Databricks adds adaptive search model to speed agent retrieval

Databricks has expanded its Adaptive Instructed-Retriever search model to help AI agents move faster when a query takes more than one round of retrieval. That matters because once an agent has to refine its search, every extra step adds delay and cost. Databricks says the new model is meant to handle that tradeoff automatically instead of forcing developers to choose one speed setting for everything.

The company is pitching the model as a retrieval building block for Genie Code, Genie One and Genie Agents. In Databricks’ own testing, it matched the retrieval quality of several leading third-party and open-source models while answering twice as fast as Anthropic’s Claude Sonnet 5, OpenAI’s GPT-5.6 Luna and DeepSeek’s V4-Flash. Those results came from seven held-out benchmarks covering different domains and search difficulty levels, and they have not been independently verified.

Adaptive Instructed-Retriever builds on Instructed-Retriever-1, a model Databricks released earlier this year to carry user instructions, examples and data-source schemas through retrieval and generation. That earlier system focused on parallel, single-step searches and, by Databricks’ account, improved performance by more than 70% over traditional RAG on enterprise question-answering tests. The new model keeps that single-step ability, but adds a way to decide when another search is worth the time.

Developers can set a maximum number of sequential steps, and the model decides how far to go on each request. It stops when the evidence looks sufficient and keeps searching when another pass should improve the answer. Databricks trained it on synthetic enterprise retrieval environments and multi-hop questions, then used online reinforcement learning with Clipped Importance Sampling Policy Optimization. The reward system pushes accurate search paths and penalizes extra steps that do not add much value.

The company says that penalty can be tuned to create different checkpoints along the quality-latency curve. A heavier penalty should favor fewer steps and quicker answers; a lighter one should allow more searching. Databricks did not share the model’s parameter count, pricing or general availability, but said it was trained with Databricks AI Runtime and can be specialized for customer data and performance needs.