The Day Your Database Woke Up

Juan Loaiza had a startling wake-up call: companies that fail to recognize the new role of databases may soon find themselves digital dinosaurs in an era where data architecture determines destiny.

In an in-depth interview, the executive vice president for Database Technologies at Oracle explained the shift in the role of databases, from the unglamorous data repositories supporting data warehouses and lakehouses to becoming the critical nexus where business data, AI capabilities, and mission-critical operations converge. He also observed that companies that don’t shift their mindset about databases will be in danger of being left behind.

Where vectors mingle business data

One reason for Loaiza’s call lies in the seemingly forced marriage of vector search and traditional business data. Yet, this union is vital in today’s GenAI era.

“All of our big customers have a ton of business data, and then they also have documents, images, all that kind of stuff that’s really good to search with vectors, Loaiza explained. “And you want to combine the two."

Traditional databases that keep vector search and business data separate face a major issue. As companies rush to implement Retrieval Augmented Generation (RAG) as they try to fit open source models to their use cases, they’re discovering that vector embeddings create fundamental business data problems.

“If you search first on vectors, you might find lots of close matches on the vector side, but then when you go to the database business side, it might eliminate everything that you matched on that side,” Loaiza says. The same is also true vice-versa.

Oracle’s solution has been to integrate vectors directly into their database (especially in their latest Oracle Database 23ai) as a native data type with built-in indexing — essentially treating embeddings as first-class citizens alongside traditional data types.

This approach allows queries to work simultaneously against business data and vector data without compromising ACID compliance or real-time performance. It avoids the trap of sequential elimination and the dreaded "N+1 query" problem that plagues many RAG implementations.

Taming the multi-modal mayhem

It’s clear that the AI landscape is veering toward multi-modal AI models that can process text, images, and more. But, these require databases that go beyond traditional BLOB and CLOB storage types.

Loaiza admits that multi-modal support varies by customer needs. “Some people just have documents and they only care about that. Others just have images. But we do support all of that,” he says.

Such support at the database level expands use cases for various industries. Take the case of a healthcare provider implementing a diagnostic system. Its database must store and index not just patient records but radiological images, doctor notes, and medical histories. Multimodal support allows the database to maintain HIPAA compliance and enable semantic search across modalities.

Beyond multi-modal support, Loaiza emphasizes the importance of combining vector search with traditional text search. “As customers get more sophisticated in AI, they are starting to realize that the combination of text search and vector search is really powerful,” he observes.

This hybrid approach allows for both conceptual matching through cosine similarity calculations and precise keyword matching through inverted indexes — a powerful combination that’s proving essential for advanced AI applications.

Oracle’s latest Database 23ai extends this capability by introducing operators that combine text, vector, and business data conditions in a single SQL statement. This allows developers to implement hybrid search without complex application-level orchestration, effectively pushing the complexity down to the database level.

Real-time inferencing comes down to database speeds

For applications like fraud detection, real-time AI inferencing is non-negotiable. Oracle’s database technology is achieving sub-millisecond searches — a crucial benchmark for interactive applications.

“Generally, what most people think about is about 10 milliseconds, which is the limit of what a human can perceive,” Loaiza states. “That’s our goal, to keep everything below 10 milliseconds.”

This speed is achieved through advanced vector indexing and leveraging Oracle’s Exadata technology to perform searches directly in storage rather than pulling vectors into the database. By doing that, it shifts processing to where the data resides.

“We’ve shown that speeds can go up to 30 times," Loaiza notes. The performance holds even as data scales massively: “That’s something that we’ve worked a lot on with a lot of customers.”

Oracle’s Autonomous Database takes a step further by automatically tuning vector indexes based on workload patterns, removing the need for data administrators to manually fine-tune.

The database as a security sentinel

Regulatory complexity and AI-driven data access are pushing databases to become the frontline defenders of enterprise information. Companies like Oracle are bringing AI to the data, rather than moving data to AI systems. It’s a philosophy that aligns with the zero-trust security model increasingly embraced by CISOs.

“When you talk about financial data and health data, there’s a lot of regulations that prevent [moving the data from the] database and send it somewhere else. That’s considered high risk," Loaiza explains.

Equally critical is dealing with prompt injection attacks, which are becoming common as companies embrace LLMs. For example, if a customer asks about his or her account balance, a prompt injection attack can potentially trick the system into revealing sensitive data.

Oracle’s solution is to embed row-level security and attribute-based access control directly into the database. “We think the right answer is to let the database know who you are… and so it knows what data you have access to. And then when an LLM queries, it will only see your data,” Loaiza explains.

This creates the guardrails to prevent prompt inject attacks — essentially implementing a form of least-privilege access at the database level rather than relying on application-level controls that an LLM might be designed or co-opted to bypass.

Preparing for the Agentic AI frontier

As companies rush to implement Agentic AI, database security and integrity become even more critical. Loaiza sees three distinct levels of AI implementation that require different approaches: one for domain experts, another for employees, and the last for the general public.

For domain experts building with tools like LangChain or AutoGPT, productivity trumps perfect accuracy. For employees using agentic assistants, a balance is needed. For the general public interacting with customer-facing AI, safety is paramount. “You can’t make something available to the public that might be misleading or might produce catastrophic effects,” Loaiza emphasizes.

For example, a retail company implementing an AI inventory management system faces different challenges than one deploying a consumer-facing recommendation engine. The former might accept occasional errors in exchange for efficiency, while the latter could face serious reputational damage from inappropriate suggestions.

Oracle’s Autonomous Database now includes features specifically designed for agentic AI workloads, including the ability to automatically generate API endpoints that can be consumed by agent frameworks. These endpoints enforce security policies while exposing only the specific data and operations that agents need to perform their functions.

The new database paradigm: The intelligent foundation

Oracle’s strategy reveals a fundamental shift in how we approach databases. We’re witnessing the transformation of databases from passive storage systems to active, intelligent foundations for AI-driven enterprises.

For database engineers and architects, this shift demands new skills and perspectives. Understanding embedding models, similarity algorithms, and security implications of AI becomes as important as traditional database expertise. The role of the database professional also changes: it isn’t just optimizing queries; it’s about architecting intelligent systems with the databases at their core. It’s also not a place for a traditional database. 

Image credit: iStockphoto/Khanchit Khirisutchalual