Skip to main content
NetApp Solutions

Vector Database

Contributors kevin-hoke nkarthik

This section covers the definition and use of a vector database in NetApp AI solutions.

Vector Database

A vector database is a specialized type of database designed to handle, index, and search unstructured data using embeddings from machine learning models. Instead of organizing data in a traditional tabular format, it arranges data as high-dimensional vectors, also known as vector embeddings. This unique structure allows the database to handle complex, multi-dimensional data more efficiently and accurately.

One of the key capabilities of a vector database is its use of generative AI to perform analytics. This includes similarity searches, where the database identifies data points that are like a given input, and anomaly detection, where it can spot data points that deviate significantly from the norm.

Furthermore, vector databases are well-suited to handle temporal data, or time-stamped data. This type of data provides information about ‘what’ happened and when it happened, in sequence and in relation to all other events within a given IT system. This ability to handle and analyze temporal data makes vector databases particularly useful for applications that require an understanding of events over time.

Advantages of vector database for ML and AI:

  • High-dimensional Search: Vector databases excel in managing and retrieving high-dimensional data, which is often generated in AI and ML applications.

  • Scalability: They can efficiently scale to handle large volumes of data, supporting the growth and expansion of AI and ML projects.

  • Flexibility: Vector databases offer a high degree of flexibility, allowing for the accommodation of diverse data types and structures.

  • Performance: They provide high-performance data management and retrieval, critical for the speed and efficiency of AI and ML operations.

  • Customizable Indexing: Vector databases offer customizable indexing options, enabling optimized data organization and retrieval based on specific needs.

Vector databases and use cases.

This section provides varies vector databases and their use case details.

Faiss and ScaNN

They are libraries that serve as crucial tools in the realm of vector search. These libraries provide functionality that is instrumental in managing and searching through vector data, making them invaluable resources in this specialized area of data management.


It’s a widely used search and analytics engine, has recently incorporated vector search capabilities. This new feature enhances its functionality, enabling it to handle and search through vector data more effectively.


It is a robust vector database with a unique set of features. It supports both dense and sparse vectors in its indexing functionality, which enhances its flexibility and adaptability. One of its key strengths lies in its ability to combine traditional search methods with AI-based dense vector search, creating a hybrid search approach that leverages the best of both worlds.

Primarily cloud-based, Pinecone is designed for machine learning applications and integrates well with a variety of platforms, including GCP, AWS, Open AI, GPT-3, GPT-3.5, GPT-4, Catgut Plus, Elasticsearch, Haystack, and more. It’s important to note that Pinecone is a closed-source platform and is available as a Software as a Service (SaaS) offering.

Given its advanced capabilities, Pinecone is particularly well-suited for the cybersecurity industry, where its high-dimensional search and hybrid search capabilities can be leveraged effectively to detect and respond to threats.


It’s a vector database that has a Core-API with four primary functions, one of which includes an in-memory document-vector store. It also utilizes the Face Transformers library to vectorize documents, enhancing its functionality and versatility.
Chroma is designed to operate both in the cloud and on-premises, offering flexibility based on user needs. Particularly, it excels in audio-related applications, making it an excellent choice for audio-based search engines, music recommendation systems, and other audio-related use cases.


It’s a versatile vector database that allows users to vectorize their content using either its built-in modules or custom modules, providing flexibility based on specific needs. It offers both fully managed and self-hosted solutions, catering to a variety of deployment preferences.

One of Weaviate’s key features is its ability to store both vectors and objects, enhancing its data handling capabilities. It is widely used for a range of applications, including semantic search and data classification in ERP systems. In the e-commerce sector, it powers search and recommendation engines. Weaviate is also used for image search, anomaly detection, automated data harmonization, and cybersecurity threat analysis, showcasing its versatility across multiple domains.


Redis is a high-performing vector database known for its fast in-memory storage, offering low latency for read-write operations. This makes it an excellent choice for recommendation systems, search engines, and data analytics applications that require quick data access.

Redis supports various data structures for vectors, including lists, sets, and sorted sets. It also provides vector operations such as calculating distances between vectors or finding intersections and unions. These features are particularly useful for similarity search, clustering, and content-based recommendation systems.

In terms of scalability and availability, Redis excels in handling high throughput workloads and offers data replication. It also integrates well with other data types, including traditional relational databases (RDBMS).
Redis includes a Publish/Subscribe (Pub/Sub) feature for real-time updates, which is beneficial for managing real-time vectors. Moreover, Redis is lightweight and simple to use, making it a user-friendly solution for managing vector data.


It’s a versatile vector database that offers an API like a document store, much like MongoDB. It stands out due to its support for a wide variety of data types, making it a popular choice in the data science and machine learning fields.

One of Milvus’ unique features is its multi-vectorization capability, which allows users to specify at runtime the type of vector to use for the search. Furthermore, it utilizes Knowwhere, a library that sits atop other libraries like Faiss, to manage communication between queries and the vector search algorithms.

Milvus also offers seamless integration with machine learning workflows, thanks to its compatibility with PyTorch and TensorFlow. This makes it an excellent tool for a range of applications, including e-commerce, image and video analysis, object recognition, image similarity search, and content-based image retrieval. In the realm of natural language processing, Milvus is used for document clustering, semantic search, and question-answering systems.

For this solution, we picked milvus for the solution validation. For performance, we used both milvus and postgres(

Why we chose milvus for this solution?

  • Open-Source: Milvus is an open-source vector database, encouraging community-driven development and improvements.

  • AI Integration: It leverages embedding similarity search and AI applications to enhance vector database functionality.

  • Large Volume Handling: Milvus has the capacity to store, index, and manage over a billion embedding vectors generated by Deep Neural Networks (DNN) and Machine Learning (ML) models.

  • User-Friendly: It is easy to use, with setup taking less than a minute. Milvus also offers SDKs for different programming languages.

  • Speed: It offers blazing fast retrieval speeds, up to 10 times faster than some alternatives.

  • Scalability and Availability: Milvus is highly scalable, with options to scale up and out as needed.

  • Feature-Rich: It supports different data types, attribute filtering, User-Defined Function (UDF) support, configurable consistency levels, and travel time, making it a versatile tool for various applications.

Milvus architecture overview

Error: Missing Graphic Image

This section provides higher lever components and services are used in Milvus architecture.
* Access layer – It’s composed of a group of stateless proxies and serves as the front layer of the system and endpoint to users.
* Coordinator service – it assigns the tasks to the worker nodes and act as a system's brain. It has three coordinator types: root coord,data coord and query coord.
* Worker nodes : It follows the instruction from coordinator service and execute user triggered DML/DDL has three types of worker nodes such as query node, data node and index node.
* Storage: it’s responsible for data persistence. It comprises meta storage, log broker, and object storage. NetApp storage such as ONTAP and StorageGRID provides object storage and File based storage to Milvus for both customer data and vector database data.