Best Vector Databases in 2023

Jonathan Yue, PhD
10 min readAug 29, 2023

Vector databases have emerged as a prevailing trend in 2023 and possibly 2024, with their significance particularly pronounced in AI applications. This article takes a closer look at the landscape of vector databases, emphasizing their horizontal scalability and robust support for diverse data types, and introduces readers to a few vector databases, offering valuable insights to empower the development of AI-driven applications.

In an era where data-driven insights fuel innovation, vector databases have swiftly gained prominence due to their prowess in handling high-dimensional data and facilitating complex similarity searches. The focal points of this article are the dual pillars of scalability and versatility. Horizontal scalability, which ensures seamless expansion as data volumes surge, is examined closely. Moreover, the article delves into the databases’ capacity to accommodate an array of data types, spanning from structured data to unstructured content like images, audio, and more.

By studying these attributes, the article extends a guiding hand to readers navigating the rapidly evolving landscape of vector databases. Whether for recommendation systems, image recognition, natural language processing, financial forecasting, or other AI-driven ventures, understanding the top vector databases is important. Through this introductory overview, readers are poised to harness the synergy of vector databases and AI methodologies, empowering them to craft innovative applications that thrive on efficiency, scalability, and data diversity.


JaguarDB vector database is not just a vector database; it is a comprehensive solution that goes beyond vector data management. While it excels at handling vector data, it also processes non-vector data within a fully integrated framework.

JaguarDB stores high-dimensional vectors with the state-of-the-art HNSW graph index store. HNSW, short for Hierarchical Navigable Small World, is a data structure and algorithm used for approximate nearest neighbor search in high-dimensional spaces. It is designed to efficiently find data points that are close to a given query point in a high-dimensional space, without exhaustively searching through all data points. HNSW creates a hierarchical structure of data points that forms a graph. Each level of the hierarchy is a different graph that represents the data points at different levels of detail. HNSW maintains a “small world” property, which means that even though the graph is not fully connected like a traditional graph, it is still possible to navigate from one node to another through a relatively small number of edges. HNSW constructs the hierarchical graph in a way that ensures data points are connected to nearby points, enabling efficient traversal of the graph to find approximate nearest neighbors.

When searching for the nearest neighbors of a query point, HNSW uses the hierarchical structure to quickly navigate through the graph, starting from coarse levels and refining the search as it descends deeper into the hierarchy. HNSW focuses on approximate search rather than exact search. It sacrifices perfect accuracy for improved search efficiency, which is valuable when dealing with high-dimensional data.

JaguarDB brings forth a capability that enables users to engage in KNN similarity searches using a wide spectrum of distance metrics, which include fundamental measures like Euclidean and Manhattan distances, along with specialized metrics such as cosine, Jaccard, Hamming, InnerProduct, and Minkowski distances. This diversity empowers users to tailor their similarity searches based on the specific characteristics of their data and the intricacies of their analytical requirements. Whether dealing with spatial relationships, binary patterns, or various dimensions of data, JaguarDB accommodates a versatile selection of distance metrics to ensure proper similarity computation.

JaguarDB offers the flexibility in hybrid search, or multimodal search, that combines multiple types of search techniques or data representations to optimize the search process for different types of queries. This approach is particularly useful in applications that handle heterogeneous data, where the data may include both vector-based embeddings and traditional structured or unstructured data.

The ZeroMove technique is a pivotal feature within JaguarDB, delivering a transformative solution. Diverging from conventional consistent hashing algorithms that demand data migration during system expansion, ZeroMove empowers seamless scalability devoid of the necessity to transfer data across machines. With the ZeroMove method, new nodes are introduced to the system by clusters — multiple storage and compute nodes are commissioned at once. ZeroMove technology aligns particularly well with vector index stores, as the process of removing vectors from such stores might necessitate a comprehensive index reconstruction when data is relocated.

Time series data in JaguarDB refers to a sequence of data points that are ordered based on time intervals. Users of JaguarDB can collect data over successive time periods and let it automatically aggregate data over multiple time windows for real-time analysis, prediction, and decision-making in AI applications. Time series data often come from mobile targets, sensors, devices, financial markets, weather stations, social media, and more.

Geospatial data support of JaguarDB is instrumental in applications such as environmental monitoring, disaster response, agriculture, and natural resource management. Geospatial data often exhibits in the form of vectors such as ling strings and polygons. AI models equipped with geospatial insights can predict the spread of wildfires, monitor deforestation, optimize irrigation strategies, and assess the impact of climate change. These applications hinge on the AI system’s ability to process and analyze geospatial data, allowing for timely and informed interventions.

JaguarDB provides data lake capability which represents a powerful feature that integrates storage capabilities directly into the database system, offering a seamless and unified solution for managing, analyzing, and retrieving both structured and unstructured data. This integration brings efficiency, flexibility, and scalability to the storage and processing of diverse data types within the JaguarDB environment.

Fault tolerance is of paramount significance in vector databases used in AI applications due to its role in ensuring system reliability, availability, and consistent performance, even in the face of unexpected errors or failures. This is particularly crucial in AI applications where accurate and timely data retrieval and processing are essential. JaguarDB offers tolerance of machine failures and network disconnections to ensure high availability of the vector database system.

Vector data can be replicated with multiple copies, a maximum of three, in JaguarDB to ensure data availability, reliability, and fault tolerance. It creates and maintains duplicate copies of data across multiple nodes. This redundancy is crucial for the effectiveness of AI systems.


KDB+ ( “KDB plus”) is a high-performance, column-oriented database management system (DBMS) and programming language developed by Kx Systems. KDB+ is known for its ability to efficiently handle and analyze large volumes of time-series and historical data. It is widely used in the financial industry, including for trading and risk management applications, due to its speed, scalability, and specialized features for handling time-series data.

Kx Systems, the company behind KDB+, was founded in 1993 by Arthur Whitney and Janet Lustgarten. They developed KDB+ and its associated programming language, q, to address the specific needs of financial institutions that required fast and powerful data analysis tools for processing massive amounts of market data. KDB scales out its cluster by adding RDB and HDB components to different server hosts.

KDB.AI is a knowledge-based vector database and search engine that allows developers to build scalable, reliable, and real-time applications by providing advanced search, recommendation, and personalization for AI applications. KDB.AI, launched in 2023, is designed for cloud-native vector data management, vector embeddings, and engineering tasks, as seen in GPT-style query exploration.

The KX stack streamlines traditional data science workflows. By bundling advanced connectivity, vector encodings, built-in algorithms, and powerful data organizational tools — all with support for Python, Java, SQL — the KX stack squishes the stack, reducing dependencies, accelerating processing, and simplifying data exploration. To deploy KDB.AI, the initial step is to ingest data from external databases, ETL, or streaming data sources through native connectors in the platform. This data is then aggregated, summarized, and cleansed for preparation and organization by the data store.

KDB’s built-in algorithms encode vector embeddings into the database, enhancing storage efficiency for accelerated query support and real-time operations. The KDB store provision of CRUD operations for vector data are further augmented by GPT, neural net-style algorithms, and other capabilities. The Consume phase supports Python, Java, SQL, and Q queries, while accommodating new prompt interfaces. These interfaces enable interactive questions on vector embeddings, handling queries from a simple command prompt.


Pinecone provides long-term memory for high-performance AI applications. As a managed, cloud-native vector database, it offers a cloud-based simple API and solution. Pinecone serves fresh, filtered query results with low latency at scale. Pinecone is a service offering a vector database for similarity search and machine learning applications. It is a platform for creating, indexing, and querying high-dimensional vector embeddings, enabling efficient similarity searches and nearest neighbor retrieval.

Pinecone is designed to handle high-dimensional vector data. It allows users to index vectors generated from sources such as images, text, audio, or other data that can be represented as vectors. Pinecone enables users to conduct similarity search, where given a query vector, it retrieves the most similar vectors from the indexed dataset. This is particularly useful in applications like recommendation systems, image search, natural language processing, and more. Pinecone is built to scale horizontally, allowing users to handle large datasets and high query loads. It handles the distribution and load balancing of data and queries across multiple nodes.

There are two approaches to scaling Pinecone indexes. If you need to scale your environment to accommodate more vectors, you can modify your existing index to scale it vertically or create a new index and scale horizontally. Scaling vertically is relatively fast and involves no downtime. This is a good choice when users cannot pause upserts and must continue serving traffic. It also allows you to double your capacity instantly. Two methods can be adopted to horizontal scaling in Pinecone: adding pods and adding replicas. Adding pods increases all resources but requires a pause in upserts; adding replicas only increases throughput and requires no pause in upserts. Adding additional replicas increases the throughput of the index but not its capacity.

Pinecone is optimized for real-time retrieval, making it suitable for applications that require fast response times, such as interactive user experiences and real-time analytics. Pinecone offers a simple API that can index vectors, perform similarity searches, and manage the index. This API-driven approach simplifies integration into existing applications. While Pinecone itself is focused on indexing and retrieval, users need to generate vector embeddings using external tools or models. Common tools include machine learning frameworks and pre-trained models. Pinecone can be integrated into machine learning pipelines to enhance model inference and predictions using similarity search capabilities. Pinecone is a cloud-based managed vector database designed for businesses and organizations to build and deploy machine learning applications. Pinecone uses closed-source code.

The Pinecone vector database has a simple and intuitive interface, which makes it developer-friendly. It hides the complexity of managing the underlying infrastructure, allowing developers to put their focus on building applications. Its support for high-dimensional vector databases makes Pinecone suitable for various use cases, including similarity search, recommendation systems, personalization, and semantic search. It also supports single-stage filtering capability. Its ability to analyze data in real time also makes it useful for threat detection and monitoring against cyberattacks in the cybersecurity industry.


Milvus is an open-source vector database management system designed to efficiently store, index, and retrieve large-scale vector data. It is specifically optimized for handling high-dimensional vector embeddings generated from various types of data, such as images, audio, text, and more. Milvus is built to enable similarity search and nearest neighbor retrieval, making it useful for applications in machine learning, computer vision, and natural language processing.

Milvus enables indexing and searching high-dimensional vector data. It provides indexing algorithms and methods to conduct similarity searches. It distributes data and queries across multiple nodes. It uses Kafka/Pulsar, query nodes, and data nodes to process data flow. Milvus is optimized for real-time retrieval, making it suitable for applications that require low-latency responses, such as interactive search and recommendation systems. Milvus supports multiple types of vector data, including float and binary data, allowing you to store a range of embeddings. The system provides tools for fine-tuning similarity searches, enabling users to achieve the desired trade-off between accuracy and speed.

Milvus adopts a shared-storage architecture featuring storage and computing disaggregation and horizontal scalability for its computing nodes. Milvus comprises four layers: access layer, coordinator service, worker node, and storage. Etcd is used as metadata engine, and MinIO is responsible for data storage. Milvus cluster consists of these components: Root coord, Proxy, Query coord, Query node, Index coord, Index node, Data coord, and Data node.

Milvus offers a set of APIs that enables user to interact with the database, including operations for indexing vectors, performing similarity searches, managing collections, and more. Milvus can be integrated into machine learning workflows to enhance model inference and predictions using its efficient similarity search capabilities. Milvus is an open-source project, which means its code is available for use, modification, and contribution by the community.


Weaviate is an open-source vector database. It allows users to store data objects and vector embeddings from ML-models, and scale into many data objects. User can generate their own vectors or use one of the vectorization modules. Users can combine multiple search techniques, such as keyword-based and vector search, to provide new search experiences. Users can pipe them through LLM models like GPT-3 to create advanced search experiences.

Weaviate works as decentralized, and cloud-native knowledge graph and contextual search engine, developed by its original company named SeMI Technologies. Weaviate is designed to store, search, and connect data in a semantic manner, enabling searches that go beyond simple keyword matching. It is particularly well-suited for applications involving complex and interconnected data, such as natural language processing, recommendation systems, and data enrichment.

Weaviate operates as a knowledge graph, storing data in a way that captures relationships and context between entities. This graph-based structure allows for meaningful searches. Weaviate specializes in similarity searches based on vector embeddings. It can handle high-dimensional data like embeddings generated by machine learning models and supports various distance metrics for similarity calculations. Weaviate’s search capabilities are designed to understand the context of the data, enabling semantic search that considers the meaning and relationships between entities. Weaviate allows users to define a schema that describes the types of data, their properties, and their relationships. This schema-driven approach helps ensure data consistency and relevance. Weaviate provides a GraphQL-based API that allows you to interact with the system, perform searches, manage data, and define schemas. Weaviate supports data enrichment by connecting external sources to augment your data with additional information.

Weaviate uses a virtual shard system to assign objects to shards. This makes it more efficient to re-shard, as minimal data movement occurs when re-sharding. However, due to the HNSW index, resharding is still a very costly process and should be used rarely. The cost of re-sharding is roughly that of an initial import with regards to the amount of data that needs to be moved.


This article has undertaken an exploration of various vector databases, focusing specifically on their horizontal scalability and the spectrum of data types they support. Should any inaccuracies or corrections be required, we encourage readers to reach out to the author directly for rectification of errors or clarifications regarding any potential misunderstandings conveyed within the content. Your engagement is vital in ensuring the accuracy and comprehensiveness of the information presented here.



Jonathan Yue, PhD

Enthusiast on vector databases, AI, RAG, data science, consensus algorithms, distributed systems. Initiator and developer of the JaguarDB vector database