Databricks adds adaptive search model to speed agent retrieval

Databricks has introduced an adaptive search model to expedite retrieval for AI agents necessitating multiple rounds of data retrieval. The new model enhances the quality of search results while reducing response times, outperforming competitors like Anthropic’s Claude Sonnet 5, OpenAI’s GPT-5.6 Luna, and Hangzhou DeepSeek’s V4-Flash.

Built upon the Instructed-Retriever-1 model, the adaptive model optimizes search processes by determining the necessary sequential steps, stopping when the available evidence suffices, or continuing when a subsequent round is likely to improve the answer. This technique caters to data agents operating in extensive, evolving workspaces, ensuring efficient and cost-effective information retrieval.

The adaptive model is trained using synthetic enterprise environments, multi-hop questions, and reinforcement learning techniques, with adjustable penalty levels that enable a spectrum of model checkpoints balancing accuracy and latency. Databricks has not disclosed specific pricing or availability details, but the model demonstrates capabilities such as verifying the absence of specific information in datasets using minimal search steps, rivaling the performance of leading AI agents.