What Is Milvus? A Distributed Vector Database

As businesses build more AI-powered semantic search applications that return results based on similarity of meaning or appearance, they’re turning to specialized vector databases.

The Milvus vector database (as well as the likes of Chroma, Pinecone, Qdrant, and Weaviate) excels at finding approximate nearest neighbors (ANN) matches in large unstructured data sets, based on relationships in meaning. The databases suit applications such as ecommerce and media streaming product recommendations, natural language semantic search, and retrieval-augmented generation (RAG).

Milvus and other vector databases store embeddings, or positions in mathematical space, of potentially billions of numerical vectors calculated from data sets. Semantic search engines built with them can quickly return highly relevant results based on what a user intended to find. Read on to learn how Milvus stacks up against the competition.

What Is Milvus?

Milvus, owned by Chinese company Zilliz, is a high performance vector database for conducting ANN searches across data sets containing up to tens of billions of vectors. Developers can mix a subset of database filters with vector similarity searches in the same operation. Milvus maintains its performance when multiple users simultaneously run these combined searches. The database uses sharding and partitioning for horizontal scalability and takes advantage of GPU acceleration.

Use cases for Milvus include fast, real-time data retrieval from large, unstructured data sets. Applications include recommendation systems, as well as similarity searches of audio and images in industry and entertainment, similarity searches of molecules for drug development, and such searches of time-stamped vectors for financial forecasting and fraud detection. Users can tune the database to strike a balance between query performance and how fresh the data is.

Milvus is available under the Apache 2.0 open source license. Companies can run it on a single server, in cloud native Kubernetes clusters, or in the managed Zilliz Cloud. There’s also a version that runs on laptops for building prototypes.

Oracle Database Versus Milvus for RAG Queries on Fast-Changing Data

Oracle Database 23ai’s AI Vector Search lets it store semantic content from unstructured data as native vector data, as Milvus does, so users can quickly run AI-powered vector similarity searches within their business database instead of having to move that data to a separate vector database. Standalone vector databases such as Milvus are optimized for efficient vector similarity searches, but they aren’t designed for traditional, ACID-compliant synchronization with other databases. That means they need to rely on batch updates or streaming to refresh their embeddings, which can lead to inconsistent data when working with fast-changing information, such as securities prices. The need for specialized vector stores to integrate multiple systems can slow performance for such real-time RAG workloads.

Development teams can instead choose a multipurpose database, such as Oracle Database 23ai, which stores relational, graph, spatial, JSON, time-series, and other types of data, in addition to vector embeddings. That’s advantageous when designing applications such as recommendation systems that look for cues from products bought by similar customers in a short window before a consumer closes a purchase. With Oracle, there’s no API call needed from an online checkout system to the vector database, supporting reduced latency.

For scalability, Oracle Globally Distributed Database supports sharding for vector searches, routing queries and update, insert, and delete operations to the shards. The Oracle database can also determine the best order for running a vector search and applying additional relational filters, often stored separately, to narrow down ANN search results. Milvus offers a more limited set of filters.

Milvus FAQs

What are Milvus’s core features?

The database is designed for high performance vector search across data sets containing billions or as many as trillions of vectors. It uses a sharding storage architecture for horizontal scalability and lets different types of computing tasks (such as storage, workload coordination, and query execution) scale independently of one another. The software can also scale up using a large web server.

How does Milvus handle large data sets and unstructured data?

Milvus’s high performance vector database can efficiently search tens of billions or even trillions of vectors in embeddings computed from unstructured data sets. It’s able to store and retrieve embeddings derived from large sets of natural language, photos, video, molecular structures, and other data.

What are the common use cases for Milvus?

Businesses are deploying Milvus for ecommerce and media streaming recommender systems, multimedia similarity search, RAG, drug development, and financial fraud detection.

Can Milvus be deployed in a distributed environment?

Zilliz’s Milvus Distributed deployment option lets customers run the database on Kubernetes clusters. The company also offers Milvus Standalone for running on a single server.

How does Milvus compare to traditional databases?

Instead of searching unstructured data with exact terms entered by users, Milvus’s search engine is designed for speed when searching for semantically similar concepts using techniques including ANN.