Databricks Releases Adaptive Retrieval Model That Beats Claude and GPT-5.6 Luna on Speed

Quick Facts

• Databricks’ Adaptive Instructed-Retriever answers queries in an average of 5.8 seconds, more than twice as fast as Claude Sonnet 5, GPT-5.6 Luna, and DeepSeek-V4-Flash.
• The model uses a reinforcement learning technique called CISPO to decide how many search steps a query needs before running, rather than applying a fixed number of steps to every request.
• Databricks says more than 100,000 agents have been built on its Agent Bricks platform, which now processes over a quadrillion tokens per year.

Databricks on Wednesday introduced Adaptive Instructed-Retriever, a retrieval model built for enterprise AI agents that matches the answer quality of Claude Sonnet 5, GPT-5.6 Luna, and DeepSeek-V4-Flash while responding more than twice as fast.

The model averages 5.8 seconds per query. Databricks says every checkpoint on its quality-latency curve outperforms those three competing models across the full range of retrieval budgets tested.

How It Works

The core distinction from earlier retrieval systems is adaptability. Traditional retrieval-augmented generation applies a fixed number of search steps to every query, which wastes compute on simple lookups and can still fall short on complex, multi-hop questions.

Adaptive Instructed-Retriever decides how many steps to take before it begins. It can combine parallel retrieval with sequential, multi-step search when additional evidence is needed, or complete a query in a single step when the answer is straightforward.

Databricks trained the model using a reinforcement learning technique it calls CISPO, or Clipped Importance Sampling Policy Optimization. The training rewards search trajectories that find the right documents and penalizes steps that do not produce corresponding gains in retrieval quality.

“We train the agent to search more, but only when needed,” said Michael Bendersky, Research Director at Databricks. “We get a penalty that balances the time spent searching with the reward of finding the right documents.”

What the Benchmarks Show

In one test, Adaptive Instructed-Retriever verified that a company did not list restructuring costs in its fiscal 2022 income statement in two search steps. Claude Sonnet 5 reached the same recall score in three steps. GPT-5.6 Luna took four.

Across seven held-out internal and external benchmarks, the model shows more than 70% improvement in end-to-end answer quality over traditional RAG systems and roughly 15% over sophisticated reranking-based RAG. Retrieval recall gains on instruction benchmarks range from 35% to 50%.

The comparisons are based on Databricks’ own testing and have not been independently verified. The company did not disclose the model’s parameter count, pricing, or general availability timeline. Retrieval is also not a primary focus for the LLM competitors cited in the comparison.

A Family of Checkpoints

Varying the size of the penalty during training produces a range of model checkpoints with different positions on the quality-latency curve. A heavier penalty means fewer search steps and faster responses. A lighter penalty allows more searching and higher retrieval quality.

Enterprises can select a checkpoint based on whether the workload is an interactive user-facing application or a slower offline task. Bendersky said the approach gives operators a defined ceiling on how much searching an agent will do, which “makes agentic search safe to run at scale rather than a runaway meter.”

Where It Fits

Adaptive Instructed-Retriever is a retrieval building block for Databricks’ Genie Code, Genie One, and Genie Agents products. Those systems need to surface the right tables, notebooks, dashboards, and documents without spending excessive time on exploration.

Databricks said its Agent Bricks platform now hosts more than 100,000 agents and processes over a quadrillion tokens per year. The new model is designed to handle the high-frequency retrieval operations that run beneath those agents without requiring the cost and latency of larger frontier models on every request.

Read more: Databricks-trained AI agents match Claude and GPT-5.6 Luna’s answer quality