Milvus

Milvus is the highest-scale open-source vector database, designed for billion-vector workloads. Its distributed architecture separates compute, storage, and coordination, and its index zoo (HNSW, IVF_FLAT, IVF_PQ, DiskANN, SCANN, GPU-CAGRA) is the broadest of any open-source engine.

For AI agents at scale, Milvus's strengths are partition-based multi-tenancy, hybrid search across multiple vector fields (dense + sparse + multi-vector reranking), and a managed serverless tier through Zilliz Cloud with pay-per-use billing. The integrated pymilvus[model] package exposes OpenAI, Cohere, Voyage, and SentenceTransformer embedders so you can ingest text directly.

The tradeoff is operational complexity. Self-hosting Milvus in production means running etcd, MinIO (or S3-compatible storage), and Pulsar or Kafka — far more moving parts than Qdrant or Chroma. Milvus Lite (single-file embedded mode) covers prototyping, but production teams typically end up on Zilliz Cloud to avoid the ops burden. Milvus does not yet ship a first-party MCP server.

Agent-specific features:

• Partition keys for cheap per-tenant isolation inside a single collection
• Hybrid search with multi-vector fields and pluggable rerankers (RRF, weighted)
• DiskANN and GPU-CAGRA indexes for billion-scale corpora that don't fit in RAM
• Built-in embedding functions via pymilvus[model]
• Serverless and dedicated tiers on Zilliz Cloud; BYOC on AWS, GCP, and Azure