Databricks has developed a new retrieval-augmented generation (RAG) agent named KARL, designed to simultaneously tackle various enterprise search tasks that traditional RAG systems struggle to handle. Most enterprise RAG pipelines are optimized for one specific search behavior, leading to failures during multi-step reasoning and cross-document tasks, which are prevalent in complex internal data environments. By employing a novel reinforcement learning algorithm and the KARLBench benchmark, Databricks claims that KARL can perform six different types of enterprise searches at a lower cost and with reduced latency, demonstrating how multi-task training can yield generalizable knowledge across diverse retrieval behaviors. The innovation also showcases advances in off-policy reinforcement learning, enabling more efficient training by reusing previous data, thereby addressing significant challenges in enterprise search capabilities.
Databricks unveils KARL, a multi-task rag agent for enterprise search optimization
