Google Cloud help our customers unify their data and connect it with groundbreaking AI to build transformative experiences. Data, whether it’s structured data in an operational database or unstructured data in a data lake, helps make AI more effective. For businesses to truly take advantage of generative AI, they need to access, manage, and activate structured and unstructured data across their operational and analytical systems.

At Next ‘23, Google laid out a vision to help developers build enterprise gen AI applications including delivering world-class vector capabilities, building strong integration with the developer ecosystem, and making it easy to connect to AI inferencing services. We’ve been hard at work delivering on that promise and today, we’re announcing the general availability (GA) of AlloyDB AI, an integrated set of capabilities in AlloyDB to easily build enterprise gen AI apps. 

Google also announcing vector search capabilities across more of our databases including Spanner, MySQL, and Redis to help developers build gen AI apps with their favorite databases, and we are adding integrations with LangChain, a popular framework for developing applications powered by language models. 

All these capabilities join our existing integrations with Vertex AI to provide an integrated platform for developers. Spanner and AlloyDB integrate natively with Vertex AI for model serving and inferencing with the familiarity of SQL, while Firestore and Bigtable integrate with Vertex AI Vector Search to deliver semantic search capabilities for gen AI apps.

Google believe the real value of generative AI is unlocked when operational data is integrated with gen AI to deliver real-time, accurate, and contextually-relevant experiences across enterprise applications. Operational databases with vector support help bridge the gap between foundation models and enterprise gen AI apps. And because operational databases typically store a majority of application data, they play a critical role in how developers build new, AI-assisted user experiences. That’s why 71% of organizations plan to use databases with integrated gen AI capabilities. Successful databases will evolve to be AI-first, and deeply integrate technologies such as vector search, with seamless connectivity to AI models, and tight integrations with AI tooling and frameworks. All these will be natively built into or around operational databases as table stakes. 

AlloyDB: A Modern PostgreSQL database for generative AI workloads

AlloyDB is Google Cloud’s fully managed PostgreSQL-compatible database designed for superior performance, scale, and availability. Today, we’re announcing that AlloyDB AI is generally available in both AlloyDB and AlloyDB Omni. Built for the future, AlloyDB:

  • is optimized for enterprise gen AI apps that need real-time and accurate responses
  • delivers superior performance for transactional, analytical, and vector workloads
  • runs anywhere, including on-premises and on other clouds, enabling customers to modernize and innovate wherever they are.

Customers such as Character AI, FLUIDEFI, B4A and Regnology, are using AlloyDB to power their applications. For example, Regnology’s regulatory reporting chatbot leverages natural language processing to understand complex regulatory terminology and queries.

“AlloyDB acts as a dynamic vector store, indexing repositories of regulatory guidelines, compliance documents, and historical reporting data to ground the chatbot. Compliance analysts and reporting specialists interact with the chatbot in a conversational manner, saving time and addressing diverse regulatory reporting questions.” – Antoine Moreau, CIO, Regnology

Vector search across all Google Cloud databases 

Vector search has emerged as a critical capability for building useful and accurate gen AI-powered applications, making it easier to find similar search results of unstructured data such as text and images from a product catalog using a nearest neighbor algorithm. Today, we’re announcing vector search across several Google Cloud databases, including Cloud SQL for MySQL, Memorystore for Redis, and Spanner, all in preview. 

Cloud SQL for MySQL now supports both approximate and exact nearest neighbor vector searches, adding to the pgvector capabilities we launched last year in Cloud SQL for PostgreSQL and AlloyDB. Developers can now store millions of vectors in the same MySQL instances they are already using. By utilizing Cloud SQL for vector searches — whether on MySQL or PostgreSQL — you can store and perform vector searches directly in the same operational database you’re already using without having to learn or set up a new system. 

In addition, to provide ultra-fast performance for your gen AI applications, Google are launching built-in support for vector storage and search for Memorystore for Redis. Each Redis instance will be capable of storing tens of millions of vectors and can perform vector search at single digit millisecond latency. This provides an ultra-low-latency data store for a variety of use cases such as LLM semantic caching and recommendation systems. 

Spanner can scale vector searches for highly partitionable workloads. Large-scale vector workloads that involve billions of vectors and millions of queries per second can be challenging for many systems. These workloads are a great fit for Spanner’s exact nearest neighbor search because Spanner can efficiently reduce the search space to provide accurate, real-time results with low latency.

Accelerating ecosystem support for LangChain

LangChain has grown to be one of the most popular open-source LLM orchestration frameworks. In our efforts to provide application developers with tools to help them quickly build gen AI apps, we are open-sourcing LangChain integrations for all of our Google Cloud databases. We will support three LangChain Integrations that include Vector stores, Document loaders, and Chat Messages Memory.

By leveraging the power of LangChain with our databases, developers can now easily create context-aware gen AI applications, faster. The LangChain integration provides them built-in Retrieval Augmented Generation (RAG) workflows across their preferred data source, using their choice of enterprise-grade Google Cloud database. Example use cases include personalized product recommendations, question answering, document search and synthesis, and customer service automation. 

Integration with specific LangChain components simplifies the process of incorporating Google databases into applications. Supported components include:

  • Vector stores, to support vector similarity queries. The LangChain Vector stores integration is available for AlloyDB, Cloud SQL for PostgreSQL, Cloud SQL for MySQL, Memorystore for Redis, and Spanner.
  • Document loaders, which allow for seamless data loading from various external sources such as web page content, or a transcript of a YouTube video. 
  • Chat Messages Memory, allowing storage of chat history for future reference by providing deeper context from past conversations.

Both the Document loaders and Memory integrations are available for all Google Cloud databases including AlloyDB, Firestore, Bigtable, Memorystore for Redis, Spanner, and Cloud SQL for MySQL, PostgreSQL, and SQL Server.  

These packages are now available on GitHub.

Embrace an AI-driven future

There’s a wealth of data in operational databases just waiting to power the next transformative gen AI models and applications. By enhancing AlloyDB AI for enterprise grade production workloads, adding extensive vector search capabilities across our database portfolio, and embracing generative AI frameworks from the community, developers have the tools they need to start adding intelligent, accurate, and helpful gen AI capabilities to their applications that are grounded on the wealth of data in their enterprise databases.