Databricks launches adaptive AI retrieval model that cuts search time in half

Databricks Introduces Adaptive AI Retrieval Model for Enterprise Search

Databricks unveiled Adaptive Instructed-Retriever, an AI retrieval model designed to balance answer quality, latency, and cost for enterprise AI applications. The model builds on the company's earlier Instructed-Retriever-1 but introduces a critical capability: it dynamically decides when queries need additional search steps and when to stop early on simpler requests1

1

. This adaptive approach addresses a persistent challenge in AI-driven search—applying multi-step search to every query increases latency and computational costs, while limiting queries to single-step searches compromises results for complex, multi-hop questions1

1

.

The model targets data agents working across large enterprise workspaces, where straightforward lookups should return quickly while complex discovery tasks can justify additional time and resources3

3

. Databricks reported that Adaptive Instructed-Retriever matched or exceeded the retrieval quality of Claude Sonnet 5, GPT-5.6 Luna, and DeepSeek-V4-Flash while completing requests in 5.8 seconds—more than twice as fast as those frontier models1

1

2

2

.

How Parallel and Sequential Thinking Enable Smarter Search

Adaptive Instructed-Retriever operates using two distinct retrieval modes. In parallel thinking, the system issues several rewritten versions of a query simultaneously through separate threads, merging results at the agent level2

2

. Sequential thinking activates when initial retrieved documents contain clues but not complete answers—the model uses that information to issue second-round queries, repeating the process up to a fixed step limit2

2

.

Michael Bendersky, research director at Databricks, explained the shift from the company's earlier work: "The initial Instructed Retriever work, where the workflow was pretty fixed, was more similar to the query plan because we fixed the structure of how the agent calls the retrieval system. In this particular work, we move away from this to a more adaptive approach, where the agent adapts to the question and changes the plan based on the question"2

2

.

In one benchmark example, when asked whether a company reported restructuring costs on a specific income statement line, Adaptive Instructed-Retriever verified the answer in two search steps while Claude Sonnet 5 required three steps and GPT-5.6 Luna needed four3

3

. The model achieved 50% higher recall than Sonnet on a customer account question while matching its two-step search performance2

2

.

Reinforcement Learning Trains Models to Balance Speed and Accuracy

Databricks trained Adaptive Instructed-Retriever using synthetic enterprise retrieval environments and an agentic data synthesis process, reusing training data from Instructed-Retriever-1 while adding synthetic multi-hop questions designed to benefit from multiple search steps1

1

. The company applied online reinforcement learning through a technique called CISPO—Clipped Importance Sampling Policy Optimization2

2

.

This training approach rewards high-performing search trajectories while penalizing steps that don't produce corresponding gains in retrieval quality1

1

. "We train the agent to search more, but only when needed," Bendersky said. "We get a penalty that balances the time spent searching with the reward of finding the right documents"2

2

.

Varying the strength of the step penalty during training produces multiple model checkpoints with different quality-latency trade-offs1

1

. A heavier penalty favors fewer steps and faster answers, while a lighter penalty allows more searching to achieve higher retrieval quality3

3

. Customers can select checkpoints suited to interactive applications requiring speed or slower offline workloads prioritizing thoroughness3

3

.

Cost Savings and Predictability for Enterprise AI Agents

The ability to limit unnecessary search steps could help enterprises manage AI costs as they scale agentic AI systems into production. "Repeated searches can improve answers to complex questions but add latency and compute costs, while a single search can be faster and cheaper but fall short on harder queries. The ability to determine when additional search is worthwhile could therefore help enterprises avoid spending resources without improving answer quality," said Ashish Chaturvedi, executive research leader at HFS Research1

1

.

Chaturvedi emphasized the financial implications: "The problem with agentic AI at scale is that consumption is hard to forecast, agents searching and re-searching create compounding, unpredictable cost and latency, and finance teams hate these variable bills. Knowing your agents will search within a defined ceiling, and that you can set that ceiling per workload, is what makes agentic search safe to run at scale rather than a runaway meter"1

1

.

Retrieval operates as a high-frequency operation for data agents, making the speed and cost advantages potentially significant at scale1

1

. The model's ability to deliver competitive retrieval quality with lower latency could allow enterprises to avoid applying the cost of larger models to every retrieval request1

1

.

Reducing Developer Orchestration Work in Multi-Hop Questions

Adaptive Instructed-Retriever shifts decision-making from developers to the model itself. Traditionally, developers handling complex, multi-hop questions design control flow for multiple search rounds themselves, deciding when to refine queries, search again, or stop1

1

. The model's learned search policy handles those decisions internally, reducing orchestration-level work1

1

.

Databricks tested the model across seven held-out internal and external benchmarks spanning different domains and levels of search difficulty3

3

. The company trained the model using Databricks AI Runtime, which customers can use to specialize models for their own data and performance requirements3

3

. The model serves as a retrieval building block for Databricks' Genie Code, Genie One, and Genie Agents3

3

.