AI agents increasingly need retrieval systems tailored to specific domains and tasks, but adapting embedding models to those needs typically requires carefully hand-crafted training data. That manual process is expensive and does not scale well, creating a bottleneck for specialised applications.

The authors propose SYNTRA, a self-improving agentic framework that generates its own training data through an iterative process. Starting from a small amount of initial input, the framework synthesises new examples and uses them to refine the model, gradually improving its own performance without ongoing human intervention.

By automating data synthesis, SYNTRA offers a path to more scalable adaptation of embedding models. The paper argues that this reduces the cost and effort associated with building specialised retrieval systems, potentially making them more accessible across a wider range of domains.