I wanted to learn Fabric so I picked the simplest free API I could find and spent the Easter evening seeing how far I could take it.
See the outcome first 🙂
I can open a Fabric Data Agent and just… ask questions.
Things like:
- If I had to pick one city to visit this week purely based on weather, which one would you recommend and why?
It answers instantly. No SQL. No dashboards. No training required.
Why weather data
The Open-Meteo API is completely free. No account & no API key. Clean JSON out of the box.
It was the fastest way to get real data flowing without authentication.
Three cities: Melbourne, Sydney, Brisbane. Current conditions, hourly forecasts, 7-day outlook.
What I learned building it
Fabric uses a medallion architecture, Bronze, Silver, Gold. I’ve heard about this pattern before and it now makes sense how each layer serves a different purpose.
Bronze is just the raw API response saved as JSON. No transformation, no cleaning. This is your safety net, if anything breaks downstream, the original data is always there.
Silver is where the transformation happens. I won’t pretend I wrote the PySpark from scratch. I used AI to help with the code. But I understood what it was doing, debugged the errors when things broke, and made the calls on what the data should look like. Three Delta tables came out of this layer, current conditions, hourly forecasts, and daily forecasts.
Lastly, for Gold, instead of hundreds of hourly rows, you end up with one clean summary per city per day. Average temp, total rain, max wind, humidity, visibility. Everything a Data Agent needs to answer questions quickly without querying the entire dataset.
The Data Pipeline schedules all of this to run automatically every night at 11pm. Set it once, forget it.
From the Gold table, one click and you have a Semantic Model. Honestly I always nodded along when people said that term in meetings. Now I get it, it is basically the bridge between your raw tables and whatever is reading the data. Power BI, a Data Agent, anything.
Power Bi or Fabric Data Agent
From the Semantic Model, one click takes you to either a Power BI report or a Fabric Data Agent. I went with the Data Agent.
A few things stood out to me about it:
- You can add Azure AI Search indexes as a knowledge source so if you have unstructured data like documents or PDFs, you can bring that into the same agent alongside your structured tables
- One-click publish and deploy straight into Microsoft 365 Copilot, no separate setup, it just shows up where your users already work
- Readily available MCP server connectable to Copilot Studio for custom flows, or the pro code route via Azure if you need more control
Why this matters for enterprises
Most organisations already have data sitting across multiple systems, invoice data in their ERP, contracts in SharePoint, sales records in Salesforce, support tickets in ServiceNow, financial data in SAP.
It is rarely in one place and rarely in one format.
Fabric’s medallion architecture gives you a clean way to automate and schedule the ingestion of all of that into OneLake, structured, cleaned, and ready to query without having to rip and replace existing tools.
Alternatives
If you are already in the Microsoft ecosystem, Azure, Teams, Power BI, M365, Fabric is a very natural starting point. Everything talks to each other out of the box.
For someone like me building data engineering foundations, that lower friction matters. I could focus on learning the concepts rather than figuring out how to connect everything together.
Keen to try Databricks and Snowflake next and see how they compare 😃
Resources
If you are also learning to build in Fabric, I am sharing the PySpark notebook here, hopefully it saves you some of the trial and error I went through.
github.com/ifiecas/australian-weather-fabric
Happy learning~~ 🙂