Databricks KARL: AI Learns Enterprise Search Skills

Databricks has introduced KARL, an innovative AI agent designed to master complex enterprise search tasks through reinforcement learning. Unlike traditional RAG systems optimized for a single behavior, KARL simultaneously handles six distinct enterprise search functions, including entity search, report synthesis, and reasoning over internal notes.

This advanced agent matches the performance of leading models like Claude Opus 4.6 on specialized benchmarks, but with a substantial reduction in query costs and latency. A key innovation is KARL's ability to train entirely on synthetic data generated autonomously, bypassing the need for costly human labeling. This approach allows it to tackle the nuanced, non-verifiable data typically found in enterprise environments.

KARL's training leverages a novel reinforcement learning algorithm, OAPL, developed by researchers from Cornell, Databricks, and Harvard. This method is highly sample-efficient, enabling comprehensive training within a feasible budget. The agent also incorporates advanced context management, compressing information to navigate extensive data without losing accuracy.

While KARL demonstrates remarkable capabilities, it still faces challenges with highly ambiguous queries where multiple valid answers exist. Future developments aim to expand its functionality beyond vector search to include SQL queries and calculations, further enhancing its utility for enterprise data teams.