Adaptive Instructed-Retriever: Frontier-Quality Search at 2x Lower Latency
Databricks, Wednesday, September 9th, 2026
A retrieval model trained to spend extra search steps only where query complexity justifies the cost.
Databricks introduces the Adaptive Instructed-Retriever, which combines parallel and sequential multi-step search and applies additional retrieval passes selectively based on query complexity, exiting early on simple requests.
It was trained with reinforcement learning to balance retrieval quality against compute cost and ships as multiple checkpoints with different quality-latency profiles.
Benchmarks claim frontier-comparable quality at roughly half the latency. The adaptive element matters for cost control in agentic systems, where uniform retrieval depth means simple queries subsidize complex ones.