DAX Authoring

Table of Contents

Must / Prefer / Avoid

Must

• Always test generated DAX before using it in app code. Execute each planned .dax file once through npx fabric-app-semantic-model-tools run-dax-query and compare its report with the schema-query plan
• Run DAX through run-dax-query, not npx rayfin connector invoke; run-dax-query runs Rayfin CLI calls one at a time.
• Use fully-qualified 'Table'[Column] for column references
• Use simple [Measure] for measure references
• Use DEFINE for VAR and local MEASURE declarations (single DEFINE block, no commas)
• Use existing model measures for requested business metrics. Raw aggregation is allowed only when the schema workflow established that no suitable shipped measure exists. Never replace a shipped measure with raw SUM, COUNT, COUNTROWS, AVERAGE, DISTINCTCOUNT, or a local approximation
• Generate DAX from the completed schema-query plan; do not rediscover or substitute model objects
• Do not implement the UI until the final DAX has executed successfully and its output passes the planned shape and semantic validation gates, with every warning resolved or explained by the request

Prefer

• SUMMARIZECOLUMNS as the primary grouping function for queries
• TREATAS for filter arguments in SUMMARIZECOLUMNS
• Variables (VAR) to improve readability and avoid repeated calculations
• Shipped measures from measure-heavy KPI tables

Avoid

• SQL keywords (SELECT, WHERE, HAVING, etc.) within DAX expressions
• EVALUATE with scalar functions directly (wrap in ROW or table function)

Generating DAX Queries

DAX Syntax Rules

• Measures are named objects in the semantic model. They specify how to aggregate data. Measures should be used first to answer user requests before a new DAX formula is used to aggregate data.
• EVALUATE Statement: Not a function but a statement. It must always precede a table expression. Avoid pairing EVALUATE with scalar functions like CONCATENATEX, SUMX, DISTINCTCOUNT, etc.
• Use the fully-qualified name 'Table'[Column] for column references, and the simple name [Measure] for measure references.
• CALCULATE takes a scalar expression as its first argument. CALCULATETABLE takes a table expression as its first argument.
• SUMMARIZECOLUMNS requires a specific order: groupby columns, then filters, then aggregations/measures.
• Do not use SUMMARIZECOLUMNS when there is no aggregation and the groupby columns belong to more than one table; use VALUES, SUMMARIZE, or SELECTCOLUMNS instead.
• When using SELECTCOLUMNS or CALCULATETABLE, include any columns needed downstream (ORDER BY, FILTER).
• Filters propagate across relationships based on unidirectional or bidirectional settings.
• INTERSECT, UNION, EXCEPT require identical column counts in both inputs.
• For current date/time, use TODAY() or NOW().

Inline Examples

Simple filtered aggregation:

// Total sales for red products
EVALUATE
  ROW("Total Sales Amount", CALCULATE([Total Amount], 'Product'[Color] == "Red"))

Multi-filter grouping with SUMMARIZECOLUMNS:

DEFINE
  VAR _Filter1 = TREATAS({"Consumer Electronics"}, 'Product'[Category])
  VAR _Filter2 = FILTER(ALL('Calendar'[Year]), 'Calendar'[Year] >= 2022 && 'Calendar'[Year] <= 2023)

EVALUATE
  SUMMARIZECOLUMNS(
    'Calendar'[Year],
    'Calendar'[Month],
    _Filter1,
    _Filter2,
    "Total Quantity", SUM('Sales'[Order Quantity]),
    "Discount", [Total Discount]
  )

TopN with filtering:

DEFINE
  VAR _Filter = TREATAS({"Red", "Black"}, 'Product'[Color])
  VAR _Core = SUMMARIZECOLUMNS('Product'[Name], _Filter, "Total Sales", [Total Amount])

EVALUATE
  TOPN(10, _Core, [Total Sales], DESC)

For full syntax reference, worked examples, and anti-patterns, see dax-query-patterns.md. For function details, see dax-core-reference.md.

Time Intelligence

Time intelligence functions enable period-based analysis (YTD, YoY, rolling windows, etc.). They require a properly configured Date table.

Consult dax-time-intelligence.md whenever the user's request involves:

• Period-to-date calculations (YTD, QTD, MTD)
• Period comparisons (Year-over-Year, Month-over-Month)
• Rolling windows (last 12 months, last 30 days)
• Opening/closing balances
• Custom date ranges

Testing & Iteration

1. Use the completed schema-query plan: exact shipped measures, grouping columns, filters, relationship path, expected columns, and row shape
2. Author one EVALUATE for each distinct query the plan requires, and execute each final query once with run-dax-query (see Query Execution)
3. Treat an execution error as a query failure. A successful result is not by itself proof of correctness, and an empty one is not an error. Compare the reported columns, row count, sample, and warnings with the plan's exact output columns, row shape, filters, ordering, grain, and value expectations
4. Reject all-blank output, unexplained all-zero grouped output, distinct requested outputs that repeat the same measure/value, and all-rank-1 output when comparison or ordering is expected. These are semantic mismatches even when the command succeeds; do not build the UI from them
5. If an error or semantic mismatch remains, inspect the output-binding map, helper or parameter filter context, relationships, and filters; consult dax-core-reference.md, make one evidence-based correction to only the failing query, and re-test it
6. Stop each distinct planned query after it returns a structurally and semantically consistent result. Across the task, run one optional grain diagnostic for a shared ambiguity; for each distinct query, run one validation and at most one evidence-based corrective retry. Do not use this allowance for exploratory model probing or repeated discovery

Preserve the selected candidate by editing the validated query rather than regenerating it.

Result reuse and stopping

• Before running a query, check whether a previous result already contains the answer. Never rerun an identical or equivalent DAX query.
• A zero scalar is valid. An all-zero grouped result is not validated for a task that expects meaningful comparison unless task or catalog evidence shows that all zeros are legitimate.
• For a blank or empty result, inspect the shipped measure, relationship path, grouping dimension, helper-table filter context, and filters. Make one correction; accept a corrected blank only when the requested semantics allow no data.
• If every grouped row returns the same measure value, first investigate a disconnected grouping table or ineffective filter propagation. Do not rewrite or replace the measure to manufacture variation.
• If every row ranks 1, verify that the rank expression evaluates over the intended comparison set and that filters did not collapse that set before accepting the result.
• On timeout, narrow filters, reduce grouping cardinality, apply TOPN, or simplify only the failing planned query. Never retry the same query unchanged.
• Preserve successful batched outputs. Correct only failed or semantically inconsistent files.

Query Execution

Use:

npx fabric-app-semantic-model-tools run-dax-query \
  --dax src/queries/revenue.dax \
  --alias <alias>

Each .dax file must contain one EVALUATE; otherwise nothing executes. For two to four planned queries, repeat --dax in one call:

npx fabric-app-semantic-model-tools run-dax-query --dax src/queries/a.dax --dax src/queries/b.dax --alias <alias>

The queries run one at a time, through the same connector operation as the running app. For each query the command prints its status, output column names, row count, first rows (--sample <n>, default 5), and warnings, not the full result. Warnings flag no rows, an all-blank column, or a measure output (not a Table[Column] grouping column) that is all zero, constant, all rank 1, or identical to another output; identical outputs are reported as one group. Warnings are nonfatal prompts for review, not proof of correctness. A shortened preview is not complete shape evidence.

A failed query keeps its own entry beside the successful ones, and the call exits 2: correct only the failed queries. Missing, malformed, unreadable, or overflowed results never pass, and input or read failures also exit nonzero; an error message alone does not show that a query ran. An overflow result has no table; see semantic-model-data.

Row count: a query returns every row unless you bound it. Bound it with the aggregation the plan calls for, and use TOPN(...) only when the request asks for a ranked subset or to bound a query that timed out. --limit <n> is a guard, not a cap: validation fails when more than n rows come back.

Troubleshooting