Published at Aug 29, 2024

Vector Store vs. Vector Database: Understanding the Connection

Written by Anya Sage

Vector embeddings—numerical representations of words, phrases, or other data in a high-dimensional space—are a critical component of semantic search and AI systems. They allow machines to capture semantic meaning by encoding relationships and similarities between concepts. Yet embedded vectors are a unique data type that requires special handling due to their high-dimensional nature. To address this need, two related data storage systems have emerged: vector stores and vector databases. 

The terms “vector store” and “vector database” are often used interchangeably, so parsing the exact connection between them can be hard. But it’s a connection that’s important to understand because it sheds light on the nature of vector data storage/retrieval and the technical details of building vector data systems. 

To explain the vector store vs. vector database connection, we first define vector stores and vector databases. Then, we examine the relationship between them and the resulting technical complexities. Finally, we consider what to look for when evaluating vector databases for your projects and show how Timescale's vector database, anchored on our mature PostgreSQL cloud platform, meets these assessment criteria. Let's dive in.

What's the Difference Between a Store and a Database?

Before we zoom into the specifics, let’s define vector stores and vector databases. Both tools are designed for storing and searching embedded vectors. However, there are subtle differences between them that outline their relationship and functionality.

What is a vector store?

A vector store is a specialized system designed for holding embedded vectors. Due to the unique properties of vector embeddings, vector stores require specific design considerations that set them apart from traditional data storage systems.

Vector embeddings are high-dimensional numerical representations of data often used in machine learning and natural language processing tasks. An embedding is a compact representation of raw data, such as an image or text, transformed into a vector comprising floating-point numbers. It’s a powerful way of representing data according to its underlying meaning.

Vector embeddings work by representing features or objects as points in a multidimensional vector space, where the relative positions of these points represent meaningful relationships between the features or objects.

The key characteristics of vector stores include:

1. Optimization for high-dimensional data: Vector embeddings typically consist of hundreds or thousands of dimensions, which pose unique challenges for storage and retrieval.

2. Specialized retrieval algorithms: Unlike traditional databases that use exact matching queries, vector stores employ nearest-neighbor searches with specific distance metrics. These algorithms, such as those found in the scikit-learn library, are designed to find the most similar vectors based on their numerical properties. As scikit-learn explains: “The principle behind nearest neighbor methods is to find a predefined number of training samples closest in distance to the new point, and predict the label from these.” 

3. Efficiency focus: Traditional databases are often inefficient when dealing with vector data. Vector stores are built from the ground up to efficiently handle the storage and retrieval of high-dimensional vectors.

4. Limited data type flexibility: To optimize performance, vector stores typically focus on supporting high-dimensional numerical data, sacrificing the versatility (of handling various data types) found in general-purpose databases.

5. Streamlined schema designs: Compared to general-purpose databases, vector stores often have less flexible schema designs, prioritizing structures that are optimized for vector data.

6. Specialized query support: Instead of supporting a wide range of query types, vector stores are optimized for nearest neighbor retrieval, which is the primary operation performed on vector data.

What is a vector database?

A vector database, on the other hand, is a more comprehensive system that incorporates the capabilities of a vector store while providing additional features and functionality. Here are the key characteristics of a vector database:

1. Extended database functionality: vector databases are often built as extensions of existing database systems, adding vector storage and retrieval capabilities to proven database technologies.

2. Integration of vector and relational data: these systems connect stored vectors to the robust, complex query systems and structured data typically found in relational databases.

3. Broader query support: vector databases allow for more complex queries that can combine vector similarity searches with traditional database operations.

4. Flexible data model: unlike pure vector stores, vector databases can often handle a mix of vector and non-vector data types, providing greater versatility for complex applications.

5. Advanced indexing and optimization: many vector databases incorporate advanced indexing techniques to improve the performance of both vector and non-vector queries.

How Vector Stores and Vector Databases Are Related

To understand the relationship between vector stores and vector databases, think of them as interconnected components within a larger system. Most available vector storage and retrieval systems are actually vector databases that contain a vector store as a core component.

The relationship can be described as follows:

• Vector store as a subsystem: the vector database has a vector store contained within it, serving as the specialized component for efficiently holding and searching vector data.

• Database wrapper: the larger database system acts as a wrapper around the vector store, providing additional functionality and integration capabilities.

• Connecting systems: the database layer includes systems that allow database queries to interact with the store's retrieval functions, bridging the gap between traditional database operations and vector-specific functionality.

A prime example of this relationship can be seen in the pgvector extension for PostgreSQL (enabling open-source vector similarity search for PostgreSQL so you can store your vectors with the rest of your data). This extension introduces tools for creating tables optimized for vector data storage and provides function calls that perform different types of nearest neighbor searches on vector tables.

In this case:

• The vector tables created using pgvector serve as the vector store component.

• The larger PostgreSQL database, with the pgvector extension, becomes a full-fledged vector database.

• SQL queries can incorporate pgvector functions, seamlessly integrating vector operations (the vector store subsystem) with traditional database queries (the larger database system design).

The pgvector extension allows developers to leverage the power of vector similarity searches while still benefiting from the robust features and familiarity of a traditional relational database system.

What to Look for in a Vector Store and Vector Database

When evaluating vector databases for your projects, there are key factors to consider. These factors help ensure that you choose a database that not only performs well but also integrates smoothly with your existing infrastructure and workflows.

Well-optimized vector store

Adding high-dimensional schema support and nearest-neighbor search capabilities to a database isn't an extremely complicated project, yet optimizing these features for production use is a significant challenge. A production-ready vector database should have a store component with the following characteristics:

1. Efficient, fast storage: The system should be able to quickly insert, update, and delete vector data, even when dealing with large datasets.

2. State-of-the-art nearest neighbor algorithms: Optimizing nearest neighbor search is an active field in algorithm research. The best systems stay at the cutting edge of these developments and implementations, continuously improving their performance.

3. Scalability: The vector store should be able to handle growing datasets without significant performance degradation.

4. Memory efficiency: Given the high-dimensional nature of vector data, efficient memory usage is needed to maintain performance as data volumes increase.

5. Support for multiple distance metrics: Different applications may require different similarity measures, so a versatile vector store should support various distance metrics (for example, Euclidean, cosine, dot product).

Clean connection with the database

While vector store efficiency and speed are critical, it's equally important that this component integrates smoothly with the broader database system. Some considerations in this area include:

1. Intuitive syntax: Vector stores designed solely with optimization in mind can sometimes come with clunky or unusual syntax. Look for systems that offer a clean, intuitive interface for vector operations.

2. Compatibility with database features: The vector store should work well with broader database tools like indexing, transactions, and backup systems.

3. Query integration: It should be easy to combine vector similarity searches with traditional database queries, allowing for complex operations that leverage both vector and non-vector data.

4. Consistent data types: The system should provide a seamless way to work with vector data types alongside standard database types.

5. Performance optimization: Look for systems that can optimize queries involving both vector and non-vector operations.

Familiar and robust database system

Some vector database products offer excellent vector store performance and database integration (the two properties mentioned above), but they may be built from the ground up as entirely new systems. This specificity can introduce a significant learning curve and potential challenges:

1. Learning new tools: Adopting a completely new database system can be time-consuming and costly, especially when it's primarily to handle one specific data type.

2. Integration challenges: New systems may not easily integrate with existing tools and workflows in your organization.

3. Limited community support: Newer, specialized systems might have smaller user communities, making it harder to find solutions to problems or best practices.

4. Uncertain long-term support: There's always a risk that a new, specialized system might not receive long-term support or updates.

An ideal vector database builds on existing, well-supported database systems to mitigate these risks. This approach offers several advantages:

• Shorter learning curve: Developers can leverage their existing knowledge of familiar database systems.

• Robust ecosystem: Established databases often have a wide range of tools, extensions, and integrations available.

• Large community: Popular database systems have large, active communities that can provide support and share knowledge.

• Long-term stability: Well-established database systems are more likely to receive ongoing updates, security patches, and feature improvements.

• Easier talent acquisition: It's typically easier to find developers experienced with popular database systems than those familiar with highly specialized new tools.

The vector store vs. vector database decision is a paradox of choice captured well in Making Postgres a Better AI Database: 

As it turns out, evolving to meet changing developer needs is exactly what PostgreSQL has done since its inception, thanks to its rich extension ecosystem and community. So, with the specialized extensions as explained below, using Postgres for AI delivers the best of both worlds—better vector storage and retrieval performance without losing time-tested reliability and convenience. 

What Timescale Offers: PostgreSQL as a High-Performance Vector Database

Timescale's vector database system lives in Timescale Cloud, which enables developers to build production AI applications at scale with PostgreSQL.

With Timescale Cloud, developers can access pgvector, pgvectorscale, and pgai—extensions that turn PostgreSQL into an easy-to-use and high-performance vector database, plus a fully managed cloud database experience. It is designed to meet the demanding requirements of modern vector data applications while building on the strengths of established database technology—something that developers in companies large and small are excited about. 

But how does Timescale address the vector database selection criteria discussed above? Let’s examine that. 

Production-level vector store performance

Timescale Cloud offers an open-source PostgreSQL stack for AI applications. With Timescale Cloud, developers can access pgvector, pgvectorscale, and pgai—extensions that turn PostgreSQL into an easy-to-use and high-performance vector database, plus a fully managed cloud database experience. These extensions make Postgres the de facto database for building AI applications because they: 

• Eliminate the need to use a standalone vector database in your AI data stack

• Lower the barriers for adopting and scaling PostgreSQL for your AI applications

• Empower you to easily build and scale RAG, search, and agents applications

Here’s a quick overview of each extension: 

• Pgvector is the popular open-source extension for vector data in PostgreSQL, enabling open-source vector similarity search for PostgreSQL.

• Pgai brings more AI workflows—like embedding creation and model completion—to PostgreSQL, making it easier for developers to build search and retrieval augmented generation (RAG) applications.

• Pgvectorscale builds on pgvector to enable development of more scalable AI applications, with higher-performance embedding search and cost-efficient storage.

Our benchmark test (using a dataset of 50 million Cohere embeddings of 768 dimensions each) compared the performance of PostgreSQL with pgvector and pgvectorscale against Pinecone, widely regarded as the market leader for specialized vector databases. The results showed that using PostgreSQL with pgvector and pgvectorscale dismantles the argument of “greater performance” often made to justify choosing a dedicated vector database. 

Compared to Pinecone’s storage-optimized index (s1), PostgreSQL with pgvector and pgvectorscale achieves 28x lower p95 latency and 16x higher query throughput for approximate nearest neighbor queries at 99 % recall.

PostgreSQL with pgvector and pgvectorscale extensions outperformed Pinecone’s s1 pod-based index type, offering 28x lower p95 latency.

The benchmark test also revealed compelling cost benefits.

Self-hosting PostgreSQL with pgvector and pgvectorscale offers better performance while being 75-79 % cheaper than using Pinecone.

PostgreSQL on Timescale Cloud has additional unique capabilities for handling vector data at scale. Its production-level performance features include:

1. Advanced indexing algorithms: Pgvectorscale’s Streaming DiskANN overcomes limitations of in-memory indexes like HNSW by storing part of the index on disk, making it more cost-efficient to run and scale as vector workloads grow. Pgvectorscale’s Streaming DiskANN includes support for Statistical Binary Quantization (SBQ), a novel binary quantization method (developed by researchers at Timescale) that improves accuracy over traditional methods of quantization. 

2. Efficient storage: Hybrid time-based vector search is optimized, leveraging the automatic time-based partitioning and indexing of Timescale’s hypertables to efficiently find recent embeddings, constrain vector search by a time range or document age, and store and retrieve LLM response and chat history with ease.

3. Simplified AI stack: PostgreSQL on Timescale Cloud provides a single place for vector embeddings, relational data, time-series data, and event data that powers next-generation AI applications.

4. Support for streaming filtering: pgvectorscale supports streaming filtering which allows for accurate retrieval even when secondary filters are applied during similarity search.

Built on Timescale's PostgreSQL foundation

A major Timescale advantage is its foundation in PostgreSQL, one of the most popular and well-supported open-source databases. PostgreSQL is emerging as the de facto database standard; as noted by Timescale CEO and co-founder Ajay Kulkarni in Why PostgreSQL Is the Bedrock for the Future of Data, “PostgreSQL for Everything” has become a “growing war cry among developers.” It’s a war cry for stack simplification at a time when specialized database proliferation has led to overly complex data pipelines.

Timescale Cloud for AI and vector data works with everything in your AI stack. Source

Timescale Cloud’s foundation in PostgreSQL provides several advantages:

1. Proven reliability: PostgreSQL has been battle-hardened by production use for over three decades. Timescale Cloud inherits that reliability.

2. Rich ecosystem: Users can leverage the vast array of tools, extensions, and integrations available for PostgreSQL, which have effectively turned PostgreSQL into a full-fledged platform.

3. SQL support: Timescale Cloud allows users to combine vector similarity searches with standard SQL queries, providing powerful data manipulation capabilities.

4. Built-for-PostgreSQL scalability: Timescale’s tiered storage architecture makes PostgreSQL big-data-ready by leveraging the flexibility of PostgreSQL and hypertables for effective data management. With tiered storage, you can automatically tier your data between disk and object storage (S3), effectively creating the ability to have an infinite table. 

5. Advanced features: Alongside vector operations, users can take advantage of PostgreSQL's features like full-text search, window functions, JSON support, and capabilities like streaming replication, hot standby, and in-place upgrade.

Robust support and functionality

As a high-performance, developer-focused cloud platform, Timescale Cloud provides PostgreSQL services for the most demanding workloads—whether AI or time-series, analytics, and event workloads. It is ideal for production applications and provides a worry-free and easy development experience, with programmatic APIs, one-click database forking, high availability (HA), read replication, seamless upgrades, and expert support—plus robust security and privacy functionality. Timescale Cloud benefits from PostgreSQL’s rock-solid foundations:  

1. Community support: users can tap into the vast knowledge base of the PostgreSQL community, which continues to make the core better and is witnessing more companies contributing, including the hyperscalers. 

2. Timescale-specific support: Timescale provides additional support through its own community and team, focusing on vector-specific features and optimizations.

3. Regular updates: as part of the larger Timescale ecosystem, Timescale Cloud receives regular updates and improvements.

4. Time-series optimization: leveraging Timescale's expertise in time-series data, Timescale Cloud for AI and vector data offers powerful optimization potential for applications that combine vector and time-series data.

Conclusion

As we've explored in this article, vector stores and vector databases are closely related tools in the domain of high-dimensional data storage and retrieval. Vector stores are storage and retrieval tools optimized around the specific technical requirements of embedded vector data, while vector databases connect vector stores to familiar structured database systems.

Understanding the connection between vector stores and databases is important for teams evaluating tools for their vector-data-based projects. Whether you're working on semantic search, recommendation systems, or other AI-powered applications that rely on vector embeddings, that understanding allows more informed decisions when choosing a system that balances performance, usability, and integration with existing workflows.

This article reviewed vector databases assessment criteria, which Timescale Cloud meets as it offers powerful, robust store and database tools built on a familiar PostgreSQL backbone. By combining cutting-edge vector search capabilities with the reliability and extensive feature set of PostgreSQL, Timescale Cloud provides a compelling option for organizations looking to implement vector-based applications without sacrificing database functionality or ease of use. 

Try Timescale Cloud for free

With one database for your application's metadata, vector embeddings, and time-series data, you can say goodbye to the operational complexity of data duplication, synchronization, and keeping track of updates across multiple systems. Let’s sum up what you get with Timescale:

• One mature cloud platform for your AI application (for vector, relational, and time-series data) 

• Flexible and transparent pricing that decouples compute and storage

• Ready to scale from day one so you can push to production with confidence

• Enterprise-grade security and data privacy, including SOC2 Type II and GDPR compliance

Ready to explore the capabilities of Timescale Cloud and see if it’s right for your vector project? You can find out by trying it for free today.