Bridging the gap between business and data experts: Databricks launches AI/BI

The opening keynote of the Data + AI Summit 2024 event in San Francisco was packed with new announcements.

Two main threads connect all these together. First, Databricks ' mission to simplify and democratize data and AI for everyone, including non-technical users. Second, a strong belief that state-of-the-art AI results are obtained by compound systems with multiple components working together with focused Large Language Models (LLMs), rather than by building a bigger and general LLM.

This is where the announcement of Databricks AI/BI fits in perfectly. This new product line addresses the main challenges in Business Intelligence (BI) - traditionally, BI requires knowing the questions you want to answer beforehand and building data and semantic models into assets, resulting in long lead times.

AI/BI takes an AI-first approach to solve these challenges, offering everyone the ability to analyze data through two key product experiences: an enhanced Dashboard functionality, and a new conversational interface called Genie.

AI/BI Dashboards (Generally Available)

AI/BI dashboards are an evolution of the previous Dashboard experience, based on feedback and customer pain points.

The most visible update is the inclusion of AI-assisted dashboard building, to meet the increasing demand and speed for more dashboards and reports.

For data teams, view-only shareable dashboards are introduced, to make sure that non-technical stakeholders have streamlined access to the information they need, without requiring workspace access.

Also, the new simplified content model with API support allows them to manage and deploy dashboards as code. With plans for a future integration with Git folders and Asset Bundles, dashboards will be as manageable as any other asset on the platform.

AI/BI Genie (Public preview)

This is where AI/BI really democratizes data for everyone. A business user can "ask questions to the data" using natural language, interacting with the model to refine the analysis. Business users can perform ad-hoc analyses or get metrics that are not available yet in the semantic model.

In their demos and sessions in the event, Databricks has put a big emphasis on how Genie is built with accuracy in mind. Data teams can use different features to improve Genie's reliability and accuracy, as well as learn what their stakeholder are looking to understand:

• Genie Spaces are domain focused collections of tables, sample queries and text guidelines to improve Genie's responses to questions on specific domains;
• When responses are slightly inaccurate, users can edit the SQL code and save it as an instruction for future queries;
• A monitoring page allows users to understand common business questions and i can step through conversations to see how Genie behaved to make improvements;
• Genie is deeply integrated with Unity Catalog and uses metadata to infer information on the available datasets. This means it is even more crucial to have a well governed and documented platform describing the data available!

Furthermore, Genie is built to ask for clarifications when it doesn't understand a query, instead of hallucinating and confidently giving a wrong answer. Users can then add and save new semantic knowledge to improve Genie's understanding - solving for the ubiquitous problem of hidden business knowledge living exclusively inside the head of domain experts.

How is this different from other text-to-sql solutions? There are plenty of vendors in this space, often with impressive demos. However, these are often "bolt-on" solutions that are a thin wrapper on top of a general LLM, fine-tuned using existing semantic models. This approach can work well in isolation, but often falls short in real world scenarios because:

1. Real world data is messy, and general LLMs are prone to hallucinate
2. Business rules are unique and similar concepts differ from company to company
3. Semantic models by their nature are never complete or up-to-date

Following their overarching design philosophy, Databricks has designed Genie as a compound AI system, composed of multiple collaborative agents tuned to specific tasks. This structure enables the system to learn and adapt continuously, while including controls and adjustments for improved accuracy and oversight.

Three other advanced features are planned in the roadmap, to further improve accuracy in the responses:

- Certified Answers: data teams will be able to prescribe consistent behavior to specific mission-critical questions.

- Canonical Metrics will be another option to prescribe behavior, specifically how to query the right tables in the right ways for metrics where rollups are tricky to get right, such as unique user counts.

- Accuracy Evaluation will be a dedicated experience to test questions before sharing.

The road ahead

The new Dashboard experience provides quality-of-life updates that will allow data teams to improve their workflows and better management of dashboards and reports. The AI capabilities of Dashboards and Genie allow for a faster feedback loop between data teams and business users and bridging both the technical as well as the domain knowledge gap between the two.

AI/BI is a much-needed revamp of the Analyst experience in the platform, and an interesting application of LLMs for increased productivity. It promises to make self-service analytics a step closer to reality, and we are excited to see it deliver on this promise.

You can see AI/BI it in action in the keynote demo here Data + AI Summit Keynote Day 1 - Full - starting at [2:39:15] (youtube.com)

If you want to know more, watch two Data + AI Summit sessions Enabling Business Self-Service with Lakeview Dashboards (youtube.com) and Introduction to AI/BI Genie - A New Conversational Analytics Experience for Business Users (youtube.com)