Databricks debuts Adaptive Instructed-Retriever to accelerate agentic search

Databricks published research this week introducing Adaptive Instructed-Retriever, a model built to decide how many search steps are necessary before execution. [1] According to the company's unverified testing, the model matches the answer quality of GPT-5.6 Luna, DeepSeek-V4-Flash, and Claude Sonnet 5. [1] The system answers questions in an average of 5.8 seconds, operating more than twice as fast. [1] Most agentic search tools previously forced a choice between missing multi-hop reasoning by stopping early or running unnecessary steps for simple queries. [1]

The previous iteration, Instructed Retriever, launched in January for curated workspaces and outperformed traditional RAG by up to 70 percent on complex enterprise questions using metadata reasoning. [1] Databricks research director Michael Bendersky stated that the new architecture moves away from a fixed workflow to an adaptive approach that changes its plan based on the user's question. [1] In one benchmark, the adaptive search matched Sonnet's two-step search while returning 50 percent higher recall on a customer account query. [1]

The shift toward dynamic retrieval step planning suggests that the next phase of enterprise RAG will focus heavily on computational efficiency rather than just recall. By matching the output quality of GPT-5.6 Luna and Claude Sonnet 5, Databricks is positioning its specialized retrieval architecture as a faster, purpose-built alternative to relying solely on generalist frontier models for multi-hop enterprise search. However, because these figures stem from unverified internal testing, independent benchmarks will be necessary to confirm the latency advantage at scale. What the sources don't address: How the adaptive search impacts total compute costs compared to running fixed query plans.