GCP Latest Announcements: Vector Support in PostgreSQL Services

Artificial intelligence (AI) is revolutionizing the way we interact with information, brands, and each other. It is anticipated that nearly all applications in the future will be AI-driven. As developers strive to build these cutting-edge applications, they require the necessary tools to incorporate new and immersive experiences seamlessly.

These experiences heavily rely on application state, user inputs, and operational data to provide context that caters to users’ needs effectively. Operational databases, which are already at the core of some of these applications, play a critical role in enabling innovative generative AI user experiences.

Today, Google Cloud announced the introduction of vector support in Cloud SQL for PostgreSQL and AlloyDB for PostgreSQL. This new capability empowers you to utilize your database for storing and indexing vector embeddings generated by large language models (LLMs), leveraging the renowned pgvector PostgreSQL extension. By harnessing this power, you can unlock the full potential of generative AI in your applications. LLMs running on Vertex AI and other AI platforms can efficiently discover similar items using both exact and approximate nearest neighbor search techniques. Furthermore, you can leverage the data and features of relational databases to enhance and process the information.

Vector embeddings are numerical representations that simplify complex user-generated information, including text, music, and video, so that it may be easily saved, edited, and searchable. These representations are produced using embedding models in a way that places their individual embeddings adjacent to one another in the embedding vector space if two pieces of material have semantic similarities. Then, you may effectively filter data based on similarity using the indexed vector embeddings.

For instance, if you are a clothing retailer, you may want to provide product recommendations that are similar to the items in a user’s shopping cart. With the help of Vertex AI’s pre-trained text embeddings model, you can generate embeddings based on product descriptions. The pgvector extension allows you to store and index these embeddings in your Cloud SQL or AlloyDB database, enabling you to easily query for similar products. By combining these embeddings with your operational data, you can tailor the results based on the user’s preferences and the product’s attributes. For instance, you can apply filters based on structured data such as price or category and even join the results with real-time operational data like user profiles and inventory information to ensure that only in-stock items of the user’s size are displayed.

Vector embeddings also play a vital role in helping developers leverage pre-trained LLMs, which have gained tremendous popularity in the past year. These LLMs are trained on vast amounts of data and can be applied to various use cases, including translation, summarization, question answering, and creative writing. Their flexibility in application development lies in the fact that they can be tailored to specific needs without necessitating machine learning expertise or custom model training. By strategically crafting a prompt, developers can customize the output of an LLM and ground it using application-specific contextual data such as documentation, user profiles, and chat history.

One important aspect to consider about LLMs is their lack of state. However, to ensure seamless conversations and provide relevant responses, the chat history is crucial. Due to strict input token limits, it is not always possible to provide the entire context in the prompt. Embeddings offer a solution by enabling the storage of extensive contexts, such as documentation or long-term chat histories, in your database. You can filter and retrieve the most relevant information from these embeddings. In essence, the context can encompass both the conversation itself and pertinent information that might not have been explicitly mentioned. By feeding this contextual data to the model at each iteration, you can simulate long-term memory and integrate business-specific knowledge effectively.

Key Benefits

By incorporating vector support directly into Cloud SQL and AlloyDB databases, you can create AI-enabled experiences that are fully aware of the application and user state. In addition to the fundamental vector support, Cloud SQL and AlloyDB offer the following benefits:

Enterprise-Grade Serving Stacks

Cloud SQL and AlloyDB are fully managed services that power low-latency serving applications with enterprise-grade capabilities. These databases offer automatic provisioning and maintenance, reducing database costs. Furthermore, they provide features such as high availability, data protection, and built-in security, all supported by our 24/7 SRE team.

Ease and Familiarity

With vector support in PostgreSQL, you can utilize your existing operational database to drive AI-enabled experiences. Leverage your familiarity with PostgreSQL and utilize the entire PostgreSQL ecosystem to its fullest potential.

Tight Integration with Operational Data

We firmly believe that the most remarkable AI-enabled experiences are seamlessly integrated with the application itself, leveraging real-time transactional data to enhance user experiences. AlloyDB and Cloud SQL for PostgreSQL make this integration possible by embedding AI capabilities directly into your operational database. This support enables powerful queries across structured and unstructured operational data, combining vector predicates with standard PostgreSQL filters.

When you choose AlloyDB or Cloud SQL for PostgreSQL for your enterprise workloads, you can leverage the power of vector embeddings alongside your fully managed operational database.

Integrating with Vertex AI

The vector support in Cloud SQL and AlloyDB truly shines when combined with the generative AI services offered by Vertex AI. Vertex AI provides a wide range of pre-trained foundational and embeddings models across text and images. With AlloyDB, you can even directly invoke custom Vertex AI models from the database, ensuring high-throughput and low-latency augmented transactions. Together, these capabilities provide a comprehensive toolkit for integrating large language models into your applications.

Moreover, you can also leverage the vector support in conjunction with the Vertex AI Matching Engine, the industry’s leading high-scale, low-latency vector database. By utilizing this capability, you can store embeddings and perform vector similarity matching. With Cloud SQL and AlloyDB, you can achieve similar functionality, empowering you to use vector embeddings directly within your application. This integration allows you to leverage real-time operational data and the rich features of the trusted PostgreSQL database, resulting in enhanced user experiences.

In conclusion, with the announcement of vector support in PostgreSQL services, a new era of AI-enabled applications is upon us. By leveraging this powerful capability, developers can create transformative experiences that resonate with users. The seamless integration of vector embeddings with operational databases provides the foundation for building advanced generative AI applications. Embrace the power of vectors and unlock the true potential of AI in your applications.