Pbi Context

pbi-context — AI Context Engine for Power BI Models

[](https://pypi.org/project/pbi-context/) [](https://github.com/Osc2405/pbi-context/actions/workflows/tests.yml) [](https://www.python.org/) [](https://github.com/Osc2405/pbi-context/blob/main/LICENSE) [](https://github.com/Osc2405/pbi-context/actions/workflows/tests.yml)

pbi-context is the zero-dependency context compiler that lets any AI agent read, query, and audit a Power BI model — via CLI, indexed JSON, or a read-only MCP server.

Who it's for: data engineers documenting dashboards, consultants auditing models they didn't build, and anyone connecting an AI agent (Claude, GPT, Copilot) to a Power BI model's structure.

Quick Start

pip install pbi-context

# From a .pbit file...
pbi-context --input "data/pbit/my-model.pbit"
# ...or a PBIP project (folder, .pbip marker, or .SemanticModel/ — auto-detected)
pbi-context --input "data/pbip/my-model/"

cat "output/my-model.pbit/model_documentation.md"

Result: 7 files in output// in seconds — human-readable Markdown, JSON/JSONL context for AI agents, and an indexed, queryable version for large models.

Full walkthrough — folder structure, expected output, and using the context in Python

The command generates a folder in output/ with all documentation files:

output/my-model.pbit/          (or output/my-model/ for PBIP)
├── metadata.json              # Structured model metadata
├── model_documentation.md     # Human-readable documentation
├── agent_context.json         # LLM-optimized context (top-20 measures)
├── model_context.jsonl        # JSONL format for embeddings/RAG
├── index.json                 # Lightweight index + pointers
├── relationships.json         # All relationships
└── tables/
    ├── Sales.json             # Full detail per table
    └── ...

cat "output/my-model.pbit/model_documentation.md" | head -n 15

# my-model - Power BI Data Model

**Generated:** 2025-12-22 14:23:29

## Model Summary

- **Business Tables:** 9
- **Total Columns:** 23
- **Total Measures:** 44
- **Relationships:** 9

Use the JSON context directly in Python (or point an AI agent at it via --query or --mcp-serve — see [Use Cases](#use-cases) below):

import json

with open("output/my-model.pbit/agent_context.json", "r", encoding="utf-8") as f:
    context = json.load(f)

print(f"Model: {context['model_name']}")
print(f"Key measures: {len(context['key_measures'])}")
print(f"First measure: {context['key_measures'][0]['name']}")

Model: my-model
Key measures: 20
First measure: Revenue Budget

Why pbi-context?

| Your Need | pbi-context Solution | |-----------|---------------------| | Document 10+ dashboards fast | Batch processing with --batch | | Support the new PBIP format | Full TMDL parser, auto-detected from .pbip or folder | | Train AI agents on your models | Indexed JSON/JSONL context, a query CLI, and an MCP server | | Let an AI agent query the model live | Read-only MCP server (--mcp-serve) — validated against a test harness, not yet a live MCP client, see MCP server | | Use it from your AI coding assistant | Chat-invocable Skill for Claude Code + prompt file for GitHub Copilot | | Actually readable DAX | Hierarchical indentation (4x better than raw) | | Compare model versions | Content-aware --diff, with impact analysis (--diff-impact) | | See the model at a glance | Embedded Mermaid ER diagram in model_documentation.md — renders natively on GitHub/VS Code | | Visualize the model in Gephi/yEd | --export-graph — JSON node/edge lists or GraphML | | Zero-cost, zero-install | Python-only, no .NET dependencies |

Perfect for: Data engineers onboarding teams, consultants auditing models, organizations building AI copilots for BI.

Project Status

v1.1.0 is published on PyPI under the name pbi-context — this project was previously published as pbi-docs (v1.0.0–1.0.1); the tool didn't change, only the name, to avoid a discoverability collision with an unrelated, similarly-named project. See CHANGELOG.md for the full rename note.

The read/context layer — PBIP/TMDL support, indexed output, query resolver, MCP server, --diff-impact, --export-graph — is done, implemented and tested (see the Tests badge above for the current count). Every claim above is backed by a dated, reproducible report, not just asserted: see [Validation](#validation) below.

Next up: validating with real human users that scoped context doesn't cost time or accuracy versus raw file dumps — the protocol is ready (docs/humanvalidation_protocol.md), currently blocked on recruiting participants, not on code.

Writing/editing TMDL models and PBIR/report-layer parsing remain deliberately out of scope (see [Roadmap](#roadmap) for why).

Last updated: 2026-08-03.

Requirements

• Python 3.10+ (3.12 recommended)
• Works on Windows, macOS, and Linux

Optional: virtual environment (venv). No external libraries required.

Installation

Just want to run pbi-context? pip install pbi-context (see Quick Start above) is all you need. The steps below are for working on pbi-context itself (editable install from a local clone).

# 1) Clone or download the repository
# 2) (Optional) Create and activate a virtual environment
python -m venv venv
source venv/bin/activate

# 3) Editable installation (development)
pip install -e .

# Verify Python version
python --version

Windows/PowerShell notes

python -m venv venv
./venv/Scripts/Activate.ps1

If PowerShell blocks activation, run as Administrator:

Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

---

Quick Usage

Basic Commands

Process a .pbit file:

pbi-context --input "data/pbit/my-model.pbit"

Process a PBIP project (new in v1.0):

# From the .pbip marker file
pbi-context --input "data/pbip/my-model.pbip"

# From the .SemanticModel folder directly
pbi-context --input "data/pbip/my-model.SemanticModel"

# From the project root folder (auto-detected)
pbi-context --input "data/pbip/my-model/"

Specify custom output directory:

pbi-context -i "data/pbit/my-model.pbit" -o "my-results"

Process multiple files (batch mode — mixed formats supported):

pbi-context --batch "data/pbit/*.pbit"

Compare two versions of a model (mixed .pbit/.pbip supported):

pbi-context --diff "data/pbit/model_v1.pbit" "data/pbip/model_v2/"

Verbose mode (more debugging information):

pbi-context --input "data/pbit/my-model.pbit" --verbose

Human-readable indexed output (indented JSON, for debugging — compact by default):

pbi-context --input "data/pbit/my-model.pbit" --pretty

Generate documentation in Spanish:

pbi-context --input "data/pbit/my-model.pbit" --lang es

Generate documentation in English (default):

pbi-context --input "data/pbit/my-model.pbit" --lang en
# Or simply omit --lang (English is the default)
pbi-context --input "data/pbit/my-model.pbit"

See docs/troubleshooting.md for the full expected-output walkthrough, how to verify a fresh install, and common errors.

Important Notes

• Supported formats: .pbit files (ZIP + JSON TMSL) and .pbip projects (TMDL folder structure). .pbix files must be exported to .pbit from Power BI Desktop (File > Export > Power BI Template).
• PBIP entry points: The --input flag accepts a .pbip marker file, a .SemanticModel/ folder, or a project root folder. Format is auto-detected.
• Microsoft Fabric semantic models: Fabric uses the same TMDL format as PBIP, so compatibility is expected but not empirically validated (no real Fabric export has been tested against this parser yet) — see docs/fabric_compatibility.md.
• Language selection: Use --lang en for English (default) or --lang es for Spanish. The language affects the generated model_documentation.md and agent_context.json files.
• Paths with spaces: Use quotes around paths that contain spaces.
• Recommended paths: Place your files in data/ or data/pbit/ to keep the project organized.

---

Project Structure

pbi-context/
├── pbi_extractor/           # Main package
│   ├── __init__.py
│   ├── cli.py              # CLI with argparse (auto-detection, --index-format)
│   ├── extractor.py        # .pbit (ZIP+JSON TMSL) extractor
│   ├── pbip_extractor.py   # .pbip / TMDL extractor (new in v1.0)
│   ├── processor.py        # Metadata processing (format-agnostic)
│   ├── indexed_output.py   # index.json + tables/*.json writer (new in v1.0)
│   ├── formatters.py       # Advanced hierarchical DAX formatting
│   ├── categorizer.py      # Table/measure categorization
│   ├── documentation.py    # Markdown generation
│   ├── diff.py             # Model comparison
│   ├── jsonl_generator.py  # JSONL generator for LLMs
│   └── i18n.py             # Translations (en/es)
├── tests/
│   ├── fixtures/
│   │   └── minimal_pbip/   # TMDL test fixtures (new in v1.0)
│   ├── test_pbip_extractor.py
│   ├── test_cli_detection.py
│   ├── test_indexed_output.py
│   ├── test_categorizer.py
│   ├── test_i18n.py
│   └── test_processor_and_context.py
├── githooks/                # Reference pre-commit hook (docs/pre_commit_hook.md)
├── data/                   # Input model files
├── output/                 # Generated results
├── pyproject.toml          # Package configuration
├── README.md
├── CHANGELOG.md
└── LICENSE

---

Generated Outputs

After running the command, a folder is created in output/ with the model name. Inside you'll find:

• metadata.json
• summary: totals of tables, visible columns, visible measures and relationships.
• tables: each table with columns (type, visibility, category) and measures (clean expression, format, display folder, category).
• relationships: from/to, cardinality, direction and active status.

• model_documentation.md
• Model summary (language depends on --lang flag, default: English).
• List of tables (hidden or business), visible columns and measures grouped by category: revenue, cost, margin, percentage, ratio, temporal, etc.
• Sections with DAX expressions formatted with hierarchical indentation.
• Relationships table with visual representation of table connections.
• AI Agent Usage Guide with sample questions (translated based on selected language).

• agent_context.json
• Model name, totals, available tables, key measures (up to 20), temporal columns and sample questions (language depends on --lang flag, default: English).

• model_context.jsonl
• Line-delimited JSON format optimized for embeddings and RAG.
• Each line is an independent object (table, measure or relationship).
• Includes formatted DAX and sample prompts.

• index.json (new in v1.0)
• Lightweight model summary with per-table metadata (column/measure counts, categories) and relative paths to all other output files.
• Allows LLM agents to navigate large models without loading the full metadata.json.

• relationships.json (new in v1.0)
• All model relationships in a single focused file.

• tables/.json (new in v1.0)
• Full detail for one table (columns + measures including DAX).
• One file per table, addressable via index.json.

---

index.json — open format specification

index.json is a small, stable contract meant to be consumed directly by any tool — not just pbi-context' own CLI/resolver/MCP server: a lightweight per-table summary (name, column/measure counts, categories, format, and a relative path to that table's detail file) plus pointers to every other output file, so an agent can navigate a large model without loading metadata.json. By default it's written compact (no indentation); pass --pretty for indented JSON.

Full field-by-field contract — top-level shape, per-table entry, the tables/.json JSON vs. TOON shapes, and the versioning policy — is documented in docs/index-json-spec.md.

---

Use Cases

1. Automatic Dashboard Documentation

Problem: Your company has multiple undocumented Power BI dashboards. Analysts waste time searching for which measures to use and how tables are related.

Solution:

# Process all dashboards in a folder (English documentation)
pbi-context --batch "data/dashboards/*.pbit"

# Or generate Spanish documentation for all dashboards
pbi-context --batch "data/dashboards/*.pbit" --lang es

Result:

• Each dashboard generates its own documentation in output/[dashboard-name].pbit/
• Documentation ready to share with the team
• Automatic identification of measures by category (revenue, cost, margin, etc.)

Output example:

output/
├── Sales Dashboard.pbit/
│   ├── model_documentation.md  # 23 documented measures
│   └── metadata.json
├── Finance Dashboard.pbit/
│   ├── model_documentation.md  # 31 documented measures
│   └── metadata.json
└── Operations Dashboard.pbit/
    ├── model_documentation.md  # 18 documented measures
    └── metadata.json

---

2. New Analyst Onboarding

Problem: New employees need weeks to understand Power BI model structure and which measures to use for each analysis.

Solution:

1. Generate the model documentation (in your preferred language):

# English documentation (default)
pbi-context --input "data/pbit/my-model.pbit"

# Spanish documentation
pbi-context --input "data/pbit/my-model.pbit" --lang es

1. Upload the model_documentation.md file to your favorite AI agent (Claude, GPT-4, etc.)

1. The agent can answer questions like:

• "What revenue measures are available?"
• "How is Gross Margin calculated?"
• "What tables are related to Customer?"

Interaction example:

User: What revenue measures does this model have?

Agent: The "my-model" model has 11 revenue measures:
- Total Revenue (simple): SUM([Revenue])
- YTD Revenue (simple): TOTALYTD(SUM([Revenue]),'Date'[Date])
- Revenue SPLY (medium): CALCULATE([Total Revenue],SAMEPERIODLASTYEAR('Date'[Date]))
- Revenue Budget (medium): CALCULATE([Total Revenue], FILTER(Scenario, Scenario[Scenario]="Budget"))
...

Benefit: Significant reduction in onboarding time by having immediate answers about the model structure.

---

3. Model Auditing

Problem: You need to compare two versions of the same dashboard to identify which measures or relationships changed between releases.

Solution:

# Compare two versions of the model
pbi-context --diff "data/pbit/dashboard_v1.pbit" "data/pbit/dashboard_v2.pbit"

Result: A diff_dashboard_v1_vs_dashboard_v2.json file is generated, content-aware — not just which measures/columns/relationships were added or removed, but which existing ones changed content (DAX expression, format string, display folder, hidden flag, category, data type, cardinality, cross-filtering, active flag).

Identity for matching an object across both models:

• Measures and columns: (table, name).
• Relationships: (from_table, from_column, to_table, to_column) — a relationship that keeps the

same connected columns but changes cardinality/cross-filtering/active flag shows up in relationships_modified, not as a remove+add.

DAX changes are flagged "semantic" or "cosmetic" via a whitespace-insensitive compari

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Osc2405
• Source: Osc2405/pbi-context
• License: MIT
• Homepage: https://pypi.org/project/pbi-docs/

Install and usage instructions live in the source repository linked above.