Blog — Advancing Analytics

Why Databricks Metric View Relationships form the Cornerstone of the Semantic Layer in the Agentic Era

Written by Ashley Warren | 17/11/2026, 11:43

The Pre-AI Semantic Layer

For most people who worked in traditional business intelligence, the goal was to build a suite of reports that provided your business users the insight they need to affect change on your organisation. Much of the business meaning behind those reports, the metric definitions, aggregation logic and the relationships between business events, often sat within the reporting tool.

The need for this context, that collectively we call the Semantic Layer, is why pretty much every reporting tool evolved ever more powerful capabilities for curating data. In my opinion, the pinnacle of this in reporting tools can be seen in the Power BI Semantic Model. The Semantic Model is the innovation that enabled Power BI to reign as the de-facto number 1 analytics tool for nigh on a decade, appearing as the most top right entry in every Gartner Magic Quadrant for Analytics and BI from 2018 to 2025.

... But then came AI.

The Conversational Analytics Revolution

The focus of analytics is currently shifting from serving insight solely through highly curated reports, to tools that enable insight to be offered in response to natural language questions.

The biggest complaint from users about reports was always that the one you’d built for them didn’t answer question X, which normally was something they hadn’t thought of asking for when you scoped the report. As an industry, we responded to this by trying to predict every question a user could ask about their data, build models that could enable us to answer them all and then build ever more new reports to serve those answers. Effectively, we created an endless dashboard sprawl, which, despite best intentions, often left users complaining about dashboards not being fit for purpose and led people to clamour for the death of dashboards.

Enter Conversational Analytics. Ask a question, the AI evaluates it, does some research, writes a query against your data, produces a summary and some visuals and responds with the answer (or at least some approximation of what that answer might be). These tools have progressed massively in the last 2 years, with the crème de la crème, in my opinion at least, being Genie One and Genie Agents in Databricks. However, as with so many areas of our industry (and countless others), this new AI use-case necessitates an evolution of what has come before. To enable conversational analytics, how we create and manage our Semantic Layer must change. Some of those changes come from demands of the technology itself, such as the increased need for context of what our data is and what it means to live within the system, rather than solely in users’ heads. But many of the changes stem from the behaviour that this new technology elicits in us as humans.

There are two key characteristic that have enabled humans to develop into the pre-eminent species on our planet. The first is our boundless curiosity. As a generalisation, we have an innate drive to find and answer those groundbreaking questions that are unique to us and our experience of the world around us. Combine that with the second key characteristic, our exceptional pattern recognition skills. Set that in the context of our businesses and the questions that we ask often span multiple domains and business events that potentially aren’t naturally connected.

For example, a human might make a link between a specific instance of poor customer service for a client and the ability of the account team to get a contract renewal over the line. They then ask, “how widespread has this issue been and what is the financial impact of it?”. These are events from two different domains of the business, and so the first time this is asked, the analyst likely has to create a new semantic model with those facts and link them. When we start to look at predicting the future impact of this issue, maybe the question strays beyond traditional reporting and into the realm of data science. Maybe it expands further as you investigate the cause of the customer service issue and find that it was caused by a supply chain issue which in turn was caused by an issue with paying your suppliers. Next, you need to act on it and end up feeding that insight back into your system of record

I’ve seen this kind of thing happen multiple times in my career. At this point, the poor analyst is maintaining a semantic model that spans most of your business, supports reporting and data science models, and has become critical to day-to-day operations. You naturally ended up in the dreaded scenario for BI teams; One semantic model to rule them all. Or more aptly, one model that performs horribly, and is quickly becoming critical business infrastructure. This is the worst-case scenario for traditional semantic models.

And there-in lies the case for a tool like Genie One. You define each business event independently and set the rules of how to join them together using Relationships. Genie One can then use all that context to bring the relevant information together when a question is asked, in the appropriate manner to solve the problem. By understanding the required outcome at query time, Genie can do this in a more natural way than a pre-defined semantic model, that's trying to solve every problem for every user, ever could.

Tangent resolved. Context loaded. Back to the Title of the Blog

At this point in the blog, you are potentially wondering “how has the author strayed so far from the blog’s title?” But I assure you, it is all relevant. To draw a parallel to how Genie One works, I was pre-loading us with all the context we’d need so that, when we ask the core question of the blog “Why do Databricks Metric View Relationships form the cornerstone of the Semantic Layer in the Agentic Era?” we have the background to understand the answer.

Fundamentally, Metric views represent business event and define your metrics over them, much like fact tables do. You establish the grain of the underlying data, the relevant dimensional fields, how those fields & sources fit together, and how metrics are calculated. You also explain when to use that metric view as part of the description.

When a user asks a question, Genie One reads the descriptions and contents of metric views to identify relevant data that will help it answer. If the question spans multiple domains and metric views, like the sprawling example from earlier in the blog, it may need to combine the metric views together.

Before explicit Metric View Relationships, Genie One had to infer how those views fit together from the available context. However, as with most things in the semantic layer, sometimes it’s not as simple as it initially seems, and these inferences from Genie were incorrect. This formed the fundamental gap in the metric view proposition.

With Metric View Relationships, we can now describe how metric views relate to one another, as part of their definition. We use metadata to specify the cardinality and join conditions between different metric views. Then our agentic systems can then just use them correctly without guessing. At first, it might seem like an inconsequential feature. There is very little historic reason for needing this within the context of human-driven semantic layer design, but to me it is essential to enabling conversational analytics that the semantic layer is traversable for an agent. This is context that traditionally lived in the minds of the analysts and was embedded into the design choices we made. But it now needs to be shared with the agentic system, and I think it will trigger a fundamental rethink of how we design a semantic layer.

Previously, we defined the semantic layer based on multi-fact table style models, that were built around how a human analyst might need to pull data together for reporting. It was focused on domains, human usability and partially relied on our ability as analysts to overcome gaps in context with our problem-solving skills. That doesn’t work in the same way for agentic systems, and we will need to adjust our design philosophy accordingly.

In the future, we will build individual metric views modelling single facts around specific business events and define the relationships between them. The system will then use those relationships to combine the relevant views as needed. Put simply, the semantic layer is set to become much more modular than it ever was, with a better way to bring those modules together.

To me, Metric View Relationships are a cornerstone feature of this vision for the future of the semantic layer. And once again, Databricks are in the vanguard, leading us towards that future!