Milvus

Milvus stores and searches the vector representations behind text, image and multi-modal data. When a retrieval system outgrows a library holding embeddings in memory, this is the layer that takes over, keeping vectors alongside scalar fields such as integers, strings and JSON objects.

Written in Go and C++, it separates compute from storage in a Kubernetes native architecture, so query nodes scale for read heavy traffic and data nodes for write heavy traffic. A standalone mode fits a single machine, and Milvus Lite installs with pip for local work in Python.

Several design choices start to matter once a workload gets large.

• Index variety: HNSW, IVF, FLAT, SCANN and DiskANN are all supported, with quantization variants and mmap.
• Hardware acceleration: CPU and GPU indexing, including NVIDIA CAGRA.
• Hybrid search: dense and sparse vectors share one collection, with BM25 full text search and functions that rerank results from several requests.
• Multi-tenancy: isolation at database, collection, partition or partition key level lets one cluster serve many tenants.
• Hot and cold storage: frequently read data stays in memory or on SSD while the rest moves to cheaper storage.

Stateless microservices recover quickly from failure and replicas add fault tolerance and throughput, while mandatory authentication, TLS encryption and role based access control cover security. The project sits under the LF AI and Data Foundation with Zilliz as its major contributor, and the tutorials work through retrieval augmented generation, semantic search, recommendation and multimodal retrieval.