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
Read, don't copy: replacing file feeds with OpenSharing - Part 2 (How to Implement)
OpenSharing lets you expose a Delta table through an open protocol so recipients read it live, replacing per-recipient file and API exports along with the file tracking and reconciliation they require. This implementation guide shows recipients consuming the shared table as a checkpointed incremental stream, receiving row-level inserts, updates, and deletes through Change Data Feed.
Read, don't copy: replacing file and API feeds with OpenSharing - Part 1 (The Why)
Instead of copying customer data via SFTP drops or API pulls, you can read the customer's shared table in place with OpenSharing and follow its changes through streaming or change data feed, with no need to track what's been processed yourself. This removes the per-feed pipelines, fragile incremental loads, backfills and duplicate storage, and it makes sharing easier for customers, since adding you as a recipient is just a grant on a table they already maintain, with audit logs recording every access.
How public sector organizations can govern AI at scale
Unity Gateway lets public sector teams govern AI the way they already govern data in Unity Catalog, with controls for model access, spend caps, and request-level tracking. Databricks also manages data residency for its hosted models by geography, giving compliance teams a documented answer on where AI requests are processed.
Real-time fraud detection with Feature Views
Feature Views let you define fraud features once and reuse them for both point-in-time training and online scoring, which reduces data leakage and training-serving skew. The example pairs rolling windows for payment bursts with a 30-day Sawtooth baseline and an on-demand feature that scores each payment against it, and a single materialize_features() call sets up managed streaming pipelines into the Online Feature Store on Lakebase with sub-200ms feature freshness.
A builder’s guide to scaling Databricks Apps
Scaling Databricks Apps starts with agreeing on target response times and peak user counts, then routing heavy work to the right service while app compute handles the UI and coordinates calls. Keep code responsive and state recoverable by caching per user and storing hard state in Lakebase or Unity Catalog tables, then size instances, workers and Lakebase connections from load tests and telemetry.
Which Data Collaboration Tool: OpenSharing, Clean Rooms, Marketplace
Databricks offers three ways to share data and AI assets, and the right one depends on your relationship with the other party. Use OpenSharing for live, read-only access with a known recipient, Clean Rooms when partners want to combine data without exposing all of it, and Marketplace to reach or find audiences you don't know yet.