triplets 0.2.0

About:

• Parses CIM RDF/XML data to pandas dataframe with 4 columns [ID, KEY, VALUE, INSTANCE_ID] (triplestore like)
• The solution does not care about CIM version nor namespaces
• Input files can be xml or zip files (containing one or mutiple xml files)
• All files are parsed into one and same Pandas DataFrame, thus if you want single file or single data model, you need to filter on INSTANCE_ID column

Documentation:

https://haigutus.github.io/triplets

Upgrading from 0.0.x? See docs/migration_0.0_to_0.1.md.

To get started:

# Core (python_lxml_pandas engine, no extra deps)
pip install triplets

# With pyarrow (enables python_lxml_arrow + cython_pugixml_arrow engines, ~10x faster)
pip install triplets[arrow]

Install extras by feature:

The embedded qlever SPARQL engine (fastest) ships in no wheel — it is a local source build, see docs/building.md.

import pandas
import triplets

path = "CGMES_v2.4.15_RealGridTestConfiguration_v2.zip"
data = pandas.read_RDF([path])

Result:

You can then query a dataframe of all same type elements and its parameters across all [EQ, SSH, TP, SV etc.] instance files, where parameters are columns and index is object ID-s

data.tableview_by_type("ACLineSegment")

Export:

from triplets.export_schema import schemas
from triplets.export import ExportType

data.export_to_cimxml(
    rdf_map=schemas.ENTSOE_CGMES_2_4_15_552_ED1,
    export_type=ExportType.XML_PER_INSTANCE_ZIP_PER_XML,
)

Export schemas are versioned and shipped per profile. Alongside the CGMES bundles (schemas.ENTSOE_CGMES_2_4_15_552_ED1, schemas.ENTSOE_CGMES_3_0_0_552_ED1, …) the versioned NC (Network Code) profiles are available as schemas.ENTSOE_NC_2_4_1_552_ED1 / schemas.ENTSOE_NC_2_4_1_552_ED2. The _ED1 / _ED2 suffix selects the serialization edition; profile resolution is schema-driven, so the right profile section is matched from the instance header.

Look into examples folders for more

Parser engines

Three parser engines with automatic fallback (fastest available):

The cython_pugixml_arrow engine is a compiled C++ extension included in published wheels. It requires pyarrow at runtime, so install with triplets[arrow] to enable it.

The cython engine is pre-built in published wheels — no compilation needed.

Engine selection is automatic across the library (parser, exports, SPARQL, validation): installing an extra makes everything that can use it faster, with no code changes. Inspect and steer it globally:

triplets.engines()                        # what "auto" resolved to, per subsystem
triplets.set_engine(parser_cimxml="python_lxml_pandas", sparql="rdflib")
triplets.set_engine(parser_cimxml="auto") # restore auto-selection

Per-call engine= arguments always win over set_engine. Operations on your DataFrame itself (filters, tableviews, references) always run in the engine of the object you call them on — pandas frames stay pandas, polars stays polars.

Polars

import polars
import triplets

data = polars.read_rdf(["grid_EQ.xml", "data.zip"])   # returns polars DataFrame

data.triplets.get_types_count()
data.triplets.tableview_by_type("ACLineSegment")
data.triplets.filter_triplets(KEY="Type", VALUE=".*Generator.*", regex=True)
data.triplets.export_to_csv(export_to_memory=True)
data.triplets.export_to_nquads("/tmp/output.nq")
data = polars.read_nquads("/tmp/output.nq")           # round-trips N-Quads back to triplets

read_nquads is registered as pandas.read_nquads / polars.read_nquads and is also exposed top-level as triplets.read_nquads.

DuckDB

import duckdb
import triplets

data = duckdb.connect()                              # default table "triplets"
data = duckdb.connect("grid.duckdb", table="grid", schema="cim")  # per-connection defaults
# explicit table/schema config is stored in the database file — reopening
# duckdb.connect("grid.duckdb") later resolves cim.grid automatically

data.read_rdf(["grid_EQ.xml", "data.zip"])           # streams into the connection's table
data.read_rdf(["update.zip"], append=True)           # adds rows instead of replacing
data.get_types_count()                               # uses connection table/schema
data.tableview_by_type("ACLineSegment").df()
data.filter_triplets(KEY="Type", VALUE=".*Sub.*", regex=True).df()
data.references_to("some-uuid").df()
data.export_to_nquads("/tmp/output.nq")

# Per-call override; rebind defaults with set_triplets_table(...)
data.types_dict(table="other", schema="main")

# Direct SQL (your identifiers — tools always quote theirs)
data.sql('SELECT VALUE, COUNT(*) FROM "cim"."grid" WHERE KEY = \'Type\' GROUP BY VALUE').df()

# The same tools are also on the `.triplets` namespace (parity with pandas/polars)
data.triplets.tableview_by_type("ACLineSegment").df()
data.triplets.get_types_count()

SPARQL queries

SPARQL 1.1 over the loaded data — SELECT → DataFrame, ASK → bool, CONSTRUCT → triplet DataFrame. Works on pandas, polars and DuckDB inputs:

PREFIXES = """
PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX cim: <http://iec.ch/TC57/CIM100#>
"""
names = data.sparql.query(PREFIXES + "SELECT ?s ?name WHERE { ?s cim:IdentifiedObject.name ?name }")

Three engines behind one API (auto picks the fastest available):

Details and measured numbers: docs/sparql.md.

SHACL validation

Validate against SHACL shape files; the result is a violations DataFrame (empty = conforms) with the same shape across all engines:

from triplets.export_schema import schemas

violations = data.shacl.validate("shapes.ttl", rdf_map=schemas.ENTSOE_CGMES_3_0_0_552_ED1)

# slower optional context pass: source file, object type/name,
# shape sh:name/sh:description, schema attribute/class definitions
violations = data.shacl.validate(shapes, context=True)

# SARIF 2.1.0 for GitHub / SonarQube / any SARIF viewer — grouped by default
# (one result per rule with occurrenceCount + sample instances)
violations.shacl.to_sarif(path="report.sarif")

# standard SHACL sh:ValidationReport — format from path suffix (or format=)
violations.shacl.to_shacl_report(path="report.ttl")
violations.shacl.to_shacl_report(path="report.xml", report_source="model.xml",
                                 report_references=["equipment.ttl"])

Engines: polars (auto, real profiles in ~2 s) → pandas → pyshacl (reference); duckdb for larger-than-memory data. sh:sparql constraints ride the SPARQL engine above (minutes → milliseconds with oxigraph/qlever). Details: docs/validation.md.

Accessor namespace

pandas and polars DataFrames use df.triplets.*; a DuckDB connection uses con.triplets.*. The same method names are available on both (DuckDB returns relations — add .df() or .pl() when needed):

# pandas / polars
df.triplets.tableview_by_type("ACLineSegment")
df.triplets.export_to_nquads("/tmp/output.nq")

# DuckDB
con.triplets.tableview_by_type("ACLineSegment").df()
con.triplets.get_types_count()

Root-level methods (df.type_tableview(...), con.filter_triplets(...)) still work for backwards compatibility.

Cache lifecycle

Engines keep internal state (compiled shapes, SPARQL indexes) cached across calls. Reset it explicitly with triplets.clear_caches(), or scope it to a block so it is cleared on exit:

import triplets

with triplets.cache_scope():
    ...            # caches populated here are dropped when the block exits

triplets.clear_caches()   # or clear everything manually

CLI tools

cim-spreadsheet -i model.xml -o output.xlsx
cim-diff original.xml modified.xml

Performance (RealGrid, 1.14M rows)

Committed benchmark results live in tests/performance_results/; re-run with pytest -m performance. Representative numbers (cython parse 1.47s → 0.157s vs the lxml engine = ~9.4x):

The old rdf_parser.py functions still work but emit deprecation warnings. See docs/migration_0.0_to_0.1.md for renames and breaking changes.