Databricks tested a stronger model against its multi-step agent on hybrid queries. The stronger model still lost by 21%.

Databricks tested a stronger model against its multi-step agent on hybrid queries. The stronger model still lost by 21%.

Databricks Tests Multi-Step Agent Against Hybrid Queries, Still Falls Short

Databricks, a leading cloud computing platform, has conducted research that tests a stronger model against its multi-step agent on hybrid queries. The results show that the stronger model still lags behind by 21%. This comes as no surprise to data teams building AI agents, who often struggle with questions that require joining structured data with unstructured content.

Background & Context

The issue at hand is that single-turn Retrieval-Augmented Generation (RAG) systems, which are commonly used in AI agents, fail when faced with hybrid data tasks. These tasks involve mixing precise structured filters with open-ended semantic searches. For instance, a question might ask for sales data from the past three months and related customer reviews. This is a common challenge for enterprises, which often struggle to answer such questions using current agent architectures.

Databricks' research team has been working on addressing this issue. Their previous research on instructed retriever models showed improvements in unstructured data retrieval using metadata-aware queries. The latest study builds on this work by incorporating structured data sources, relational tables, and SQL warehouses into the reasoning loop. This allows the agent to consider both structured and unstructured data when answering hybrid queries.

Impact on Swiss SMEs & Finance

The implications of this research are significant for Swiss SMEs and the finance sector. Many businesses in Switzerland rely on AI agents to analyze complex data and make informed decisions. However, these agents often struggle with hybrid data tasks, which can lead to inaccurate or incomplete results. By developing more effective multi-step agents, Databricks is helping to address this issue and improve the accuracy of AI-driven decision-making.

What to Watch

The next step for Databricks will be to refine its multi-step agent and continue to test its performance on hybrid queries. The company's research team will likely work on addressing the remaining 21% performance gap and exploring new applications for this technology. As the AI research community continues to evolve, it will be interesting to see how Databricks' work on multi-step agents contributes to the development of more effective AI agents.

Source

Original Article: Databricks tested a stronger model against its multi-step agent on hybrid queries. The stronger model still lost by 21%.

Published: April 14, 2026

---

Disclaimer: This article is for informational purposes only and does not constitute financial advice. Consult a licensed financial advisor before making investment decisions.