What Milvus does
Milvus is an open-source vector database for approximate nearest-neighbor, filtered, sparse, dense, multimodal, and hybrid search, with managed Zilliz Cloud available.
Milvus separates vector search infrastructure from the model that creates vectors. It supports dense and sparse vectors, scalar fields, metadata filtering, multiple index types, hybrid ranking, partitions, replication, and distributed scaling. Developers can run lightweight local variants, self-host a cluster, or use the related Zilliz Cloud managed service.
The open-source license removes a software fee but not compute, memory, storage, networking, backup, observability, upgrades, and engineering work. Managed pricing depends on capacity or serverless usage, storage, reads, writes, regions, backups, and support. Benchmark the actual vector count, dimension, filter selectivity, recall target, ingest rate, query concurrency, and growth rather than relying on a small demo.
A vector database does not determine whether an embedding is accurate or a result is authorized. Deleted source material can survive in old vectors or backups, and metadata filters can leak tenants if applied inconsistently. Encrypt data, isolate credentials, enforce permissions at every query, version embeddings and indexes, test backup restoration, and monitor recall, latency, stale records, and compaction.
How Milvus works
Applications transform approved text, images, audio, or other objects into vectors with an external embedding model, attach scalar metadata and permissions, and insert them into a Milvus collection. Milvus builds selected indexes, searches nearest vectors, applies filters or hybrid ranking, and returns identifiers and scores. The application fetches source content and may pass it to a model. Operators own schema, sharding, index tuning, backups, upgrades, ACL enforcement, embedding versions, and relevance review.
How to set up Milvus
Choose the operating model
Compare local, self-hosted, and managed deployment for scale, regions, networking, skills, recovery, support, and total cost.
Design schema and permissions
Define vector fields, dimensions, source IDs, versions, tenant ACLs, filters, retention, and deletion before ingesting data.
Benchmark indexes
Test representative corpus size, recall, filter selectivity, latency, memory, build time, ingest, concurrency, and failure recovery.
Build versioned ingestion
Pin embedding models and chunking, deduplicate sources, make writes idempotent, validate counts, and support staged re-indexing.
Operate and evaluate
Configure encryption, backups, alerts, quotas, compaction, upgrades, and disaster recovery while tracking retrieval quality and stale data.
Milvus FAQs
Is Milvus an embedding model?
No. It stores and searches vectors created by an external model or application.
Is Milvus free?
The open-source software is available without a license fee; hosting and operations still cost money. Managed Zilliz Cloud is commercial.
What is hybrid search?
It combines multiple retrieval signals, such as dense semantic vectors, sparse lexical vectors, and metadata filters or reranking.
Can Milvus enforce application permissions automatically?
Schema and platform controls help, but the application must consistently attach and apply tenant and document-level authorization filters.
Do higher search scores mean true answers?
No. Scores indicate similarity under the configured model and index, not factual truth or calibrated confidence.
Listing reviewed 2026-07-30. Product details and pricing can change; verify important terms on the provider's website.
Related Coding AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.