fabric.guru
Sep 27, 2026
3/5 · Tier B
AI ReadinessPySpark
deep-dive
<p><strong>Fabric-RLM</strong> improves deterministic tasks by giving a language model an executable Python workspace, a limited turn loop, and an output contract. Instead of asking the model to reason directly from token patterns, the harness lets it construct a list, run code, inspect the result, and submit a typed answer.</p><p>Across nine small and midsize models, only one answered the weekday-letter question correctly without the harness, while all nine succeeded through Fabric-RLM. The result demonstrates tool-assisted verification rather than increased model knowledge; it does not establish reliability for open-ended generation or complex data work.</p>
Measured model comparison demonstrates how executable tools and guardrails improve deterministic data tasks, though the example is intentionally simple.
fabric.guru
Sep 25, 2026
4/5 · Tier A
AI ReadinessAPI AutomationData Quality
deep-dive
<p><strong>Fabric-RLM</strong> runs recursive language-model tasks beside Fabric data instead of placing entire files into a prompt. The model writes Python, executes it in the notebook environment, inspects results, repairs failures, and returns only outputs matching a typed contract and optional validators.</p><p>Inputs can include Lakehouse files, Delta tables, and semantic models; skills provide reusable procedures for exploration, Excel, PDFs, Delta, and semantic-model work. Each run records generated code, outputs, errors, tokens, timings, and validation repairs, enabling inspection and replay.</p><p>A 140 MB IMF CSV containing 1.57 million rows was processed with DuckDB and openpyxl in 78 seconds, producing a validated workbook. The approach still depends on model-generated code, notebook execution limits, validator quality, and model-endpoint or external-model costs.</p>
Concrete Fabric-native RLM implementation includes measured performance, validation, cost comparison, and operational traceability.
SQLBI
Sep 22, 2026
2/5 · Tier B
AI ReadinessSemantic Model
opinion
<p>AI-generated DAX has become more useful in 2026, particularly for drafting measures, transforming models, and handling longer development tasks. The productivity gain is strongest when the request clearly states business intent and the developer can inspect the generated semantics.</p><p>AI still does not reliably resolve filter context, data-model design, edge cases, or performance without validation. The article is primarily an opinion and training-position update rather than a Fabric feature or measured production study.</p>
The opinion reports improving AI-generated DAX but emphasizes human validation, modeling judgment, and performance review.
