Google Cloud is strengthening its analytics and transactional databases, including BigQuery, AlloyDB, and Spanner, with the aim of driving the development of its customers' ERP applications.
BigQuery, Google Cloud's top-of-the-line database for supporting analytics and AI workloads, has developed several AI enhancements. First, the company launched a preview of BigQuery and Vertex AI's integration in text and speech. Google Cloud says this will allow users to extract insights from unstructured data such as images and documents.
Gemini, the company's largest and most powerful AI model, is also available to BigQuery customers through Vertex AI. Last week, the model sparked some controversy by underperforming when it debuted in the consumer market.
These AI features follow BigQuery's earlier announced vector search feature. The vector search feature in the preview supports key components of the Genai application, such as similarity search and retrieval enhanced generation (RAG) using large language models.
Direct access to Vertex AI in BigQuery provides ease of use to Google Cloud AI customers in a number of ways, said Gerrit Kazmaier, general manager and vice president of data analytics at Google Cloud AI.
"As a data analytics practitioner, you can access all Vertex AI models, including our Gemini models, via the SQL command line or the BigQuery embedded Python API," Kazmaier said at a press release. "It's amazing because it means you don't need to go to a data scientist or a machine learning platform. You can access it in the field you're working on, on the data you have at hand. ”
The second benefit of integration, Kazmaier said, is better access to data from AI models. Prior to this integration, transferring data to AI models often required building and manipulation as well as data pipelines to move the data. No longer needed, he said. "All the complexities are gone. ”
The ability to combine text- and image-based AI models in Vertex (now available to data analysts via BigQuery) would also be of great benefit to customers.
This opens up a whole new phase of analyzing the scenario. He said that summarizing, extracting, classifying, condensing, and translating structured and unstructured data is a big deal. Roughly speaking, 90% of the data is unstructured. This data is often not used for enterprise data analysis because you can't process it in a meaningful way.
On the transactional (or operational) side, Google Cloud announced the general availability of AlloyDB AI, a dedicated version of the AI-hosted Postgres database that the company announced at last year's Next 23 conference. AlloyDB AI's ability to store vector embeddings and perform vector search functions is a core component of Google Cloud's customer GenAI use case.
Google Cloud has also launched a new integration with Langchain, a popular open-source framework that can help connect customer data into large language models (LLMs). Andi Gutmans, general manager and vice president of databases at Google Cloud, said that all of Google Cloud's databases will be integrated with Langchain.
Gutmans said the new features are in response to customer needs to find a way to get more GenAI value out of their data.
The company also announced that it will add vector search capabilities to other databases hosted by customers on its cloud, including Redis and MySQL. Cloud Spanner, Firestore, and BigTable will also get vector capabilities, Gutmans said.
What's special about Spanner is that it will have a Nearest Neighbor search feature, which is a slightly different variant. "What's really exciting are customers who have very, very large use cases — for example, trillions of vectors, like user-based height partitioning." As you can imagine, some of Google's internal apps are segmented by user – they will be able to store and search vectors in trillions (vectors) scale. ”
"Our belief is that any database, anywhere where operational data is stored, that you might need to use in a genai use case, should also have vector capabilities," he said. "It's no different than 15 to 20 years ago when JSON support was added to databases. We believe that a good vector function should just keep the basic functionality of the database. ”