fabric-semantic-model

name	fabric-semantic-model
description	Infer and generate Power BI semantic models (TMDL format) from Microsoft Fabric lakehouse tables. Covers star-schema detection, table classification (fact/dimension), relationship inference, DAX measure generation, data type mapping, and deployment via REST API. Use when: create semantic model, build data model, star schema, Power BI model, TMDL, fact table, dimension table.
compatibility	Requires the Fabric MCP Server VS Code extension for table discovery. Generates TMDL files deployable to Microsoft Fabric via REST API.

Fabric Semantic Model Creator — Reference

Generates Power BI semantic models in TMDL format by analyzing lakehouse table schemas and sample data. The agent reasons about table names, column names, data types, cardinality, and sample values to infer the best star-schema model.

Table Classification Rules

Classify each lakehouse table as fact or dimension based on these signals:

Signal	Fact Table	Dimension Table
Name patterns	Contains fact_, fct_, sales, orders, transactions, events, log, movements	Contains dim_, lookup, calendar, date, product, customer, employee, geography, category, status
Column patterns	Has multiple _id/_key foreign keys + numeric measure columns (amount, quantity, price, total, count)	Has a single primary key + descriptive text columns (name, description, label, title, code)
Row count	Typically high (transactions)	Typically low (reference data)
Numeric density	Many numeric columns relative to total	Few numeric columns, mostly strings

If ambiguous, default to dimension. Better to have an extra dimension than miss a fact table's relationships.

Relationship Inference Rules

Detect relationships by matching column names across tables:

1. Exact key match: orders.customer_id → customers.customer_id (or customers.id)
2. Name-based match: column named {table_name}_id or {table_name}_key in a fact table points to the dimension with that name
3. Single-column PK in dimension: if a dimension has exactly one _id/_key column and a fact table has a column with the same name, infer the relationship
4. Date keys: columns named date, date_id, fecha, order_date, etc. connect to a date/calendar dimension

All relationships are many-to-one (fact → dimension) unless the agent detects a bridge/junction table (many-to-many).

Cross-filter direction: Single by default. Only use Both for bridge tables.

DAX Measure Generation Rules

Generate measures based on column classification:

Column Pattern	Measure	DAX Expression
amount, total, revenue, sales, importe	Sum	SUM('{table}'[{column}])
quantity, qty, count, cantidad, units	Sum	SUM('{table}'[{column}])
price, cost, precio, coste	Average	AVERAGE('{table}'[{column}])
discount, descuento	Sum + Average	Both measures
Any numeric in fact table	Row Count	COUNTROWS('{table}') (one per fact table)
id/key in dimension	Distinct Count	DISTINCTCOUNT('{table}'[{column}])

Naming convention: Total {Column}, Avg {Column}, # {Table} (row count), # Unique {Column} (distinct count). Use business-friendly names, not technical ones.

Data Type Mapping

Map lakehouse (Spark) types to TMDL types:

Spark Type	TMDL dataType	Notes
string	string
int, integer	int64	TMDL uses int64 for all integers
long, bigint	int64
float	double	TMDL has no float, use double
double	double
decimal	decimal
boolean	boolean
date	dateTime	TMDL uses dateTime for dates
timestamp	dateTime

TMDL File Structure

The semantic model is a folder of .tmdl files plus metadata:

{model_name}/
├── definition/
│   ├── database.tmdl          # Database-level settings
│   ├── model.tmdl             # Model-level settings (culture, access)
│   ├── relationships.tmdl     # All relationships
│   └── tables/
│       ├── {table1}.tmdl      # Table definition + columns + measures + partition
│       ├── {table2}.tmdl
│       └── ...
└── definition.pbism            # Semantic model settings

database.tmdl template

database '{MODEL_NAME}'
	compatibilityLevel: 1604

model.tmdl template

The model must include a dataSource block pointing to the lakehouse's SQL analytics endpoint. The agent must fetch this endpoint from the Fabric REST API: GET /v1/workspaces/{id}/lakehouses/{id} → look for properties.sqlEndpointProperties.connectionString.

model Model
	culture: es-ES
	defaultPowerBIDataSourceVersion: powerBI_V3
	discourageImplicitMeasures
	sourceQueryCulture: es-ES

	dataSource '{LAKEHOUSE_NAME}'
		type: structured
		connectionDetails
			protocol: tds
			address
				server: {SQL_ENDPOINT}
				database: {LAKEHOUSE_NAME}
		credential
			AuthenticationKind: OAuth2

Table .tmdl template (for each table)

Partitions use Power Query m expressions with mode: directQuery connecting to the SQL analytics endpoint. Do NOT use entity partitions or expressionSource: DatabaseQuery — those fail when deploying via REST API.

table '{TABLE_NAME}'

	measure '# {TABLE_NAME}' = COUNTROWS('{TABLE_NAME}')
		formatString: #,##0

	column {COLUMN_NAME}
		dataType: {TMDL_TYPE}
		sourceColumn: {SOURCE_COLUMN}
		summarizeBy: {none|sum|average|count}

	partition '{TABLE_NAME}' = m
		mode: directQuery
		source =
			let
				Source = Sql.Database("{SQL_ENDPOINT}", "{LAKEHOUSE_NAME}"),
				dbo_{TABLE_NAME} = Source{{[Schema="dbo",Item="{TABLE_NAME}"]}}[Data]
			in dbo_{TABLE_NAME}

relationships.tmdl template

relationship {GUID}
	fromColumn: '{FACT_TABLE}'.{FK_COLUMN}
	toColumn: '{DIM_TABLE}'.{PK_COLUMN}

definition.pbism

{
  "version": "4.0",
  "settings": {}
}

Sample Data Analysis

When classifying tables and inferring relationships, analyze sample data (first 100 rows) to:

1. Confirm cardinality: A column with 10 distinct values in 10000 rows is likely a FK to a dimension
2. Detect date formats: Sample values reveal if a string column is actually a date
3. Detect key patterns: IDs that are sequential integers vs UUIDs vs composite
4. Validate relationships: Check that FK values actually exist in the referenced dimension's PK

Gotchas

• Star schema is mandatory. DAX is optimized for star schema. Flat/denormalized tables work but give poor performance and confuse AI tools. If the lakehouse has a single flat table, the agent should recommend splitting it into fact + dimensions (but not do it automatically).
• Partitions must use Power Query m with mode: directQuery. Do NOT use entity partitions, expressionSource: DatabaseQuery, or mode: directLake — those all fail when deploying via the REST API. The correct pattern is a Power Query Sql.Database() expression pointing to the lakehouse SQL analytics endpoint. See the table template above.
• The model needs a dataSource block. Without it, the partition's Sql.Database() call has no endpoint. Fetch the SQL analytics endpoint from GET /v1/workspaces/{id}/lakehouses/{id} → properties.sqlEndpointProperties.connectionString. Use credential: AuthenticationKind: OAuth2.
• Use the script for SQL endpoint and schema discovery, not MCP. The MCP server doesn't expose the SQL analytics endpoint and onelake_get_table fails for some lakehouses. Use the script's sql-endpoint command for the endpoint and list-tables command for table schemas. These use the REST API + Delta log reading and are reliable.
• summarizeBy: none for all non-measure columns. Only measure columns should have sum/average. Setting summarizeBy on regular columns creates implicit measures which pollute the model.
• Use discourageImplicitMeasures in model.tmdl. This forces explicit DAX measures and prevents automatic SUM/COUNT on numeric columns.
• Spanish locale: Use culture: es-ES and sourceQueryCulture: es-ES in model.tmdl. Format strings use . for thousands and , for decimals in Spanish.
• Relationship GUIDs: Generate a unique UUID for each relationship. Use Python's uuid.uuid4().
• Column names with spaces or special chars: Wrap in single quotes in TMDL: column 'Total Amount'.
• Key columns should be and : Primary keys in dimensions should be hidden from report view and marked as the table key.

Deployment, Management & Cleanup

Use ./scripts/fabric_semantic_model.py to manage semantic models in Fabric via the REST API. Uses DefaultAzureCredential (Azure CLI session).

Install dependencies (one-time):

pip install azure-identity requests

Deploy a semantic model

Reads a local TMDL folder, base64-encodes each file, and creates a SemanticModel item via POST /v1/workspaces/{id}/semanticModels. Includes the .platform and definition.pbism files.

python scripts/fabric_semantic_model.py deploy <workspace_id> <model_name> <tmdl_folder_path>

List semantic models

python scripts/fabric_semantic_model.py list <workspace_id> [<model_name>]

Delete a semantic model

python scripts/fabric_semantic_model.py delete <workspace_id> <model_id>

List tables with schemas

Returns all tables in a lakehouse with their column names and types as JSON. Uses the Fabric REST API to list tables, then reads each table's Delta log to extract the schema. Use it instead of MCP onelake_get_table if it fails.

python scripts/fabric_semantic_model.py list-tables <workspace_id> <lakehouse_id>

Get lakehouse SQL analytics endpoint

Returns the SQL analytics endpoint required for the dataSource block in model.tmdl. Run this in Phase 1 and use the output when generating TMDL files.

python scripts/fabric_semantic_model.py sql-endpoint <workspace_id> <lakehouse_id>

Deployment gotchas

• The definition field is required for semantic model creation — unlike notebooks, you cannot create an empty model and update it later.
• displayName must be unique in the workspace.
• TMDL format is the default and recommended. Use definition/ folder structure.
• The API may return 202 for provisioning — the script polls automatically.

Global Rules

• TMDL format only. Generate TMDL files, not model.bim (TMSL). TMDL is the modern format, text-based, and Git-friendly.
• Star schema. Always model as star schema with clear fact and dimension tables.
• Explicit measures only. Use discourageImplicitMeasures and create all measures explicitly in DAX.
• Business-friendly names. Tables, columns, and measures should use clear names like "Total Revenue" not "TR_AMT".
• Non-destructive. Never overwrite an existing semantic model. Deploy with a unique name or ask the user first.