Databricks Adds Adaptive Enterprise Search, Claims 5.8-Second Answers

Databricks has introduced Adaptive Instructed-Retriever, a model that changes the number and type of document searches for each enterprise question. The company says it matched several frontier models across its tested retrieval budgets while averaging 5.8 seconds per answer.

The product is aimed at a familiar retrieval problem: one search can miss answers spread across documents, while a fixed multi-step process spends extra time on questions that did not need it. Adaptive Instructed-Retriever is designed to decide how many search steps a question warrants.

One task, two search paths

For questions that could benefit from different wording, the system can rewrite a query several ways and search in parallel. When initial documents offer a clue but not the answer, it can instead run sequential follow-up searches, using the retrieved material to guide the next query.

How it learns when to stop

• Databricks trained the model with online reinforcement learning using a technique it calls CISPO.
• Additional search steps carry a penalty, intended to balance finding the right documents against the time spent searching.

A narrow comparison with frontier models

Databricks compared its retriever with Claude Sonnet 5, GPT-5.6 Luna and DeepSeek-V4-Flash on the same search task. It says Adaptive Instructed-Retriever matched those models across the retrieval budgets tested, presenting the result as a case for a specialized retriever rather than a larger general-purpose model.

In a customer-account benchmark example, Databricks says its system used the same two search steps as Claude Sonnet 5 while returning 50% higher recall. In another example concerning restructuring costs, it says the system reached perfect recall in two steps, versus three for Claude and four for GPT-5.6 Luna.

The efficiency case still needs outside testing

The figures are from Databricks’ testing and have not been independently verified. The release therefore establishes the company’s approach and its reported quality-latency results, not that those results will hold across other enterprise document collections.

Databricks introduced its earlier Instructed Retriever in January and said it improved on traditional RAG by up to 70% for complex, instruction-heavy enterprise questions. Adaptive Instructed-Retriever extends that retrieval-focused line of work with a variable search plan.