Databricks research shows multi-step agents consistently outperform single-turn RAG when answers span databases and documents

To confirm this is an architecture problem rather than a model quality problem, Databricks reran published STaRK baselines ...