Milvus is an open-source vector database for searching unstructured data. Initiated by Zilliz and donated to the LF AI & Data Foundation, it uses the Apache 2.0 license. It stores vectors with scalar fields and provides collections, insert, upsert, delete, query, metadata filtering, approximate nearest-neighbor, full-text and hybrid-search operations. Milvus is not an embedding model or an answer generator; ingestion, authorization, vectorization and generation remain separate parts of a production RAG stack. The brand boundary matters: Milvus is the community-governed open-source project, while Zilliz Cloud is Zilliz’s managed commercial Milvus service with its own plans, SLA, console and compliance features.
Quick Verdict
- Best for: infrastructure teams needing a self-hostable, horizontally scalable vector store with substantial hybrid-search control.
- Worth using? Yes when scale, index choice or deployment ownership matters. Start with Milvus Lite for prototypes; compare Zilliz Cloud when operational simplicity matters more.
- Related tools: Jina AI supplies embeddings/reranking, while RAGFlow, Dify and LightRAG sit higher in the stack.
Best For
Milvus fits RAG platforms, recommendation systems, image or video search, similarity and anomaly systems, and organizations controlling data placement and index parameters. A small local dataset may only need Lite. Teams without Kubernetes, capacity planning, monitoring and backup expertise should compare a managed service. Milvus does not create vectors by itself; Jina AI or another embedding provider fills that role.
Key Features
- Database lifecycle operations: Collections and schemas plus insert, upsert, delete, get/query, bulk import and migration tooling support changing corpora, not just one ANN call.
- Indexes and search: FLAT, IVF, HNSW, SCANN, DiskANN and quantized variants offer different recall, memory, build-time and update tradeoffs.
- Hybrid retrieval: Dense vectors, sparse vectors, BM25 full text, scalar filters and multiple vector fields can be combined and reranked.
- Deployment choices: Milvus Lite is a local-file developer path, Standalone runs on one machine, and Distributed targets horizontal Kubernetes scale. Their operating envelopes are not interchangeable.
- Reliability and security primitives: Sharding, replicas, coordinator high availability, authentication, TLS and RBAC can support production designs, but configuration and recovery drills determine whether they work in practice.
Use Cases
- Store document-chunk vectors with tenant, authorization and time metadata, then filter before semantic search.
- Combine dense, BM25 sparse or multimodal retrieval and fuse or rerank the candidate lists.
- Upsert changed entities and delete withdrawn documents while measuring how quickly changes become visible.
- Design multi-tenancy with databases, collections, partitions or partition keys instead of relying on unsafe string filters alone.
- Scale query and data paths independently while monitoring ingestion backlog, search latency and retrieval quality.
Production metrics should include Recall@k, P50/P95/P99 search latency, QPS, write throughput, index build duration, freshness lag, delete visibility, memory/disk use, recovery point and recovery time. Benchmarks must use representative dimensions, filters, top-k and concurrency. Vendor examples cannot validate a particular workload.
Pricing
Milvus software is free, but compute, object storage, disks, networking, Kubernetes, backups and operators are not. Zilliz Cloud is a separate service whose public offering currently includes Free, Standard, Enterprise and Business Critical plans, with Serverless, Dedicated and BYOC choices; costs vary by region, cloud, capacity and compute type.
| Option | Software/service price | Responsibility |
|---|---|---|
| Milvus Lite | Apache 2.0, free software | Local file and application lifecycle |
| Standalone or Distributed | Apache 2.0, free software | Deployment, upgrades, scaling, security and recovery |
| Zilliz Cloud Free/Serverless | Current allowance and consumption rates | Managed platform; user controls schema and usage |
| Zilliz Cloud Dedicated/enterprise | CU, storage, region or contract based | Higher isolation, SLA, networking and governance |
Do not present a Zilliz Cloud starting price as the Milvus license price. Compare total cost with staffing, redundancy, backups, transfer and idle capacity on the same basis.
Pros
- Apache 2.0 and LF AI & Data governance reduce dependence on one proprietary data API.
- A progression from local Lite to distributed deployment.
- Broad dense, sparse, full-text, filter and multi-vector retrieval support.
- Insert, update, delete and query operations suit continuously changing production corpora.
- A managed Zilliz path exists without erasing the project/service distinction.
Cons
- Distributed operation has meaningful upgrade, capacity and incident-response complexity.
- Index and search parameters create substantial recall, latency and cost tradeoffs.
- The database does not solve chunking, embedding quality, permission inheritance or answer hallucination.
- Results from Lite or Standalone cannot automatically predict distributed production behavior.
- Reindexing, re-embedding and large migrations need capacity and maintenance planning.
Alternatives
| Tool | Better for | Advantage | Tradeoff |
|---|---|---|---|
| RAGFlow | Document-to-answer workflows | More complete application workflow | Not the same general database layer |
| Dify | Low-code AI applications | App orchestration and knowledge UI | Less low-level index control |
| LightRAG | Experimenting with open RAG | Flexible application approach | Storage and operations still need assembly |
| Open WebUI | Self-hosted chat access | Convenient UI and model connection | Does not provide equivalent database operations |
FAQ
Are Milvus and Zilliz Cloud the same product?
No. Milvus is open source; Zilliz Cloud is a managed service operated by Milvus’s major contributor.
Does Milvus generate embeddings?
Its core role is storage and retrieval. Ecosystem helpers can call models, but model version, licensing and cost remain separate.
Can Milvus Lite run a large production service?
Lite is intended for local and lightweight use. Evaluate Standalone, Distributed or managed options for high concurrency and availability.
Should document updates use insert or upsert?
It depends on primary-key and version design. Stable keys with upsert, or delete-then-reinsert for changed chunks, are common; verify that stale vectors disappear.
How do I choose HNSW, IVF or DiskANN?
Benchmark recall, latency, memory, build time and update behavior on real data. No index wins every workload.
Is hybrid search always better than dense search?
No. Exact terms and identifiers often benefit from BM25 while conceptual questions benefit from dense retrieval. Tune fusion and reranking on labeled queries.
Bottom Line
Milvus is a capable open vector database with a broad deployment path, but good outcomes come from data lifecycle design, retrieval evaluation and reliable operations rather than a scale slogan. This page was verified on 2026-07-15 against the Milvus repository, official documentation and Zilliz Cloud pricing, with the project and managed service kept explicitly separate.