milvus-io/milvus

Milvus is a high-performance, open-source vector database designed to store and search vast collections of vector data, the kind of mathematical representations (called embeddings) that AI models use to understand text, images, audio, and other unstructured content. The core problem it solves is that traditional databases like PostgreSQL or MySQL are built for exact matches or range queries on structured data, but AI applications need a different kind of search: finding items that are semantically similar rather than exactly equal. When an AI model converts a phrase like "What is machine learning?" into a long list of numbers (a vector), Milvus can efficiently find all stored vectors that are mathematically closest to it, a technique called Approximate Nearest Neighbor (ANN) search. This is the foundation of features like semantic search, recommendation engines, image similarity finders, and retrieval-augmented generation (RAG), where a chatbot fetches relevant documents before answering a question. Milvus works by organizing vectors into collections, building specialized index structures (such as HNSW, DiskANN, or IVF variants) that allow it to skip most of the data during a search and still return accurate results quickly. It supports metadata filtering alongside vector search, so you can combine similarity ("find documents like this one") with traditional filters ("only from the last 30 days"). Under the hood it is written in Go and C++, with GPU acceleration support for even faster indexing via NVIDIA's CAGRA library. The system comes in three deployment sizes: Milvus Lite runs entirely in Python for quick experiments; Standalone mode runs on a single machine via Docker; and the fully distributed Kubernetes-native mode scales horizontally to handle billions of vectors across many machines. Zilliz Cloud offers a fully managed hosted version for teams that want to skip infrastructure management entirely. Developers building AI-powered search, recommendation, or question-answering products would reach for Milvus when they need production-grade reliability and throughput beyond what smaller in-process libraries like FAISS can provide.