Milvus

Milvus is an open-source vector database built for massive-scale similarity search, capable of handling billions of vectors with millisecond query latencies. Developed by Zilliz, it's designed as a cloud-native, distributed system from the ground up, making it the go-to choice for enterprise deployments that need to scale beyond what single-node vector databases can handle.

Milvus uses a disaggregated architecture with separate components for coordination, data storage, query execution, and indexing. This design allows independent scaling of each component — you can add more query nodes for higher throughput without provisioning additional storage. The system supports multiple index types including IVF (Inverted File), HNSW, DiskANN (for disk-based indexing of datasets that exceed memory), and GPU-accelerated indexes for extreme performance requirements.

The data model in Milvus is collection-based with a schema definition that specifies fields, data types, and index parameters. Unlike simpler vector stores, Milvus supports multiple vector fields per collection, scalar field filtering, dynamic schemas, and partition-based data organization. Partitions are particularly useful for multi-tenant agent applications where each customer's data lives in a separate partition for isolation and efficient querying.

For AI agent stacks, Milvus integrates with LangChain, LlamaIndex, Haystack, and other frameworks through official connectors. The PyMilvus SDK provides both ORM-style and functional APIs. Milvus Lite, a lightweight version that runs in-process, serves as a development and testing environment with API compatibility to the full distributed deployment. Zilliz Cloud offers a fully managed Milvus service for teams that want the power without the operational overhead.

Key strengths include proven scalability (billions of vectors in production at companies like eBay and Shopee), flexible indexing strategies for different performance/cost trade-offs, and strong consistency guarantees through a WAL (Write-Ahead Log) and timestamp-based MVCC. The active open-source community and LF AI & Data Foundation governance provide long-term project stability.

The trade-offs are significant operational complexity for self-managed distributed deployments (MinIO, etcd, and Pulsar/Kafka dependencies), a steeper learning curve compared to simpler alternatives like Chroma or Pinecone, and higher minimum resource requirements. Milvus is best suited for teams with the infrastructure expertise to manage distributed systems or those using Zilliz Cloud for a managed experience.