Introducing OfficeQA Pro V2: A New Benchmark for Enterprise Grounded-Reasoning
Summary
Databricks introduced OfficeQA Pro V2, a new enterprise grounded-reasoning benchmark generated from roughly 120,000 pages of U.S. Treasury PDFs using an internal synthetic data pipeline. While out-of-the-box agents averaged only 26.0% accuracy on the benchmark, Databricks’ Genie achieved up to 60% accuracy using an optimized agent harness.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from Databricks Blog
Five AI Questions We're Hearing from Financial Services Leaders
Transitioning financial services AI from pilot projects to governed production deployments hinges on addressing five key questions shaping the industry. Learn how Databricks unifies data, AI, and governance across financial crime, banking, and wealth-management workflows, and connect with leaders at Sibos 2026 in Miami.
A practical approach to end-to-end Solvency II reporting in Databricks
Databricks now supports Solvency II end to end, connecting data ingestion, reserving, capital calculation, governance, and disclosure into a single workflow instead of fragmented systems and teams. A demo shows how this unified approach delivers one control view, governed automation, AI-assisted review, and faster scenario analysis for reporting readiness.
Beyond embedding: How to secure AI/BI Dashboards for every viewer
A single AI/BI dashboard can now serve every viewer securely: one entitlements table plus the signed embed token's __aibi_external_value control row-level access per viewer, so teams avoid duplicating dashboards or repeating filters across queries. Access is driven by identity-provider group membership rather than manually maintained user lists, and the pattern layers default-deny protections—masked columns, refused tokens for unentitled viewers, and Unity Catalog row filters—for defense in depth.
Introducing Consort: Test-driven development on a branching database
Consort brings test-driven development to database branching, giving each branch its own isolated copy of data so tests run against realistic state without touching production. The post walks through what changes when your database can branch like code, demonstrates an agentic development workflow built on it, and shows how to try Consort yourself.
Adaptive Instructed-Retriever: Frontier-Quality Search at 2x Lower Latency
Effective enterprise data agents require search that is both accurate and fast. Earlier...
Evaluation-First AI Agents: How Zepto Scales Customer Support on Databricks and MLflow
Zepto built its customer support agents on Databricks and MLflow using an evaluation-first approach, treating traces, golden datasets, and LLM-as-judge checks as core infrastructure and connecting development and production through a dual-loop architecture with a strict quality gate. The result: 65% lower support costs and payback in under a month, giving other AI builders a reusable blueprint for engineering reliability, cost, and risk into production-grade agents.