Milvus, Joute's review

The essentials

• Open source vector database designed for large-scale production
• Free (self-hosted) or Zilliz cloud version
• Billions of vectors indexed, high availability, ACID-compliant
• For engineering teams deploying RAG or semantic search applications in production

What is Milvus

Milvus is the reference open source vector database for applications requiring large-scale similarity search. When you're building a RAG system, a semantic search engine, or a recommendation system with millions or billions of vectors, Milvus is built for that. Originally developed by Zilliz, the project is now an Apache-incubated solution with an active community.

Strengths

Production scalability

Milvus handles billions of vectors with sub-millisecond latency. The indexing algorithms (HNSW, IVF, DiskANN) are optimized for both GPU and CPU.

Distributed architecture

Milvus is designed for horizontal scaling. You can distribute load across multiple nodes with high availability.

Rich ecosystem

Official connectors for Python, Java, Go. Integrations with LangChain, LlamaIndex, Haystack.

Limits

High operational complexity

Deploying Milvus in production requires Kubernetes and DevOps expertise. The learning curve is significant.

Oversized for small projects

For a few thousand vectors, Chroma or Qdrant are simpler and sufficient.

Pricing

Free as open source (self-hosted). Managed cloud version via Zilliz with usage-based pricing.

Alternatives

For more simplicity in development: Chroma. For a managed cloud vector database: Qdrant. For the managed Pinecone ecosystem: Pinecone.

Verdict

Milvus is the right choice when you have serious scale constraints in production. For exploration and small projects, start with Chroma or Qdrant and migrate to Milvus when you need it.

FAQ

What's the difference between Milvus and Zilliz?

Milvus is the open source project. Zilliz is the company that maintains it and offers a managed cloud version (Zilliz Cloud) with less operational management.

Can Milvus be used locally in development?

Yes, a Milvus Lite version runs in pure Python without a server, ideal for development and testing.

Does Milvus support hybrid search (vector + metadata filters)?

Yes, hybrid search with filtering on scalar fields is a key Milvus feature.

What types of vectors does Milvus support?

Dense vectors (standard embeddings), sparse vectors, and binary vectors. Multi-vector support per entity.

Joute may earn a commission on subscriptions taken out via the links in this article. This doesn't change our reviews.