Back to Blog

Enriching a Fabric IQ Ontology - Adding Entities, Bindings and Time Series Data

November 23, 2026•7 min read•Michael Ridland

The first time most teams build a Fabric IQ ontology, they generate it from an existing semantic model and feel quite pleased with themselves. Then they try to point an AI agent at it and realise the ontology only knows what the Power BI model knew. Sales, products, stores. The operational data that actually explains why sales dropped in one store last Thursday lives somewhere else entirely.

That gap is what the second part of Microsoft's ontology tutorial covers. The Enrich the ontology tutorial takes the ontology you generated in part one and extends it with new data that was never in the semantic model, including streaming telemetry from an eventhouse. I've been through it a few times now, both on my own and with clients, and I think it is the most useful part of the series because it shows what an ontology can do that a semantic model can't.

Here is what the tutorial walks you through, and what I would add from our project work.

Why enrichment is the point

A semantic model is built for aggregation. It answers "what were total sales by region last quarter?" very well. An ontology is built to describe things and how they connect. A store has freezers. A freezer has a temperature reading every minute. A temperature excursion affects the products stored in that freezer, which affects sales and waste.

You can't easily express that chain in a star schema. In an ontology, it is just entity types, properties and relationships. Once it exists, an agent can follow it. Ask "why was dairy waste high at the Parramatta store?" and an agent grounded on the ontology has a path from the store to its freezers to their temperature history.

So enrichment isn't a nice extra. Without it, your ontology is a slightly different view of your Power BI model, and you could have just pointed the agent at the semantic model.

Step one - add a new entity type

The tutorial has you create a new entity type for something that didn't exist in the semantic model. In the retail scenario it's equipment in the stores, specifically freezers. You give the entity type a name, then define its properties.

The decision that matters most here is the key. Every entity type needs a property, or combination of properties, that uniquely identifies each instance. For a freezer, that is the asset ID. It sounds trivial. It isn't. On a client project for a distribution business, we spent half a day working out that their "asset ID" was unique per site, not globally, so we needed a composite key of site and asset. If you get the key wrong, relationships quietly join the wrong things together and your agent gives confident wrong answers.

My rule: before you define a key in the ontology, run a distinct count query against the source table and prove it is unique.

Step two - bind static properties to a lakehouse table

An entity type with no data binding is just a definition. The tutorial has you bind the new entity type to a lakehouse table that holds the static details: asset ID, model, store, install date and so on.

The binding experience is reasonably straightforward. You pick the lakehouse, pick the table, then map source columns to the properties you defined. Where it gets fiddly is naming. If your lakehouse columns are called asset_id and store_cd, and your ontology properties are AssetId and StoreCode, the mapping is manual. Not hard, just tedious on wide tables.

One thing I'd flag from earlier work: binding behaviour depends on workspace networking. In previous rounds of testing, bindings to lakehouses in workspaces with inbound public access disabled didn't behave the way you'd want. Check the current documentation for your setup before you assume a locked-down production lakehouse will just work.

Step three - add time series properties from an eventhouse

This is the interesting part. The tutorial has you bind a time series property to data in an eventhouse, which is where real-time telemetry lands in Fabric. Each freezer gets a temperature property that isn't a single value but a history of readings over time.

To set it up you choose the eventhouse table, map the column that identifies the entity (the asset ID again, which is why the key matters so much), map the timestamp column and map the value columns.

When this works, it's genuinely good. You end up with a single entity that has both its static description from the lakehouse and its live behaviour from the eventhouse. That combination is exactly what operations teams want and almost never get, because the two kinds of data usually live in completely different systems with completely different owners.

A few practical notes:

  • Get your timestamps in UTC at ingestion. Mixing local times from devices across Australian time zones and daylight saving is a mess you don't want to debug inside an ontology.
  • Think about volume. Telemetry every few seconds from thousands of devices adds up quickly. The ontology doesn't copy the data, it binds to it, but queries still hit the eventhouse, so retention and caching policies matter.
  • Make sure the device identifiers in the telemetry actually match the asset IDs in the lakehouse. On one project, the IoT platform used MAC addresses and the asset register used internal codes. We needed a mapping table before any of this could work.

Step four - create the relationship

Finally the tutorial has you add a relationship type between the existing Store entity and the new Freezer entity. You define the relationship, then bind it to data that shows which freezer belongs to which store, usually a foreign key column in the freezer table.

Name relationships the way people talk. "Store has Freezer" or "Freezer located in Store" reads well when an agent explains its reasoning. Generic names like "Relationship1" (which I've seen in real ontologies) make everything harder to read and debug.

After this step you can explore the graph in the ontology view: click a store, see its freezers, click a freezer, see its temperature history. That moment is usually when stakeholders in a demo start leaning forward.

What's still rough

Fabric IQ is still maturing, and I'd be doing you a disservice if I pretended otherwise.

The authoring experience is fine for a tutorial-sized ontology and gets clunky when you have dozens of entity types. There's a lot of clicking. I'd like better bulk editing and better support for treating ontology definitions as code, so they can be reviewed and deployed through proper pipelines.

Error feedback when a binding fails can be thin. You sometimes find out a binding is wrong only when a query returns nothing.

And the bigger challenge isn't technical. Ontologies force you to agree on what things are. What is a "customer" in your business? Is a franchise store a store? Those conversations take longer than the build, and they should. Skipping them gives you a pretty graph full of ambiguity.

Should you bother?

If you're planning to put AI agents in front of operational data, yes. An enriched ontology gives those agents a map of your business that's far more useful than a pile of tables. If all you need is dashboards, stick with your semantic model.

We help Australian organisations design ontologies and connect them to agents that people actually use. Our Microsoft Fabric consultants handle the data platform side, and our enterprise AI agents team builds the agents that sit on top. If you're still working out where an ontology fits in your plans, an AI strategy session is a good place to start.

Do the tutorial yourself first, with the sample data. It takes an afternoon, and you'll come out of it with a much clearer sense of whether this fits your organisation than any slide deck will give you.