Databricks Unveils KARL Agent That Cuts Enterprise Search Costs 33% While Boosting Performance

Quick Facts

• KARL agent delivers 33% cost reduction per query and 47% lower latency while matching Claude Opus 4.6 performance on enterprise benchmarks
• Instructed Retriever architecture provides 70% improvement in answer quality over traditional RAG systems
• Knowledge Assistant is now generally available across 10 regions after Databricks expanded enterprise search capabilities

Databricks launched KARL (Knowledge Agents via Reinforcement Learning), an enterprise search agent that the company claims matches top-tier AI performance while cutting costs by a third. The system delivers 33% lower cost per query and 47% lower latency compared to Claude Opus 4.6 on purpose-built benchmarks.

KARL uses reinforcement learning to train across six distinct enterprise search behaviors simultaneously. The agent handles complex tasks requiring up to 200 sequential vector database queries, refining searches and cross-referencing documents before providing answers.

“A lot of the systems that were built for retrieval before the age of large language models were really built for humans to use, not for agents to use,” Michael Bendersky, research director at Databricks, told VentureBeat. “What we found is that in a lot of cases, the errors that are coming from the agent are not because the agent is not able to reason about the data. It’s because the agent is not able to retrieve the right data in the first place.”

The company also introduced Instructed Retriever architecture, which addresses traditional RAG limitations. This system provides 35-50% improvement in retrieval recall on instruction-following benchmarks and 70% improvement in end-to-end answer quality compared to standard RAG architectures.

Databricks built KARLBench to measure performance across enterprise search behaviors including constraint-driven entity search, cross-document report synthesis, and procedural reasoning over technical documentation. Some benchmark tasks exhausted context windows multiple times over before reaching conclusions.

The North Dakota University System reported success using Agent Bricks to build an information extraction agent that parsed legislative calendars, saving 30 days of manual optimization work. “Agent Bricks allowed us to build a cost-effective agent we could trust in production,” said Ryan Jockers, assistant director of reporting and analytics.

Knowledge Assistant is now available in 10 regions including us-east-2, ca-central-1, eu-central-1, eu-west-1, ap-southeast-1, and ap-southeast-2. The system turns documents into cited answers through a fully managed AI agent.

“Enterprises are finding that simple retrieval-augmented generation breaks down once you move beyond narrow queries into system-level reasoning, multi-step decisions, and agentic workflows,” said Phil Fersht, CEO of HFS Research.

The developments represent a shift from model-focused improvements in 2023-2024 toward system-level enhancements in 2026. Databricks trained KARL entirely on synthetic data generated by the agent itself, requiring no human labeling.

Read more: Databricks built a RAG agent it says can handle every kind of enterprise search