fabric-ai-meta

Make any Power BI semantic model readable by AI, from your laptop.

Point it at a .pbip folder and you get a classified schema, an AI readiness score, and framework-native exports for LangChain, OpenAI, Semantic Kernel, and AutoGen. No Fabric tenant, no notebook, no sign-in.

Try it

In Power BI Desktop: File > Save As > .pbip. Then:

pip install fabric-ai-meta
fabric-ai-meta analyze "Sales" --pbip ./Sales.SemanticModel

That reads the local TMDL, classifies every table and measure, scores the model, and writes ./output/sales/:

ai-ready-schema.json          # tables, measures, relationships, all classified
readiness-score.json          # {"score": 0.82}  <- how AI-ready this model is
langchain-tool.json           # drop straight into LangChain
openai-function.json          # and into OpenAI function calling
semantic-kernel-plugin.json   # and Semantic Kernel  (export autogen adds the fourth)
measure-dependency-graph.json
extraction-raw.json

It parses your DAX, so a TOTALYTD(...) measure comes back understood, not guessed:

{ "name": "Sales YTD", "category": "time_intelligence", "requires_date_filter": true }

No model handy? fabric-ai-meta analyze "Adventure Works" --mock runs the same flow on bundled fixtures.

Your models never leave your machine

Table names, measure logic, and business rules describe how your company works. You never have to trust this tool with any of it.

• Local by default. --pbip reads TMDL off your disk, --mock uses bundled fixtures. Neither touches a network or an account.
• No telemetry. The only outbound calls in the codebase go to the Fabric REST API and to the LLM provider you configure. Both are opt-in.
• LLM enrichment is opt-in and capped. Nothing is sent anywhere without --llm-enrich. You pick the provider (10+, including local Ollama), you supply the key, and max_cost_per_run stops an overspending run.
• Writeback is dry-run by default. apply-descriptions and apply-copilot show the diff and change nothing until you pass --no-dry-run.

Live workspace extraction is the one thing that needs Fabric, because sempy.fabric only exists in the notebook runtime. Everything else runs anywhere Python does.

Commands

Every command takes --pbip <folder>, --mock, or --workspace <name>. Worked examples for all of them are in the user guide.

Add --llm-enrich to any extraction command to fill in missing descriptions and detect fact-table grain. It is off unless you ask, cost-capped, and works with 10+ providers including a local Ollama.

Do you even need an ontology?

Knowledge graph projects are expensive and most semantic models do not need one: a star schema with good descriptions already answers the questions people actually ask. Run this before you fund the project.

fabric-ai-meta governance --pbip ./models --graph-necessity --report ./gov.json

The report gains a verdict per model:

{
  "name": "Contoso Sales",
  "tier": "GRAPH_UNNECESSARY",
  "pressure": 0.0,
  "confidence": "evidenced",
  "evidence": [
    "4 of 4 questions resolve within <=2 tables (flat aggregation)",
    "no bridge or many-to-many relationships",
    "relationship graph is a shallow star (diameter 2)",
    "single fact table"
  ],
  "recommendation": "Described schema suffices; defer ontology/graph adoption."
}

Tiers are GRAPH_UNNECESSARY, GRAPH_OPTIONAL, and GRAPH_WARRANTED, blended from four signals: how many real questions traverse three or more tables, bridge and many-to-many presence, relationship graph depth, and multi-fact complexity. No Fabric capacity, no LLM calls.

Pass --questions ./questions.txt (one per line, or a JSON list) to judge against what your users actually ask. Without it the check falls back to Copilot example prompts, then to measure dependencies, which is a conservative proxy: DAX names columns in one or two tables while real traversal happens through relationships at query time. That fallback biases toward GRAPH_UNNECESSARY, so supply real questions when you want the verdict to carry weight.

Why this exists

Microsoft's AI features for semantic models (Prep for AI, Copilot descriptions, Data Agents, Fabric IQ Ontology) share three limits. Prep for AI is configured by hand, one model at a time, with no bulk API. Nothing exports outside Fabric, so a LangChain or function-calling pipeline starts blind. And nothing compares models, so Total Sales in one and Sum of Sales in another stay invisibly identical.

This is not a replacement for those tools. It is an automation layer on top of them and a bridge to the AI ecosystem outside Fabric.

Library API

from fabric_ai_meta import MockExtractor, score_model, generate_ai_ready_schema, to_openai_function

model = MockExtractor(fixture_path="tests/fixtures/adventure_works.json").extract("Adventure Works")
score, breakdown = score_model(model)
schema = generate_ai_ready_schema(model)
openai_fn = to_openai_function(model)

47 public exports; see fabric_ai_meta.__all__.

Custom exporters

Subclass BaseExporter, register it under the fabric_ai_meta.exporters entry point group, and it appears as fabric-ai-meta export <name> with the same flags as the built-ins. No fork needed. Worked dbt example: docs/plugin-development.md.

Docs

Contributing

Issues and pull requests welcome. Before opening one:

pip install -e ".[dev]"
pytest tests/ -x -q     # 558 tests, no Fabric runtime or network needed
ruff check .

New exporters ship as plugins rather than PRs here. Fixture models under tests/fixtures/ and doc fixes are the easiest first contributions.

If this saved you time, star the repository. That is the signal I use to decide what to build next.

License

MIT. See LICENSE.

Built by Prasanth Sistla. Not affiliated with, endorsed by, or sponsored by Microsoft. "Microsoft Fabric", "Power BI", and "Copilot" are trademarks of Microsoft Corporation.