Problem: A team with dozens of Power BI semantic models had no documentation. New starters spent weeks working out what the measures and relationships did. Keeping docs current was impossible because the models changed faster than the writing.

Approach: We built a pipeline that connects to a semantic model, extracts the full metadata (measures, relationships, DAX expressions), and generates readable documentation in several formats at once.

Technical details:

• Metadata extraction through the Power BI REST API and TMDL
• Descriptions written by GPT-4 with context-aware prompting
• Graphviz diagrams of measure dependencies and table relationships
• Output as Markdown, HTML, JSON, or wiki pages
• Runs from the command line or on a schedule

Outcome: Onboarding dropped from weeks to days. The pipeline regenerates when a model changes, so the documentation ships alongside the model update.

Stack: Python, Power BI, GPT-4/Azure OpenAI, Graphviz

Sandboxed plugin ran for projects/bi-documentation and read “BI Documentation Pipeline” through the mediated API.