Examples

The repository includes three complete example projects that you can run as-is.

Every invocation needs an explicit --config; the option has no default. Build the jar first (see Installation) — mvn clean package produces target/owlsda.jar.

A run overwrites the example's output

OWL-SDA deletes the file at output-path before it starts, and archives any previous benchmark directory. Point output-path somewhere else first if you want to keep what is already there.

Project 1 - Glass Production Facility

examples/project-1/ demonstrates a production-process ontology with temporal variants.

What it generates

Given the prompt:

"Create a data example for a glass production facility. Introduce multiple versions of the operation in different years with different systems, processes and subprocesses."

OWL-SDA generates RDF instances for facilities, systems, processes, and subprocesses across multiple years, illustrated using a glass manufacturing scenario (installations like furnaces, processes such as melting/annealing, and measurement points for emissions).

The config also shows off context-window-tokens, an openai-compatible provider pointed at a self-hosted gateway, and a long list of extract.mirrors for external ontologies.

Running the example

java -jar target/owlsda.jar --config examples/project-1/config.yml

Key files

File	Description
examples/project-1/config.yml	Run configuration.
examples/project-1/input.ttl	OWL ontology used as generation input.
examples/project-1/input-example.txt	Extra user context passed to the agents.
examples/project-1/reasoner.rules	Custom Jena rules for the reasoner.
examples/project-1/shacl/	Cached SHACL shapes (default-shacl.ttl, inferred-shacl.ttl), reused across runs.
examples/project-1/benchmark_paper/	Curated benchmark snapshot used for the plots below.

A run additionally writes output.ttl (the generated triples), inferred.ttl (the reasoner's inferred model), external/ (cached external ontologies) and benchmark/ (snapshots) into examples/project-1/. None of these are shipped in the repository.

Project 2 - Patient History Record

examples/project-2/ demonstrates richer generation on a larger medical ontology input.

What it generates

Given the prompt:

"Create a patient history record for the ontology."

OWL-SDA generates linked patient history data aligned with the classes and properties in owlapi.xrdf.

This config mixes providers: ollama workers with a copilot supervisor and reviewer.

Running the example

java -jar target/owlsda.jar --config examples/project-2/config.yml

Key files

File	Description
examples/project-2/config.yml	Run configuration.
examples/project-2/owlapi.xrdf	Main ontology input (RDF/XML).
examples/project-2/2504.08750v1.pdf	Domain paper passed as user context.
examples/project-2/reasoner.rules	Custom Jena rules for the reasoner.
examples/project-2/shacl/	Cached SHACL shapes.
examples/project-2/output.ttl	Output kept from an earlier run, so you can inspect the result without running anything. Deleted at the start of your next run.
examples/project-2/benchmark_paper/	Curated benchmark snapshot used for the plots below.

inferred.ttl, external/ and benchmark/ appear here after a run; they are not shipped.

Project 3 - LDES Parking Garage Observations

Work in progress - not fully benchmarked yet.

examples/project-3/ demonstrates LDES-oriented dummy data generation for parking availability observations. It also showcases URL-based user context: the LDES specification is fetched live from the web at runtime rather than loaded from a local file.

What it generates

Given the prompt:

"Generate dummy LDES data for a parking garage with observations about available spaces over time."

OWL-SDA generates parking observation instances linked to a stream resource, with timestamps and availability values aligned to the ontology constraints.

Running the example

java -jar target/owlsda.jar --config examples/project-3/config.yml

Key files

File	Description
examples/project-3/config.yml	Config using both local and URL-based user context (https://semiceu.github.io/LinkedDataEventStreams/).
examples/project-3/input.ttl	Main ontology input for parking and observation terms.
examples/project-3/input-example.txt	Domain constraints passed as extra context.
examples/project-3/reasoner.rules	Empty placeholder — the reasoner runs without custom rules.
examples/project-3/shacl/	Empty until the first run generates and caches the shapes.
examples/project-3/output.ttl	Empty placeholder; filled by a run.

inferred.ttl, external/ contents and benchmark/ are likewise produced by a run.

Benchmark Result Images

The plots below are generated from each example's benchmark-summary.json with scripts/plot_benchmark.py.

Benchmark time is calculated as the sum of durationMs for all iterations except the final REVIEW_* iteration, because that final review snapshot is already an aggregate of the review stages.

TIP

examples/project-N/benchmark_paper/ is a hand-curated snapshot kept for the paper. It was produced by an earlier version of OWL-SDA and uses a flat layout; a run today writes live/ and archive/<timestamp>/ under benchmark.output-dir instead. See Benchmarking.

Project 1 benchmark

Glass Production Facility · 11 rounds · ~27.9 min total · 297 final triples · 0 violations at acceptance · ~2.1M tokens

PNG fallback

Project 2 benchmark

Patient History Record · 22 rounds · ~54.8 min total · 2,698 final triples · 0 violations at acceptance · ~2.7M tokens

PNG fallback

Compare the benchmark raw data

• examples/project-1/benchmark_paper/benchmark-summary.json
• examples/project-2/benchmark_paper/benchmark-summary.json
• examples/project-1/benchmark_paper/benchmark_stats.txt
• examples/project-2/benchmark_paper/benchmark_stats.txt

Watch a run of your own live in the browser with --web-ui — see the Web UI guide.