Milvus

Milvus is an open-source vector database designed to organize and search large volumes of unstructured data, including text and images, by storing them as vectors. The software is deployed as a server and can be run in a distributed Kubernetes-native architecture, a standalone single-machine mode, or as a lightweight Python library via Milvus Lite.

Users interact with the system through the pymilvus Python SDK to create collections, ingest data, and perform similarity searches. It is primarily used for building AI applications that require fast retrieval of embeddings, such as recommendation engines, image search tools, and multimodal retrieval systems. The system handles the storage of learned representations of data, allowing for efficient nearest-neighbor search across billions of vectors.

Key features

• Support for multiple index types including HNSW, IVF, FLAT, SCANN, and DiskANN
• Hardware acceleration for both CPU and GPU indexing
• Hybrid search combining dense vectors with sparse vectors and BM25 full text search
• Distributed architecture that separates compute from storage for horizontal scaling
• Multi-tenancy isolation at the database, collection, partition, or partition key level
• Hot and cold storage tiers to balance performance and cost
• Role-Based Access Control and TLS encryption for data security
• Integration with AI frameworks such as LangChain and LlamaIndex

Written in Go and C++, the system is built for high availability and fault tolerance. It allows developers to scale query nodes for read-heavy workloads and data nodes for write-heavy workloads independently. The database supports storing vectors alongside scalar data types like JSON objects, strings, and integers, enabling metadata filtering during search operations. This architecture ensures that the system can maintain performance levels even as the dataset grows to billions of vectors.

Milvus is a scalable infrastructure component for developers building Retrieval-Augmented Generation (RAG) and other generative AI applications.