Milvus

Milvus is an open source vector database built specifically for storing, indexing, and querying high-dimensional vector embeddings at massive scale. As a Cloud Native Computing Foundation (CNCF) incubating project, Milvus is designed with a cloud-native, distributed architecture that enables horizontal scaling to handle billions of vectors with high throughput and low latency.

The architecture of Milvus is organized into separate components with distinct responsibilities. The query nodes handle vector search and scalar filtering. The data nodes manage data persistence and compaction. The index nodes build vector indexes asynchronously. The proxy nodes handle client requests and route them appropriately. This separation of concerns enables each component to be scaled independently based on workload characteristics, making Milvus well-suited for large-scale production deployments.

Milvus supports a comprehensive set of vector index types to accommodate different performance and accuracy trade-offs. The HNSW (Hierarchical Navigable Small World) index provides excellent recall and query performance for in-memory workloads. The IVF (Inverted File) family of indexes (IVF_FLAT, IVF_SQ8, IVF_PQ) provides different accuracy-speed-memory trade-offs using partitioning and quantization. The DiskANN index enables efficient approximate nearest neighbor search over datasets that exceed available memory by leveraging solid-state disk storage. GPU-accelerated indexes (GPU_IVF_FLAT, GPU_IVF_PQ, GPU_CAGRA) leverage NVIDIA GPU hardware for dramatically faster index building and querying on supported hardware.

Milvus supports multiple distance metrics for vector similarity measurement including L2 (Euclidean distance), Inner Product (cosine similarity with normalized vectors), Hamming distance, and Jaccard similarity. This flexibility makes Milvus suitable for diverse embedding types including text embeddings, image embeddings, audio embeddings, and binary feature vectors.

The data model in Milvus is organized around collections (similar to tables), schemas that define the fields of each entity, and partitions for logical data organization within a collection. Milvus supports multiple vector fields within a single collection, enabling multi-vector search scenarios where each entity has embeddings from multiple models or modalities.

Hybrid search in Milvus combines dense vector search with sparse vector search (BM25 or SPLADE) and scalar filtering in a single query. This capability enables search results that combine the semantic understanding of dense embeddings with the keyword precision of sparse representations, making Milvus particularly effective for information retrieval applications.

Milvus integrates natively with LangChain, LlamaIndex, Haystack, and other popular RAG and LLM application frameworks through dedicated client libraries. The PyMilvus Python client is the primary SDK for interacting with Milvus, and clients are also available for Java, Go, Node.js, and other languages.

Zilliz Cloud is the fully managed, enterprise-grade cloud service built on Milvus, offering automated scaling, multi-cloud deployment, and enhanced performance through proprietary optimizations. Milvus Lite is a lightweight version that can be embedded in Python applications for development and small-scale deployment without a standalone server.