No Need to Train Your RDB Foundation Model

Relational databases (RDBs) contain vast amounts of heterogeneous tabular information that can be exploited for predictive modeling purposes. But since the space of potential targets is vast across enterprise settings, how can we avoid retraining a new model each time we wish to predict a new quantity of interest? Foundation models based on in-context learning (ICL) offer a convenient option, but so far are largely restricted to single-table operability. In generalizing to multiple interrelated tables, it is essential to compress variably-sized RDB neighborhoods into fixed-length ICL samples for consumption by the decoder. However, the details here are critical: unlike existing supervised learning RDB pipelines, we provide theoretical and empirical evidence that ICL-specific compression should be constrained within high-dimensional RDB columns where all entities share units and roles, not across columns where the relevance of heterogeneous data types cannot be determined without extensive label information. Conditioned on this restriction, we then demonstrate that encoder expressiveness is actually not compromised by excluding trainable parameters. Hence we arrive at a principled family of RDB encoders that can be seamlessly paired with already-existing single-table ICL foundation models, whereby no training or fine-tuning is required. From a practical standpoint, we develop scalable SQL primitives to implement the encoder stage, resulting in the easy-to-use open-source RDBLearn foundation model capable of robust performance on unseen datasets out of the box.

Businesses store data across many linked tables (e.g., customers in one, orders in another, products in a third) that together form a relational database (RDB). In order to exploit these rich data sources for making predictions such as future sales, the traditional requirement has been to build a custom model each time. Recent so-called AI ``foundation models'' can skip this retraining step, but they typically only work with one table at a time. Hence it is natural to ask, how can we bring together information from a full RDB without confusing the model, while still avoiding additional task-specific training steps? The answer provided by this work is to keep it simple and avoid trying to blend everything into one dense summary up front. Instead, we demonstrate that it can be advantageous to summarize each RDBcolumn on its own as an intermediate featurization step, since values within a column share the same meaning. We exploit these findings through RDBLearn, a free, open-source multi-table foundation model that handles new databases out of the box with no training required.