANN indexing that trades a little recall for large speed gains, which you tune to your latency and accuracy needs. Choose it when the scale is real, and design for approximate results rather than exact ones.
Independent software engineer in Nairobi specialising in Acumatica customisations, Laravel backends, and tax fiscalisation integrations across East and Southern Africa.