Powering Sigma Agents with Data Models, Warehouse Search & Usage Insights

Earlier this month during our “Month of Agents” series, we introduced the new Agent Builder Assistant (“ABA”) for building Sigma Agents and how to govern and permission them. This week, our team shows you how data models, warehouse search services, and observability come together to power Sigma Agents with live, trusted data. In this week’s episode, you’ll see how to:

• Connect a data model directly as an agent’s data source, so it inherits relationships, metrics, and governance already defined in your warehouse
• Add a warehouse search service as a tool, for fast, accurate answers over unstructured or text-based data
• Monitor agent usage and cost with a secure AI usage dashboard stored in your own warehouse

Watch the video above to see it all in action, or read the accompanying transcript to understand how our Product and Engineering teams approached data models as agent sources, added search services as tools, and built observability into Sigma Agents.

As a public beta feature, Sigma Agents are accessible by default to any Sigma customer with an active AI provider enabled. Reach out to your Sigma representative with any questions. If you’re not a Sigma customer yet, you can request a demo or start a free trial to get started.

Our “Month of Agents” series concludes next week with a roadmap for Sigma Agents, highlighting the transition toward more autonomous, complex agent behaviors and future capabilities. Follow Sigma on LinkedIn, X, or YouTube to see the updates as they come.

Read on to see the full video transcript.

The fuel for Sigma Agents: data, search, and observability

Zalak: Welcome, everyone. My name is Zalak Trivedi. I’m a Director of Product at Sigma. This is episode 3 of our agent series. We learned about how to build agents, and how to govern and permission them, over the last couple of weeks. This week is all about the fuel: the data sources, the warehouse search services as tools, and observability.

With me, I have Sigma Software Engineers Rounak Salim and Madison Chamberlain. Let’s see some demos.

Madison: Definitely.

Demo: How do data models power Sigma Agents?

Rounak: For this example, I have a CodeProvisions data model set up for my store, with various tables covering orders placed, inventory, promotions, and so on. The one we’ll primarily work with is the orders fact table, which has a list of all the orders.

Accompanying this, I have an empty workbook that doesn’t have any data in it, except for a bare-bones agent. Once we open that up, we can see the agent’s instructions are fairly simple: answer questions using the data it has access to. But you’ll notice we don’t have any data sources set up, so let’s do that.

We now have the option to pull in data models that are already set up inside Sigma. Let’s select the data model we were just looking at, and add just the orders fact table we want to work with. I’ll also update the greeting so the agent tells us what relationships and metrics it has access to.

Once we save that, we can see the agent telling us it has access to the orders data, the table we added. But it also has access to other information, such as products, stores, and promotions. That’s being pulled in from the relationships and metrics already set up in the data model.

Let’s ask the agent how much in sales our promotions have contributed, and which one performed best. While that’s running, if we go back to the orders table, we can see we don’t actually have any promotion information here, other than a campaign ID.

But you’ll notice we do have relationships set up, and this one has a relationship on the promo campaign table. We also have metrics around promotion sale numbers and sale values that the agent can use to solve this problem. If we look at the promo campaigns table, we can see there are a couple of different promotions we’ve run. Let’s see if the agent was able to use the relationship and this data to answer our question.

Switching back, we can see the agent gave us an answer about our total sales via promotions, and broke down what sales look like for each individual promo. It also told us the best-performing promo is the Meta Event promo. As you can see, we were able to set up an agent just by giving it access to the data model, and in turn it was able to use all the relationships and metrics already set up in that data model.

Why use data models as a source for Sigma Agents?

Zalak: This is a big shift, being able to use data models as a source. What’s it about, and why is it useful for our customers?

Rounak: Great question. Up until now, when users wanted to build agents, they had to pull all the raw tables into their workbook to provide that as context for the agent.

That puts a lot of pressure on the user, because organizations can have hundreds of thousands of tables, if not more. The user has to figure out which tables are appropriate, which ones are live and kept fresh and up to date, and which represent the ground source of truth, then pull those in.

That problem has already been solved. We have data models, which are essentially Sigma’s semantic layer. They let users organize and store tables in a structured, governed way, so you always have a source of truth that represents data that’s fresh, up to date, and ready for analysis.

So now, agents can pull in data models directly as data sources, without users needing to pull those tables in separately into a workbook.

Zalak: And with our warehouse partners, that means Snowflake semantic views and Databricks Unity Catalog metrics, the semantic layer defined at the warehouse level, can be pulled all the way through to the agent.

Rounak: Exactly. All that setup you’ve already done can now be used by agents.

How do metrics keep agent answers grounded?

Madison: And what about metrics? I know we’ve added additional support for that as well.

Rounak: Great question. Data models support things like relationships, metrics, and AI context, and metrics are a core part of this.

Metrics represent reusable business definitions that are central to your business. LLMs are phenomenal tools with a great general worldview, but they struggle with internal context. For example, something like ARR might have one definition at one organization and a different one at another, and an LLM can’t infer which definition applies without that context. That’s what metrics provide.

So when you pull in data models that have metrics attached, an agent can see all of those metrics. When a question touches on something represented by that internal context, the agent uses the appropriate metric instead of reasoning through its own calculation.

It uses the ground truth, which gives users a more grounded and accurate answer.

Zalak: So essentially, you’re inheriting governance all the way from the warehouse and bringing it into the agent as a source. That’s a pretty powerful feature.

What are warehouse search services, and how do they work as tools?

Zalak: Madison, I know you’ve been working on more sources around tools, specifically the warehouse search service. Let’s talk about that.

Madison: Warehouse search services are a huge unlock, mostly for speed. SQL is great for things with a deterministic answer, where you want to filter or complete a known task, but it’s not as fast or performant for something less defined.

Say I have a transcript and I want to ask, “How is my feature being used?” and I want to look through all the transcript calls. That’s not something you can do with SQL. But with a vector embedding, like a warehouse search service, it can look through all of that text-based data and return an answer in milliseconds.

You get accurate answers in milliseconds, in a way an LLM couldn’t previously do in a timely manner.

Zalak: Got it. And these search services can be brought in as tools?

Madison: Yes, exactly.

Demo: Adding a warehouse search service to an agent

Madison: Here I have a Sigma agent. I’m going to click into it and select tools to add a warehouse search service. I’m going to select the Gong call search service.

Zalak: And now you can immediately query it from your agent.

Madison: Now I’ll ask the agent how customers feel about being able to add a warehouse search service to their Sigma custom agent. As you can see, the agent is using the search service tool to find Gong calls and run sentiment analysis.

Results come back extremely quickly and accurately.

Zalak: Today we support the Cortex search service, and the Databricks version which is Vortex search service, as part of tools.

Very cool.

How do you monitor agent usage and cost?

Zalak: Let’s talk about observability, a very important part for our customers. I know you’ve been working on usage metrics around agents. Tell me about that.

Madison: We found that a lot of our customers had questions like, “how much is this costing me?”, “how many people are using the agents I’m building?”, and “how are the agents I’m building being used?”

We wanted to provide a way that was both secure and easy for customers to answer those questions on their own. So we built an AI usage table, where we store the data we’re collecting in the user’s warehouse. That covers everything from tokens used, to the tools being called, to the user’s thumbs up or thumbs down feedback and reasoning.

You can ask pretty much any question you want about how AI is being used by Sigma, or by you within Sigma, and see the answers in the AI usage dashboard.

How does Sigma keep usage data secure?

Zalak: I want to double-click on the security aspect of this. You mentioned we store the data in the customer’s warehouse.

Madison: Yes.

Zalak: Talk to me a little bit about that. How does that work?

Madison: We’re in charge of the schema, so we know what we’re storing, and we emit the events, so we’re storing it for the customer. But it’s purely in their warehouse, so we don’t hold any of the data. We don’t touch any of it. It just passes right through.

So it’s your data, you own it. We never see it, but we put it in the correct format for you.

Zalak: That’s a really secure way to show users metrics, where we control the template, and the data lives in the customer’s warehouse.

Demo: Inside the AI usage dashboard

Madison: The AI usage dashboard gives you insight into how Sigma’s AI is being used by your organization. You can get an overview of how many tokens and questions are being asked, and who’s asking what, and dive into specifics on how features like agents are being used. If you have more questions you want to ask, or specific ways you want to query the data, you can add this to your workspace and alter the workbook as you see fit.

This is your data, so you can own the product experience as well.

Zalak: Over the last 3 weeks, we talked about building agents, governing and permissioning them, and using warehouse agents as tools. This week, we covered the fuel: data models and warehouse search services as sources, and observability into how agents are being used.

Next week, we wrap the series with Jisha, who is the Product Manager on the agents team, and Chris Messer, who is a Tech Lead on the team. Thank you, Rounak and Madison, for this week.

Rounak: Thank you.