Databricks introduced Adaptive Instructed-Retriever, a new retrieval model designed to let AI agents dynamically decide how much searching a query requires. Simple requests can stop after an initial retrieval step, while more complex questions trigger additional sequential searches.

Databricks says the model achieved comparable retrieval quality to leading third-party and open-source models while delivering more than 2x lower latency in its benchmark suite. The technology is designed for retrieval layers behind products including Genie Code, Genie One and Genie Agents.

Source: databricks.com