Databricks adds adaptive retrieval model for AI agents

Databricks has introduced Adaptive Instructed-Retriever, a retrieval model designed for AI agents that can adjust the amount of search it performs according to the complexity of a query.

The model combines parallel retrieval with sequential, multi-step search. It can stop after fewer retrieval rounds when enough evidence has been found. More complex questions can trigger additional searches when further evidence is likely to improve the result.

Databricks said the approach is designed to balance retrieval quality with latency and search costs. The company tested the model across seven held-out internal and external retrieval benchmarks covering different search difficulty levels and domains.

Adaptive Instructed-Retriever achieved retrieval performance comparable to Claude Sonnet 5, GPT-5.6 Luna and DeepSeek-V4-Flash in the tests. Databricks reported an average end-to-end latency of 5.8 seconds, which it said was more than twice as fast as the comparison models.

"Effective enterprise data agents require search that is both accurate and fast. Earlier this year we released Instructed-Retriever-1, a retrieval model that can incorporate enterprise data schemas and custom instructions while using parallel test-time scaling to improve retrieval accuracy with low latency. This single-step search approach works well for a large share of user requests. However, more complex, multi-hop questions can still benefit from sequential search, where the model iteratively gathers evidence and refines its queries over multiple steps – at the cost of additional latency," said the Databricks team.

Databricks trained Adaptive Instructed-Retriever using online reinforcement learning to teach the model when additional search steps are worthwhile.

The training process uses a fixed upper limit on sequential searches. The model can end its search once it determines that sufficient evidence has been gathered. It can also continue searching when additional retrieval is expected to improve performance, subject to the step limit.

Databricks said the training uses Clipped Importance Sampling Policy Optimization, or CISPO. The reward system balances the quality of the retrieval trajectory against the cost of additional search steps.

The model was developed from a pretrained base model using synthetic enterprise retrieval environments. Databricks also reused training data from Instructed-Retriever-1 to retain its ability to perform parallel single-step searches.

Additional synthetic multi-hop questions were introduced to train the model on tasks that benefit from sequential retrieval.

"That's why we're now introducing Adaptive Instructed-Retriever, which combines the speed of parallel retrieval with the performance of sequential search while maintaining strict cost and latency guarantees. The goal is simple: spend additional search steps only when they are useful. This is the same retrieval-efficiency problem faced by Genie Code, Databricks' data agent: it must find the right tables, notebooks, dashboards, and documents in a large, changing workspace without wasting turns on brute-force exploration. Adaptive Instructed-Retriever is designed for this retrieval layer, returning quickly when the evidence is clear and using sequential search only when a harder request requires it. As we show in this post, the trained model matches the quality of leading third-party models at 2x lower latency and improves substantially over single-step search on our retrieval benchmarks," said the Databricks team.

The model is intended for AI agents that search across large collections of enterprise information. These can include tables, notebooks, dashboards and documents.

Retrieval is a key component of AI agents because systems often need to locate relevant internal information before producing an answer or completing a task.

A single search round can be sufficient for straightforward requests. More complex questions can require an agent to gather evidence from several sources and refine its search strategy between rounds.

Databricks said Adaptive Instructed-Retriever is intended to handle both cases without applying the same search process to every query.

The company also demonstrated examples involving negative-answer verification and customer discovery. In one example, the model checked whether a company explicitly reported restructuring costs as an income-statement line item. Databricks said Adaptive Instructed-Retriever reached the same retrieval reward as the comparison systems while using fewer search steps.

In another example, the model was asked which customers were using or had considered using LiteLLM Proxy. Databricks said Adaptive Instructed-Retriever achieved a Recall@10 score of 0.75 after two search steps, compared with 0.62 for GPT-5.6 Luna after four steps and 0.50 for Claude Sonnet 5 after two steps.

Databricks said the reinforcement learning process can be adjusted to produce different quality and latency trade-offs.

The company varies the penalty applied to additional search steps during training. A lower penalty allows the model to use more steps when seeking higher retrieval scores. A higher penalty encourages the model to return results with less search activity.

This produces multiple model checkpoints that can be selected according to the requirements of a particular workload.

"Training Adaptive Instructed-Retriever allows us to choose the quality-latency trade-off by adjusting the magnitude of the step penalty used during ORL. Therefore, we can train a family of checkpoints that powers Adaptive Instructed-Retriever, and pick whichever one fits the production workload," said the Databricks team.

Databricks said its trained checkpoints delivered higher quality at similar or lower latency than the untrained Instructed-Retriever base model across the tested retrieval workloads.

The company is also positioning the model as a retrieval layer for its data agent products, where systems need to search changing enterprise workspaces without spending excessive computation on every request.

"Our contribution provides a practical retrieval building block for data agents at Databricks operating over large, constantly changing workspaces. Experiences such as Genie Code, Genie One and Genie Agents must find the right tables, notebooks, dashboards, and documents without spending excessive time and effort on deep exploration. Adaptive Instructed-Retriever offers a bounded policy for this setting: it is fast on common lookups, while being thorough on difficult discovery tasks, and controllable within a product latency budget. More broadly, these results illustrate the competitiveness of small specialized models even for tasks that require multi-step reasoning. Databricks AIR allows our customers to tailor specialized models for their own domains and performance requirements," said the Databricks team.