Session

Thinking Fast & Slow: How Databricks Built High-Speed and Deep Research Agents

Overview

ExperienceIn Person
TrackArtificial Intelligence & Agents
IndustryEnterprise Technology
TechnologiesUnity Catalog, Databricks Agents
Skill LevelAdvanced

In today's competitive agentic search landscape, two operating modes are mission-critical. The first is a low-latency, low-cost mode for consumer-facing scale. It must meet strict tail-latency budgets for real-time, high-throughput use cases at minimal per-query cost, without sacrificing quality.

 

The second is a compute-intensive deep research mode enabling expert-level analysis—from financial due diligence and technology mapping to clinical review and manufacturing diagnostics. The system must plan multi-step retrieval, triangulate sources, and synthesize coherent answers.

 

Using our Instructed Retriever and Aroll frameworks, we built a unified agentic harness supporting both modes in a single architecture. Our system sits at the Pareto frontier of cost, speed, and quality. Instant mode delivers single-digit second latency maintaining accuracy, while Thinking mode achieves SoTA-level performance across enterprise domains.

Session Speakers

Speaker placeholderIMAGE COMING SOON

Michael Bendersky

/Director (Research)
Databricks

Full Summary

Thinking fast and slow: How Databricks builds instant and deep research agents on one foundation

Enterprise AI is converging on two user expectations: instant grounded answers for everyday questions and deliberate, multi-step investigations for complex work. The talk argues these modes should not live on separate stacks. A single foundation of shared tools, synthetic training data, and unified evaluation can power both, which now underpins Genie and knowledge assistant experiences on the Databricks Platform. Organizations exploring how this converged approach fits into a broader strategy will find that the Databricks Platform provides the governed, unified foundation that makes both instant and deep research modes practical at enterprise scale.

FAQ


Grounded reasoning couples retrieval with specification-aware reasoning over heterogeneous enterprise data. Specifications such as instructions and index schemas are propagated through query generation and re-ranking. Simple RAG often ignores these constraints, which leads to weaker performance in both single-step and multi-step settings.

A single specialized model shares representations across tasks, simplifies serving, and enables end-to-end optimization. Training first maximizes recall for query generation, then maximizes precision for re-ranking using high-reward trajectories from stage one.

CARL learns context compression as part of end-to-end RL. Compression triggers automatically once a token threshold is crossed, and because the same model performs solving and compressing, the summaries are directly optimized to preserve information that maximizes task reward.

CARL produces long, citation-grounded medical syntheses, resolves multi-hop structured queries that combine several constraints, and pinpoints exact figures in lengthy financial filings by cross-validating numbers across tables, all with the same model.

No. Instructor Retriever 1 and CARL are available inside Databricks products, delivering fast grounded answers in Knowledge Assistant and deep research via a supervisor agent powered by CARL.