During a keynote at Oracle CloudWorld 2023 in Las Vegas, Oracle Executive Vice President Juan Loaiza explained how the newest set of features in Oracle Database 23c integrate generative AI to make life easier for developers, data professionals, and application users.
The features include offering a much more efficient way for developers to work with JSON documents and new vector database capabilities that support natural language queries and help customers build apps that combine generative AI models with their own data, Loaiza said.
More specifically, Oracle Database 23c now offers support for vectors as a native data type within the database. The result, Loaiza said, will make it easier to combine searches on semantic and business data—allowing developers to take advantage of large language models (LLMs) in their applications.
A vector, Loaiza said, represents complex unstructured data in a way that makes it easy to find based on natural language queries. For example, a picture of a house will have numbers assigned to aspects of the house, such as construction materials, number of stories and windows, and so forth—around a thousand numbers that describe the house in the image. That string of numbers is the image’s “vector.” A string of numbers can also be used to describe the content and themes in text documents and videos.
With vectors stored natively in Oracle Database, Loaiza said, developers can build enterprise apps that use generative AI. “Not only can the database store and query these vectors, but it can also match them up with business data,” he said. “By adding AI vector search to Oracle Database, we help customers to quickly and easily get the benefits of artificial intelligence without sacrificing security, data integrity, or performance.”
“By adding AI vector search to Oracle Database, we help customers to quickly and easily get the benefits of artificial intelligence without sacrificing security, data integrity, or performance.”
To make his point, Loaiza used an example of a shopper who likes a house, snapping a picture of it and uploading it to a real estate app. Generative AI can use its own training to find similar houses. Meanwhile, Oracle Database will pull in business data about similar houses, such as price and location. This lets the AI quickly find matches that look like the desired house and are in the price range and neighborhood the shopper wants. This is a small example of a big transformation of how data is searched and used, Loaiza said. “This is how the world of data meets the world of generative AI.”
Loaiza was joined onstage by Aidan Gomez, CEO of Cohere, which builds LLMs designed to help enterprises take advantage of AI. Gomez is co-inventor of the Transformer architecture, the foundation of generative AI. He agrees that AI vector search in Oracle Database 23c will drive a new era of AppDev productivity when combined with another new feature in Oracle Database 23c called retrieval-augmented generation (RAG). RAG is a breakthrough generative AI technique that uses vectors to combine LLMs and private business data to deliver responses to natural language questions. RAG is interesting to Cohere engineers for two reasons, Gomez said. It gives the company’s LLMs access to highly secure enterprise data without needing to include it in the LLM training data. RAG also helps with the thorny problem of AI hallucinations, which are produced when LLMs generate false information presented as fact.
With RAG, an LLM can note where it sourced a piece of information. That means users of an enterprise application can easily verify that an output is free of AI hallucinations. “RAG helps with AI reliability because now an AI can cite where it got the information it’s sharing with you,” Gomez said.
The upshot of having AI vector search and RAG in the database, Loaiza said, is that Oracle Database users can add generative AI capabilities to their own applications and workflows through simple APIs while retaining security, performance, and cost benefits.
Loaiza said his team is taking these capabilities a step further, imbuing Oracle Database tools such as low-code platforms Oracle APEX and SQL Developer with generative AI capabilities. This lets developers use natural language to generate applications or SQL queries without writing code. “We’re working on a number of projects to improve productivity using generative AI,” he said.
The focus is on simplifying the way data professionals, developers, and data users interact with data, he said. Oracle will generate a “blueprint” of an application based on a developer’s natural language, Loaiza said, letting the developer declare the intended outcome and get a first draft of the application instead of hand coding it.
Loaiza detailed other significant developments in Oracle’s flagship database, such as JSON-Relational Duality, a feature that gives developers the best of JSON and relational data types in their applications. Of course, “JSON documents are just one data type that developers want to use in their full-featured applications,” Loaiza said. There’s also relational data, graph, spatial, blockchain, and now vectors.
Managing all these data types separately can lead to applications with higher complexity and lower security, he said. The way to make application development simpler is to bring them together in one full-featured converged database, he said. “When you look at Oracle Database 23c, you can see all the concrete ways that a database can help developers, data professionals—by simplifying development and giving them tools to use data with LLMs,” Loaiza said. “These are huge changes, and we’re bringing it all together.”