Kinetica Launches Native Large Language Model for Language-to-SQL on Enterprise Data

Embedded LLM allows organizations to run SQL-GPT with enhanced privacy, security and greater fine-tuning

 Kinetica, the speed layer for generative AI and real-time analytics announced a native Large Language Model (LLM) combined with Kinetica’s innovative architecture that allows users to perform ad-hoc data analysis on real-time, structured data at speed using natural language. Unlike with public LLMs, no external API call is required and data never leaves the customer’s environment. This announcement follows Kinetica’s earlier innovation as the first analytic database to integrate with OpenAI.

Top Insights: Best Practices for Implementing Transaction Monitoring

In the wake of the excitement around LLMs, enterprises and government agencies are exploring innovative ways to automate certain business tasks, while keeping sensitive information that might be revealed through fine-tuning or prompt augmentation secure. Public LLMs, like Open AI’s GPT 3.5, raise privacy and security concerns that can be addressed by native offerings integrated into the Kinetica deployment and contained within the customer’s network perimeter.

In addition to being more secure, Kinetica’s native LLM is tailored to its syntax and industry data definitions such as telecommunications, financial services, automotive, logistics and others, creating more reliable and accurate SQL generation. This capability extends beyond standard SQL, ensuring efficient handling of intricate tasks for enhanced decision-making for time-series, graph, and spatial questions. Kinetica’s approach to fine-tuning prioritizes SQL generation optimization to ensure consistent and accurate results, versus more common techniques that prioritize creativity that yield diverse but unpredictable responses. This guarantees ongoing functionality for businesses and users, offering peace of mind in SQL query outcomes.

“Kinetica has led the market with next-level capabilities for analyzing sensor and machine data with our vectorized, real-time analytic database,” said Nima Negahban, Cofounder and CEO, Kinetica. “With the integration of SQL-GPT, we extend this capability to an entirely new horizon, empowering organizations to unleash the true potential of their real-time, structured data like never before.”

The U.S. Air Force has been working with Kinetica, leveraging advanced analytics on sensor data to quickly identify and respond to potential threats, helping to keep the skies safe and secure for all users of the national airspace system. The U.S. Air Force is now using Kinetica’s embedded LLM to detect threats in our airspace and identify anomalies using natural language.

“At Kinetica, we believe in fostering openness and embracing the diversity of generative AI models,” said Amit Vij, Cofounder and President, Kinetica. “We expect there will be different LLM platforms that emerge and we want to provide our customers with choice. While currently supporting two models, our commitment lies in continuously expanding our offerings to accommodate client-driven preferences and seamlessly integrate with a wide array of future models. Towards that end, Kinetica will roll out integration with other LLM platforms like NVIDIA NeMo later this year for language to SQL as new state of the art models emerge.”

Read More: CIO Influence Interview with Sumeet Arora, Chief Development Officer at ThoughtSpot

The Kinetica database converts natural language queries to SQL, and returns answers within seconds, even for complex and unknown questions. Further, Kinetica converges multiple modes of analytics such as time series, spatial, graph, and machine learning that broadens the types of questions that can be answered. What makes it possible for Kinetica to deliver on conversational query is the use of native vectorization. In a vectorized query engine, data is stored in fixed-size blocks called vectors, and query operations are performed on these vectors in parallel, rather than on individual data elements. This allows the query engine to process multiple data elements simultaneously, resulting in radically faster query execution on a smaller compute footprint. Vectorization is made possible by GPUs and the latest advancements in CPUs, which perform simultaneous calculations on multiple data elements, greatly accelerating computation-intensive tasks by allowing them to be processed in parallel across multiple cores or threads.