ARF Layer
Context Stability
Keeps filter context predictable so the same question produces the same answer every time.
Layman explanation
Context Stability is about how filters, relationships, and time logic shape results. When relationships are ambiguous or filters interact unexpectedly, AI answers change across runs or users. This layer makes context explicit and deterministic.
A sequence showing question → context → model → answer, emphasizing stable context.
What breaks when this layer is weak
- • AI answers differ between users with similar questions.
- • The same prompt returns different values minutes apart.
- • Trend explanations flip because the filter context is unstable.
Symptoms you can observe
- • Many-to-many relationships without clear bridge logic.
- • Inactive relationships that are inconsistently activated.
- • Heavy use of bidirectional filtering to “make it work.”
- • Complex slicer interactions that are hard to reason about.
Root causes
- • Relationships are designed for visual reports, not deterministic queries.
- • Role‑playing dimensions are not clearly separated.
- • Time intelligence relies on implicit date fields or hidden logic.
- • Security filters are not tested with AI queries.
What good looks like
- • Stable filter paths from dimensions to facts.
- • Clear use of active vs inactive relationships.
- • Deterministic time intelligence using a dedicated date table.
- • Documented slicer behavior and evaluation order.
Remediation checklist
- • Map and document filter paths for critical metrics.
- • Reduce bidirectional relationships; use bridges instead.
- • Create a context test harness with representative queries.
- • Validate RLS behavior for AI‑driven access.
Metrics to track
- • # of ambiguous relationships
- • % of measures with deterministic filter paths
- • # of bidirectional relationships
- • Context variance across repeated queries
Foundational articles
Why AI Answers Change When Your Data Didn’t
Inconsistent context, not data changes, often causes fluctuating AI answers.
Filter Context in Plain English
Filter context determines what data a calculation sees; it must be predictable.
Context Volatility: Hidden Interactions Between Slicers and Measures
Volatile context is caused by slicer interactions, hidden filters, and ambiguous paths.
Ambiguous Relationships: The Silent Context Killer
Ambiguous relationships create multiple filter paths, leading to unpredictable answers.
Many-to-Many Relationships and AI: What Can Go Wrong
Many‑to‑many relationships can produce unexpected filter behavior for AI queries.
Inactive Relationships and USERELATIONSHIP: When Intent Gets Lost
Inactive relationships require explicit activation, which AI often misses.
Bidirectional Filtering: Convenience vs Predictability
Bidirectional filters can make models easier to use, but less predictable for AI.
Role-Playing Dimensions: Dates, Regions, and Other Multipliers
Role‑playing dimensions require clear naming and explicit usage to avoid confusion.
Row-Level Security and AI: What You Must Validate
RLS affects AI answers and must be validated with realistic AI queries.
Time Intelligence: Why ‘Last Month’ Is Harder Than It Sounds
Time intelligence depends on clean date tables and clear definitions of time.
Deterministic Slices: Designing Questions AI Can Ask Reliably
Deterministic slices constrain questions so answers stay consistent and explainable.
A Context Test Harness for Power BI Models
A test harness validates that key questions return stable results.