Blog | FELD M

Switching BI tools in the AI world: how good is Apache Ossie? | FELD M

Written by Sandy Bekheet | Nov 20, 2026, 2:35:58 PM

After FabCon, my colleague Ataberk Teknekaya came back with the news that Microsoft FINALLY agreed to be a part of Apache Ossie. It’s welcome news, as some of our clients are already looking to switch to Power BI and would also like to bring the business logic from their current BI tool into Power BI semantic models.

If you haven’t come across Ossie yet, it’s an open standard for semantic models. It started out as Open Semantic Interchange (OSI) and was renamed Apache Ossie when it was accepted into the Apache Incubator in June. Not to be confused with Aussie, which is a brand of curly hair products that I discovered because of the name!

The semantic model is the layer between the data warehouse and the dashboards. It’s where you decide which tables exist, how they connect and what “an order” or “revenue” actually means. Every BI tool stores it in its own format, and that’s the problem Ossie is trying to solve. Instead of translating directly from one tool to another, you translate into Ossie first and then into the target tool.  

Snowflake, Databricks, and Salesforce were already part of it, but Microsoft stayed out for about a year before joining this October. The converter is still in beta, but I still wanted to try it out and share the outcome with you.

I chose Omni for this experiment for two reasons. Firstly, it’s a great tool and thanks to our partnership, we have sandbox access; and secondly, I had already modelled data and a semantic model in Omni, which gave me a clear picture of the test I wanted to run: take one of our Omni models, push it through Ossie and see if it comes out the other side as a working Power BI model.

 

The setup: Testing whether Ossie can produce a working Power BI model

Because Ossie sits in the middle, the test needed two converters: the Omni one takes the model into Ossie, and Microsoft’s brand-new converter takes it from Ossie into Power BI. Both live in the apache/ossie-converters repository and install as two command-line tools, ossie-omni and ossie-microsoft.

For the model, I kept things small on purpose. The model covers 11 tables, with 3 relationships and 11 measures, reading from a dbt marts dataset in Google BigQuery. Every measure is a simple row count that Omni creates automatically. My thinking was: if the converters can’t handle this, there’s no point trying anything bigger.

To get the Omni files onto my laptop, I synced the model to a GitHub repository and cloned it. You can see two files below: models and relationships in YAML. 

 
Screenshot: Omni model synced to GitHub

 

The easy part: Omni to Ossie

In Omni, every table is described in a small YAML file called a view. This is the one for our orders table, shortened a bit:

 

yaml schema: dbt_prod_analytics_marts table_name: orders dimensions: order_id: primary_key: true customer_id: {} order_total: {} measures: count: aggregate_type: count

 

It points at the table in BigQuery, lists the columns, and ends with the automatic count measure.

 

Then I ran the first converter:

 

bash ossie-omni import -i "PATH_TO_OMNI_MODEL" -o model.ossie.yaml --name abc_xyz  

A few seconds later, I had an Ossie file with 11 datasets, 3 relationships, and 11 metrics. Ossie comes with its own validation script, and the file passed on the first run. I got a bit suspicious, so I counted the fields per table on both sides and they matched.

The conversion worked. One detail worth knowing: anything only Omni is able to understand gets stored in a separate section marked OMNI under custom_extensions, so it can be put back later. That includes which table each measure belongs to. Keep that in mind – it’ll matter later.

 

Yaml semantic_model: - name: abc_xyz datasets: - name: orders source: abc_xyz-snowplow.dbt_prod_analytics_marts.orders primary_key: [order_id] fields: - name: order_id expression: dialects: - dialect: ANSI_SQL expression: order_id - name: customer_id expression: dialects: - dialect: ANSI_SQL expression: customer_id # ...one entry per column relationships: - name: orders_to_customers from: orders to: customers from_columns: [customer_id] to_columns: [customer_id] metrics: - name: orders__count expression: dialects: - dialect: ANSI_SQL expression: COUNT(*) custom_extensions: - vendor_name: OMNI data: '{"view": "orders", "aggregate_type": "count"}'

 

Now onto part two: Converting the Ossie file into a Power BI model

Power BI stores its semantic model in a JSON file called model.bim. Running the second converter looked like this:

 

bash ossie-microsoft export -i model.ossie.yaml -o model.bim  
 
The converter ran without complaining. Then I opened the output, and every single measure just said BLANK().

The root cause turned out to be how Omni’s count measure is stored in Ossie, as SQL: COUNT(*). Power BI doesn’t use SQL for measures. It uses its own formula language, DAX. The converter didn’t know how to turn COUNT(*) into DAX, so it just left the measures empty.

DAX is exactly what the converter was supposed to handle, as the idea of the Ossie model is to hold both SQL and DAX. With the help of Claude, we wrote a short Python script that adds a DAX version next to the SQL one for each measure, in this case COUNTROWS('orders'). With both in the file, the converter happily used the DAX one.

 

YAML metrics: - name: customers__count expression: dialects: - dialect: ANSI_SQL expression: COUNT(*) - dialect: DAX expression: COUNTROWS('customers') custom_extensions: - vendor_name: OMNI data: '{"_v": 1, "view": "customers", "name": "count"}' - name: dim_ga4__client_keys__count expression: dialects: - dialect: ANSI_SQL expression: COUNT(*) - dialect: DAX expression: COUNTROWS('dim_ga4__client_keys') custom_extensions: - vendor_name: OMNI data: '{"_v": 1, "view": "dim_ga4__client_keys", "name": "count"}' - name: dim_ga4__sessions__count expression: dialects: - dialect: ANSI_SQL expression: COUNT(*) - dialect: DAX expression: COUNTROWS('dim_ga4__sessions') custom_extensions: - vendor_name: OMNI data: '{"_v": 1, "view": "dim_ga4__sessions", "name": "count"}'
Screenshot of the Ossie model holding both ANSI_SQL and DAX

 

The next surprise was the data source. The converter always sets up the tables for Microsoft Fabric, in a mode called Direct Lake, and fills in placeholder IDs where the Fabric details would go. There’s no option to pick anything else. I get why Microsoft would build it that way, but our data lives in BigQuery. So I had to swap those settings for queries that import each table straight from BigQuery.

With both fixes in place, I had a model.bim that looked ready. There’s no “import model” button in Power BI Desktop, so this needed a small detour. I turned off the preview feature that saves models in the newer TMDL format, since the converter writes the older model.bim format. Then I saved an empty report as a Power BI Project (.pbip), swapped its model.bim for mine, deleted the cached data file, and opened the project. But first, here’s what the orders table looked like in model.bim after all the fixes: 

 

Json { "name": "orders", "columns": [ {"name": "order_id", "dataType": "int64", "sourceColumn": "order_id", "isKey": true}, {"name": "customer_id", "dataType": "int64", "sourceColumn": "customer_id"}, {"name": "order_total", "dataType": "decimal", "sourceColumn": "order_total"} ], "partitions": [ { "name": "orders", "mode": "import", "source": { "type": "m", "expression": [ "let", " Source = GoogleBigQuery.Database([BillingProject=\"abc-xyz-snowplow\"]),", " Project = Source{[Name=\" abc-xyz -snowplow\"]}[Data],", " Dataset = Project{[Name=\"dbt_prod_analytics_marts\",Kind=\"Schema\"]}[Data],", " Data = Dataset{[Name=\"orders\",Kind=\"Table\"]}[Data]", "in", " Data" ] } } ] }

 

After opening the file, I got multiple errors.  

I fixed them one by one; here is each one in order: 

  • “Culture and Collation”: I had no idea what this error meant at first. It turned out the converter had fixed the model’s language setting to en-US, and Power BI won’t open a model that disagrees with your laptop’s regional settings. I removed the setting and let Power BI use its own.

  • The version number: The file declared compatibility level 1702, which is basically a version number for the model format. My Power BI Desktop starts at 1600, so I set that.

  • Missing data types: Power BI needs a data type for every column, and the converter hadn’t set any. So, with the help of Claude, we added the relevant data types.

  • Two roads to the same table: Our orders_with_profit table connects to customers directly, and also through orders. Omni is fine with that, but Power BI wants to know exactly which path a filter should take, and two active paths make that impossible. I set the direct relationship to inactive, so filters now go through orders.

  • A couple of missing settings: Two model-level settings Power BI normally expects had to be added: defaultPowerBIDataSourceVersion and dataAccessOptions.

And finally, no more errors. I signed in to BigQuery, hit refresh, and waited. All 11 tables loaded with the right data types, the relationships were in place, and when I compared the measures with Omni, the numbers matched. 

One issue remained, though. All 11 measures were sitting on the customers table, including the ones counting orders. Remember the OMNI section that stores which table each measure belongs to? Microsoft’s converter doesn’t read other tools’ sections, so it never found out. They still calculate correctly, but nobody looking for the orders count would think to check under customers. I haven’t found a way to fix this through the converter, so it has to be done through a script. 

 

 The model in Power BI Desktop

 

Summary: Fixes needed, but the future is promising

To sum it up, the first part of the test went really smoothly, but the second part took some effort. 

None of the fixes were hard – especially with the help of AI – but getting there meant opening the file, reading an error, fixing one thing and trying again until it’s done. Two of the fixes required custom scripts. Most of the issues were Power BI specifics: DAX, where measures live, data types, regional settings, and relationship rules. And the converter assumes you’re on Microsoft Fabric, which plenty of Power BI users aren’t. 

To be fair, none of this is very surprising. Microsoft’s converter is brand new, and Ossie is still an early-stage project. Our model also made things easy, since every measure was a plain row count with no sums, ratios, or custom logic. 

So, if you’re thinking about switching to Power BI, it’s not a one-click migration yet. But I ended up with a working Power BI model, built from an Omni model without rebuilding a single table by hand, and the numbers match. That’s a really good start for a converter this new. If it keeps improving, moving a semantic model between tools could soon be a matter of a few commands.