Skip to main content
relativedb logo

RelativeDB

Predictive queries over relational data. Declare the shape of your relational data, wire small retrievers over your storage, and ask questions about the future:

PREDICT NOT EXISTS(orders.*)
FROM customers

"For every customer, the probability they place zero orders"

Claude Code
Install the plugin
> /plugin marketplace add RelativeDB/RelQL-cc
> /plugin install RelQL@RelQL
Then ask
>/relql which of my customers are about to stop ordering?
relqlscored 12,304 customers — top risk Jane Doe · 0.94

See it in action

Video coming soon

What are Relational Transformers?

A transformer normally attends over word tokens. A relational transformer attends over a small subgraph of your database. The relational analogue of prompting an LLM, in a 86M-parameter model that scales.

CapabilityRelational TransformerRelativeDBGBDTsXGBoost & co.Graph neural netsper-schema GNNsLLMs on rowsserialized to text
No hand-built featuresyes — raw cells. no — feature tables. partial — graph wiring. partial — prompt design.
No per-task trainingyes — pretrained. no — every task. no — schema + task. yes — pretrained.
Zero-shot on new tasksyes — in-context. no. no. partial — if it fits text.
Schema-native structureyes — keys, rows, cols. no — flat table. yes — graph edges. no — flattened text.
Typed cells & real timeyes — native. partial — manual. partial — manual. no — lost as text.
No train/serve skewyes. no — common. no — common. yes.
Small & scalableyes — 86M params. yes — tiny. partial — grows. no — billions.
built in partial / manual not really

RelativeDB ships RT-J inference as a highly optimized, dependency-light C++ engine, with several quantized models for highly constrained environments. Read more →