Databricks announces Adaptive Instructed-Retriever for agent search

Databricks announced Adaptive Instructed-Retriever on September 9. The retrieval model for data agents combines parallel, single-step retrieval with sequential, multi-step search. The company said it returns early when it has enough evidence and keeps searching on harder requests, up to a fixed step limit.

In Databricks’ tests across seven held-out internal and external benchmarks, the model answered in an average of 5.8 seconds. The company said it matched the retrieval performance of Claude Sonnet 5, GPT-5.6 Luna and DeepSeek-V4-Flash while running more than twice as fast as each comparison model. Those results are vendor benchmarks. The post does not disclose complete per-benchmark scores, test hardware or model access sufficient to replicate them.

The adaptive stopping mechanism assigns more computation only when another search step is likely to improve retrieval quality. Straightforward requests can finish after one round, while multi-hop questions can trigger additional searches to gather evidence and refine later queries. The fixed upper bound is intended to keep latency and cost predictable even when the model uses sequential search.

Databricks said it trained the model with online reinforcement learning using Clipped Importance Sampling Policy Optimization, or CISPO. Its reward balances retrieval quality against the cost of additional search steps. Changing the penalty for extra steps produced checkpoints at different points on the quality-latency frontier, allowing a workload to favor lower latency or higher retrieval quality. The announcement follows Databricks’ work on Omnigent and Nimble agent-search tooling.

The company said it used Databricks AI Runtime to train the model and that AI Runtime is available to customers. The announcement did not state that Adaptive Instructed-Retriever itself is generally available, and it did not provide access terms or a release date for the model.