PBIXRay

Read Power BI semantic models without Power BI.

No Power BI Desktop. No Analysis Services instance. No XMLA endpoint, no Premium capacity, no workspace, no restore. Just a file path.

from pbixray import PBIXRay

model = PBIXRay("sales.pbix")

model.dax_measures         # every measure, as a DataFrame
model.relationships        # the full model graph
model.power_query          # all M code
model.get_table("Sales")   # actual decoded rows

PBIXRay is a Python reader for the VertiPaq engine's on-disk format. It parses .pbix files, .abf Analysis Services backups and .xlsx PowerPivot workbooks, which are three containers around the same storage engine, and returns pandas DataFrames. Read-only, offline, cross-platform, with no Microsoft runtime involved at any stage.

Every other way to inspect a model needs something running. Desktop open, a server connected, a backup restored, a capacity licensed. PBIXRay needs a file.

Install

pip install pbixray

Python 3.8 through 3.13, on macOS, Linux and Windows. Prebuilt wheels, so there is nothing to compile.

What you can pull out

Full reference and examples at pbixray.com/docs.

Three formats, one API

File type is detected from the contents, so the same code works across all three.

PBIXRay("report.pbix")
PBIXRay("nightly-backup.abf")
PBIXRay("legacy-powerpivot.xlsx")

ABF support turns a backup archive into something queryable without a restore. That covers retention audits, migration inventories, and answering "what was in this model in 2019" without standing anything up.

The 40 DMVs, from a file

Analysis Services exposes model metadata through $System.TMSCHEMA_* DMVs, normally reachable only over a live connection. PBIXRay reads all forty straight from the embedded metadata database.

model.tmschema_refresh_policies    # incremental refresh configuration
model.tmschema_role_memberships    # who is in which security role
model.tmschema_column_permissions  # object-level permissions
model.tmschema_partitions          # partition definitions and sources

This makes governance tooling possible in places a live connection is not available. A pull request, a Lambda function, an air-gapped audit, a laptop with no license.

Models bigger than your RAM

Decompressed models are memory-mapped from disk rather than loaded whole, and tables stream by VertiPaq segment. Column projection and categorical strings cut memory further.

with PBIXRay("20gb-model.pbix", on_disk=True) as model:
    for chunk in model.iter_table("FactSales", chunk_size=1_000_000):
        process(chunk)

Dictionary decoding runs on a native Huffman kernel (xmhuffman) and fans out across cores.

Built on PBIXRay

🖥️ PBIXRay for macOS

A native model inspector for Mac. Open a .pbix and browse tables, measures, relationships and storage statistics, with no Windows VM, no Parallels and no Power BI Desktop. It streams one table at a time, so models too large for memory open fine. Spotlight integration finds measures by name.

Built for the Mac-based BI consultants Microsoft has never shipped a tool for.

🌐 pbix.info

Drop a model into the browser and explore it straight away. Metadata only, with no data or statistics extracted, focused on data origin and Power Query lineage. Nothing to install.

🦆 DuckDB extension

Query PBIX files directly in SQL.

Scope

PBIXRay is a read-only extractor for the data model. It does not:

• write, modify or repack files
• evaluate DAX, so expressions come back as source text
• run a query engine
• connect to Power BI Service, Analysis Services, gateways or workspaces
• refresh anything
• parse the report layer, meaning visuals, pages, bookmarks and themes
• support .pbit, .pbids or .pbip

Read-only is a deliberate choice. A library that cannot write to a model also cannot corrupt one, which is what makes it safe to point at production artifacts and client files.

Contributing

The PBIX format is undocumented and reverse-engineered, so the test corpus is the specification. The most valuable contribution is a model that parses incorrectly.

If you find one, open an issue with the failure output and the structural details, such as encoding type, column metadata and offsets. Please never send file contents you do not own. A minimal reproduction is more useful than a real model and safer for everyone.

Built on

Decompression uses Microsoft's own MIT-licensed Xpress reference implementations, wrapped for Python as xpress8 and xpress9. Huffman dictionary decoding lives in xmhuffman. All three are kept in separate repositories so the Cython build and wheel distribution stay out of the main library.

Links

Documentation · Interactive demo · PyPI · Support

MIT licensed. Built in London by Alphaverse Limited.

Not affiliated with or endorsed by Microsoft. Power BI, Excel and Analysis Services are trademarks of Microsoft Corporation.