powerbi

Overview

Power BI Dashboard Generator is an open, educational reference implementation of an agent-driven BI authoring pipeline. It demonstrates how a Large Language Model (LLM) agent - equipped with a set of well-scoped, file-based skills - can take a natural-language request such as "Build me a sales performance dashboard from sales.xlsx" and emit a complete, openable Power BI Desktop Project (PBIP) without manual modeling, DAX, or visual configuration.

The project is intentionally structured for learning:

• Every skill is a self-contained folder under skills/ with a Markdown SKILL.md contract, reference docs, and runnable Python scripts.
• The pipeline is split into single-responsibility stages so each step can be studied, tested, and replaced in isolation.
• All artifacts the agent produces (data-model.json, TMDL files, PBIR visual.json) are plain text and human-readable.

Use this repo to learn: prompt-engineering for code-generation agents, the PBIP/TMDL/PBIR file formats, source-agnostic semantic modeling, SQL → DAX translation, and visual-bucket selection heuristics.

🚀 Quick Start

Install

git clone https://github.com/bcastelino/powerbi-dashboard-generator.git
cd powerbi-dashboard-generator
pip install -r requirements.txt

Register the skills

Drop or symlink the skills/ folder into your agent runtime. The agent will discover each SKILL.md automatically - no further configuration required. See the Installation Guide below for runtime-specific instructions.

Ask for a dashboard

User > Build me a sales performance dashboard from sales.xlsx

Agent > [invokes nlq-dashboard-orchestrator]
        [invokes data-source-connector → introspects sales.xlsx]
        [Gate A] Confirms data model with user
        [NLQ loop] Clarifies visuals, filters, layout
        [Gate B] "Ready to scaffold? (yes/no)"
        [pipeline] Generates PBIP → applies theme → assembles dashboard
        ✓ Dashboard ready: generated-dashboards/SalesPerformanceDash/

Open the resulting .pbip file in Power BI Desktop - the semantic model loads, the report renders, and visuals connect to live data.

🏗 Architecture

The system is a staged pipeline orchestrated by a top-level NLQ skill, with two human-in-the-loop confirmation gates and file-based handoffs between every skill.

---
config:
  layout: elk
  look: classic
---
flowchart TD
    User([User NLQ])
    Orchestrator[nlq-dashboard-orchestrator]
    Connector[data-source-connector]
    GateA{{Gate A<br/>data model OK?}}
    NLQ[NLQ Q&A loop<br/>visuals · filters · time]
    GateB{{Gate B<br/>ready to scaffold?}}

    subgraph Pipeline[query-to-pbip pipeline]
        direction LR
        Semantic[semantic-mapper]
        Selector[visual-selector]
        Generator[visual-generator]
        Packager[project-packager]
        Semantic --> Selector --> Generator --> Packager
    end

    Theme[theme-branding]
    Dash[bi-dash-creator]
    Output[(.pbip<br/>TMDL · PBIR · theme)]

    User --> Orchestrator --> Connector --> GateA
    GateA -->|confirmed| NLQ --> GateB
    GateB -->|yes| Pipeline --> Theme --> Dash --> Output

    classDef gate fill:#fff4ce,stroke:#d4a72c,color:#5c4a06;
    classDef skill fill:#dbeafe,stroke:#2b6cb0,color:#1e3a5f;
    classDef output fill:#dcfce7,stroke:#16a34a,color:#14532d;
    class GateA,GateB gate;
    class Orchestrator,Connector,NLQ,Semantic,Selector,Generator,Packager,Theme,Dash skill;
    class Output output;

The two confirmation gates

The orchestrator never generates files silently. It pauses at two explicit checkpoints: