sdcreader

Read a published Semantic Data Charter model's package for projection into other standards.

A projection describes a model's governed data records in another standard's terms: CDIF, DCAT-US 3.0, DCAT-AP 3.0.1, HealthDCAT-AP, and whatever comes next. What every projection describes is the same evidence, the model's package, and reads it into the same shape. sdcreader is that evidence and that shape, once, with nothing about any projection in it. The four projection writers (sdc_cdif, sdc_dcat3_us, sdc_dcat3_ap, sdc_healthdcat_ap) carried it as identical copies until this package.

pip install sdcreader            # the reader, with HTTP for fetching from the public catalog (since 4.1.0 no extra is needed)

from sdcreader import fetch_package, load_package, read_model

pkg = load_package("samples/nhanes-participant")          # a package on disk
pkg = fetch_package("xy8upneajsb8vdcmnve01g6g", save_to="pkg/")   # or from SDCStudio's public catalog, no account
model = read_model(pkg)

pkg.schema_url_pinned        # https://sdcstudio.axius-sdc.com/dmlib/dm-xy8u...xsd?sha256=0df4...   the exact bytes the records are validated against
model.title, model.leaves    # 'NHANES Participant', 153 leaves in document order, each with its path of clusters
model.codes[ct_id]           # the enumerated values of a component with the IRI, definition and code the schema annotates them with
model.dc("publisher")        # what the modeler wrote, None where SDCStudio's default sits

What a package is

SDCStudio generates a package for every published model and serves it publicly for models in public projects:

A package directory holds those four files. The reader refuses a directory without the JSON-LD and the schema, a JSON-LD whose identifier is not the directory's model, and a schema whose bytes are not the published current version: a projection cites the schema by URL and SHA-256, so the bytes have to be the ones that URL serves.

Two gaps in the package are on SDCStudio's issue tracker: the JSON-LD metadata drops fields the schema header carries (#748), and the header omits dc:source (#749). The reader takes the model's Dublin Core from the header, which is complete.

Related MCP server: TokenLab MCP Server

What the reader gives

• ModelPackage: the four files, the SHA-256, the schema URL and its pinned form (?sha256=), the public catalog URL, and check().

• Model: title, description, root (the governed record's cluster), components by identifier, leaves in document order (each a Leaf with its component, its path of clusters and a slug of that path, unique in the record), codes (the enumerated values per component as Code: value, IRI, definition, the code it is defined by), metadata (the JSON-LD's) and header (the schema's Dublin Core).

• Model.dc(name): what the modeler wrote, the header first and the JSON-LD second, with SDCStudio's field defaults as unset: "Universal" for coverage, "None" for relation, "SDC Data Model (DM)" for type, blanks. A projection never publishes a default as a fact. contributors, subjects (the semicolon list), rights_url and rights_statement derive from it.

• Component: sdc_type, label, description, data_type, constraints, links by predicate (link(p) for the first), contains, iri.

• Constants: the leaf, numeric and quantified types; the predicate IRIs IDENTIFIER, SOURCE, PUBLISHER, EXACT, CLOSE, SEE_ALSO, HAS_UNIT; DC_DEFAULTS.

Reading a record (4.3.0)

read_record(source, model) reads a Governed Data Record, the XML instance, and joins each value to the component that defines it. source is the record's bytes, its text or a path; a record of another model is refused (RecordError), as is anything that is not an sdc4:dm-<ct_id> record.

• Record: attributes (instance_id, instance_version, creation_timestamp, current-state, ...), values in document order, and exceptional, the values that carry an Exceptional Value.

• RecordValue: the component, its path of clusters (root first), instances parallel to the path (which occurrence each cluster is among the same cluster's repeats under the same parent; instance_key pairs them with the clusters' identifiers, one key per cluster instance), an occurrence index for a component repeated within its cluster instance, its position, its elements as the record holds them (a quantity's units nested: {"xdquantity-value": "1", "xdquantity-units": {"label": "Trade Units", "xdstring-value": "EA"}}), and its context (act, vtb, vte, tr, modified, latitude, longitude).

• ExceptionalValue: the ISO 21090 null flavor (code, name) and its origin: source when it stands in place of the value, recovery when sdcvalidator's recovery inserted it beside a value that failed its constraint. For recovery, RecordValue.rejected is that value, kept so a person can see it.

Cluster members sit inside adapter elements that are not components of the model; the reader passes through them. It does not type, validate or interpret values: that is the caller's work (a mapper, an exporter, a checker).

MCP server

sdcreader is also an MCP server, registered as io.github.SemanticDataCharter/sdcreader, so an assistant can read a published model the way sdcvalidator, sdcgovernance and sdcreceipt let it validate, govern and settle a record. stdio, no SDK, no account, read-only.

uvx sdcreader                    # or: sdcreader-mcp, or python -m sdcreader.mcp_server
uvx sdcreader --host https://your-sdcstudio.example   # a sovereign SDCStudio's public catalog

Each model tool takes a ct_id (fetched from the host the server was started with) or a package directory on disk. No tool takes a URL: the host is fixed at start, so a caller cannot point the server elsewhere. Every tool calls the same functions as the library; the two interfaces describe one model the same way.

Also listed on mcprush.

Tests

pip install -e ".[dev]"
python -m pytest tests -q                    # includes one fetch against production
python -m pytest tests -q -m "not network"   # offline, as CI runs it

The fixture is the NHANES Participant package of the FAIR Data Demo, the record every projection is proven on.

Versioning

The leading 4 is the SDC4 reference model, as with every Semantic Data Charter package. An addition to what the reader exposes is a minor version, a fix a patch; the major changes only when the reference model does.

License

Apache-2.0 (see LICENSE, NOTICE).