Oracle HeatWave database gets host of generative AI features – Business
Continuing its Swiss Army knife strategy with its managed MySQL database service called Heatwave, Oracle Corp. today is endowing the product with generative artificial intelligence capabilities.
They include in-database large language models, automated in-database vector store, scale-out vector processing, the ability to query in natural language, and tools that enable customers to build their own natural language chat applications. HeatWave GenAI is available immediately in all Oracle Cloud Infrastructure regions and across other cloud platforms at no extra cost to HeatWave customers.
New features enable developers to create a vector store for unstructured content with a single SQL command using built-in embedding models. A vector store is a type of database optimized for storing and processing vector data types, which are high-dimensional representations of data items often used for tasks like recommendation systems and similarity searches.
Oracle said data never has to leave the database, and training and inferencing tasks can be accomplished without the need for graphic processing units.
In-database LLMs enable users to search data, generate or summarize content and perform retrieval-augmented generation with HeatWave Vector Store. They can also combine generative AI with other built-in HeatWave capabilities, such as machine learning, and tap into the OCI Generative AI service to access pretrained foundational models.
Steadily growing functionality
Oracle has been steadily building out HeatWave features since it debuted the database-as-a-service in late 2020 with hybrid transactional and analytical features. It added advanced automation in 2021. In 2022 it folded in machine learning support and lakehouse features and made the service available on the Amazon Web Services Inc. cloud. HeatWave is now also available on Microsoft Corp. Azure via the Oracle Interconnect for Azure. Oracle recently dropped “MySQL” from the product’s name.
Although HeatWave is increasingly resembling Oracle’s flagship Autonomous Database, the company is positioning it as an offering for companies with less ambitious processing needs.
“For a customer who wants the best price/performance and doesn’t have the Oracle database, this is the best solution,” said Steve Zivanic, global vice president of database and autonomous services product marketing. “Naturally, the Autonomous Database is our flagship.”
Oracle typically positions new features as being easy to use. For example, creating a vector store database is simply a matter of pointing to a file or folder. “It’s fully automated,” said Nipun Agarwal, senior vice president of MySQL Database and HeatWave.
Smaller LLMs
The in-database LLMs are small versions of Mistral AI SAS’ namesake model and the open-source Llama 3 developed at Meta Platforms Inc. “We further quantize them and we find the accuracy and performance is about the same,” Agarwal said. “Only one invocation is required to create a vector store compared to nine in a typical scenario.”
Quantizing is a technique to reduce a model’s size and computational demands while maintaining performance by converting parameters from floating-point representations to lower-precision formats, such as integers.
In-database LLMs can also run in combination with other functions like machine learning, Agarwal said. “For example, in anomaly detection, machine learning can be used to find documents with anomalies and feed them to the LLM to generate faster results,” he said. “The size of the input context is greatly reduced, and time to run goes from six hours to three-and-a-half minutes.”
Oracle has always made bold claims about HeatWave’s performance and they have mostly held up to scrutiny. This time is no different. The company claims HeatWave can process similarity searches, which are widely used in generative AI, 30 times faster than data management products from Snowflake Inc. and 15 times faster than Databricks Inc., and do it at a lower cost.