Schema Discovery

Table of Contents

Must / Prefer / Avoid

Must

• Use fully-qualified 'Table'[Column] for column references
• Use simple [Measure] for measure references

Prefer

• INFO.VIEW functions for initial metadata discovery (read access, lightweight)
• Progressive schema discovery over full schema dumps
• SELECTCOLUMNS + FILTER to narrow INFO results

Avoid

• Fetching full schema upfront — discover incrementally based on need
• Re-fetching metadata already discovered in this conversation
• Using GetSemanticModelSchema, DiscoverArtifacts, or GenerateQuery MCP tools (these are not available)

Progressive Schema Discovery

Discover metadata incrementally based on what the user actually needs.

Strategy (Decision Tree)

User asks a question
  → Do I know which tables are relevant?
    → NO: Run INFO.VIEW.TABLES() to get table inventory
    → YES: Do I know the columns/measures for those tables?
      → NO: Run filtered INFO.VIEW.COLUMNS() and INFO.VIEW.MEASURES() for those tables
      → YES: Do I need relationships?
        → YES: Run INFO.VIEW.RELATIONSHIPS() filtered to those tables
        → NO: Do I need advanced metadata?
          → Have elevated INFO functions already failed in this session?
            → YES: Skip — assume no permission for all elevated queries
            → NO: Try the relevant elevated query:
              → Calculation groups: INFO.CALCULATIONGROUPS() + INFO.CALCULATIONITEMS()
              → Calendars: INFO.CALENDARS()
              → User-defined functions: INFO.USERDEFINEDFUNCTIONS()
              → Variations: INFO.VARIATIONS()
              → If query fails with permission error: mark elevated access as unavailable
  → Use discovered schema to write DAX queries (see dax-authoring skill)

Rules

• Start with scope estimation — Run the scope probe query first to understand model size
• Discover on demand — Only fetch tables/columns/measures relevant to the current user request
• Use INFO.VIEW functions first (read access) — tables, columns, measures, relationships
• Use INFO functions (may need elevated access) for: calculation groups, calculation items, variations, user-defined functions, calendars
• Handle permission failures gracefully — If an elevated INFO function fails but INFO.VIEW.* functions succeeded, assume the user lacks elevated permissions and skip all elevated discoveries entirely
• Always narrow results — Use SELECTCOLUMNS + FILTER to fetch only needed columns and rows
• Cache discovered schema mentally — Don't re-fetch what you've already discovered in this conversation

Recommended Discovery Order

1. Scope estimation — Count tables, columns, measures, relationships
2. Tables — Get table names, identify relevant ones
3. Columns — Get columns for relevant tables only
4. Measures — Get measures (prefer using these over raw aggregations)
5. Relationships — Get relationships between relevant tables
6. Advanced metadata (if needed) — Calculation groups, calendars, UDFs, variations

Scope Estimation Query

EVALUATE
ROW(
    "TableCount", COUNTROWS(INFO.VIEW.TABLES()),
    "ColumnCount", COUNTROWS(INFO.VIEW.COLUMNS()),
    "MeasureCount", COUNTROWS(INFO.VIEW.MEASURES()),
    "RelationshipCount", COUNTROWS(INFO.VIEW.RELATIONSHIPS())
)

Metadata Object → INFO Function Map

For the full query catalog and output column details, see discovery-queries.md.

Running Discovery Queries

Use npx fabric-app-data query <alias> --query '<DAX>' to execute INFO queries against a semantic model. For full CLI options (profiles, file input, result limits), see the fabric-cli skill.

npx fabric-app-data query <alias> --query "EVALUATE INFO.VIEW.TABLES()"

Troubleshooting