Introducing generative relational models and launching SDV 2.0
Today, we are introducing a new class of models called Generative Relational Models, or GRMs, for short.
We're also launching SDV 2.0, a major new release of SDV Enterprise, that enables organizations to build GRMs at enterprise scale with greater automation.
What is a GRM?
A GRM is a relational database captured in a generative model.
Trained from scratch, it learns the database as a connected whole, capturing the statistical patterns, data structure, relationships, context, and business rules embedded in the data and spanning its tables.
Once trained, it can generate an entirely new synthetic database from the model alone or, given a small number of rows, predict and generate the related rows and tables.
As a result a GRM can support analytical queries, make predictions, and simulate scenarios in addition to generating synthetic databases.
What is its significance for enterprise relational data?
An enterprise's relational data is proprietary and specific to that enterprise. It has been collected over years, if not decades. It contributes to the enterprise's competitive advantage. Unlike LLMs, which are trained on broadly available data, each GRM learns from an enterprise's proprietary data and the resulting model is owned by the enterprise.
SDV 2.0 enables enterprises to build GRMs locally, privately, and efficiently. While LLMs can bring the world's intelligence to an enterprise, GRMs bring enterprise intelligence to LLMs while keeping enterprise data within their controlled environments.
With these models, enterprises don't have to reconstruct an understanding of their data one application at a time or one use case at a time. They can build them once, and then use them for several downstream applications. To date, one of the most valuable things they can do is providing synthetic data as a service to support AI development, train and test AI agents, and test software applications.
What is new with SDV 2.0?
For more than a decade—first at MIT and then at DataCebo—we have focused on how to help enterprises unlock the power of their relational data.
We built one of the first algorithms to train generative models for a relational database. That early work became the foundation for what we now call Generative Relational Models.
Three years ago, we publicly released version 1.0 for our software The Synthetic Data Vault, (or SDV), so that anyone can build their own GRMs and use them for their applications.
Over the past two years, our enterprise deployments showed us that building generative relational models at scale requires much more automation.
Real enterprise databases are messy, poorly documented, and deeply interconnected. Much of their business rules are implicit in the data rather than documented in the schema.
That is why we are making SDV 2.0 generally available today. It automatically:
- Infers complex schemas and data structures
- Detects business rules embedded in the data and automatically applies them as constraints
- Learns statistical patterns across 100's of tables
- Ensures generated data adheres to those constraints
- Enables creation of data for specific scenarios
Start exploring SDV 2.0 today:
