Databricks Adaptive Instructed-Retriever doubles speed vs Claude Sonnet 5

Databricks' announcement builds on its earlier Instructed-Retriever-1 release, which aimed to improve on traditional retrieval-augmented generation by carrying user instructions, examples and data-source schemas through the retrieval and response-generation process. That architecture reportedly improved performance by more than 70% over traditional RAG on enterprise question-answering tests. The new model extends this line of work by addressing the latency and cost of multi-step searches, a task that typically requires an agent to refine queries as it gathers evidence, with each step increasing response time and computational cost.

In a test described by Databricks, Adaptive Instructed-Retriever verified a company did not explicitly list restructuring costs in its fiscal 2022 income statement in two search steps, while Claude Sonnet 5 took three and GPT-5.6 Luna took four. The company trained the model using synthetic enterprise retrieval environments and applied online reinforcement learning, with a reward system that favors accurate search trajectories while penalizing additional steps that fail to produce quality gains. Databricks also noted that customers can use its AI Runtime to specialize smaller models for their own data, suggesting that task-specific models can compete with frontier systems by learning when further searching is worth the delay.

The stakes for Databricks and its customers revolve around the quality-latency tradeoff. The model's family of checkpoints offers choices along that curve, but the company has not disclosed parameter count, pricing, or availability. The claims also rely on Databricks' own benchmarks, so independent verification is pending. How well these models perform in real enterprise environments will likely determine whether this approach gains traction among developers building data agents.