Introduction

OWL-SDA (OWL Synthetic Data AI-Agent) is a Java tool that generates synthetic RDF instance data from OWL ontologies using Large Language Models (LLMs).

Given an input ontology (in Turtle or RDF/XML format) and a natural-language prompt describing the desired data, OWL-SDA:

1. Loads and optionally reasons over the ontology (including fetching external imports).
2. Derives SHACL shapes from the OWL class descriptions.
3. Dispatches a supervisor agent that coordinates a pool of worker agents, each responsible for generating instances for a subset of SHACL shapes.
4. Validates the generated triples against the SHACL shapes after each round.
5. Passes the final output to a reviewer agent that either accepts it or requests targeted revisions.

How a run flows

The generation loop is bounded: consecutive rounds that delegate to nobody, make no progress, or leave the same violations untouched abort the run instead of looping forever. Review iterations are capped by client.reviewer.max-review-attempts, and the last attempt must end in ACCEPTED or REJECTED.

Key Concepts

Agents

Role	Responsibility
Supervisor	Orchestrates the generation loop, assigns shapes to workers, and finalises the output.
Worker (pool)	Generates RDF triples for its assigned shapes and writes them to the shared triple store.
Reviewer	Reviews the fully assembled output and signals ACCEPTED, REJECTED, or REVISION_REQUESTED.

Each role has its own model, timeouts, and tool permissions. See the Configuration Reference and Architecture.

SHACL Shapes

OWL-SDA automatically generates SHACL NodeShapes from owl:Class definitions, including property shapes derived from owl:Restriction axioms. These shapes are used both to guide the workers and to validate the output after each generation round. Shapes are derived twice — from the asserted model and from the reasoner's inferred model — and can be cached to disk via shacl.output-dir.

Triple Store

Workers write triples directly into a shared in-memory triple store. After each delegation round the store is flushed to the configured output file so that the supervisor and reviewer can read it.

User Context

Any number of local files or live URLs can be attached as extra context for the agents — domain notes, worked examples, a PDF, or a specification fetched at runtime. External ontologies referenced through owl:imports are fetched, mirrored, and cached separately; see External Ontologies.

Observing a run

Runs are long, so OWL-SDA can show you what it is doing while it works:

• Benchmark snapshots capture every round — duration, shapes processed, triple count, violations, per-role token usage, and the full message log. See Benchmarking.
• A built-in web dashboard, started with --web-ui, reads those snapshots live and serves them at http://localhost:8080: agent messages and tool calls, the generated output, the triple store, trend charts, and a SHACL violations view. It needs benchmark.enabled: true to have anything to show. See Web UI.

Supported LLM Backends

OWL-SDA supports GitHub Copilot (via the standalone Copilot CLI), Ollama (self-hosted models), and any OpenAI-compatible endpoint (bring-your-own-key — OpenAI itself, Azure OpenAI, or self-hosted gateways) as language model backends. Providers can be mixed per role: cheap local workers with a stronger hosted supervisor, for example. Support for additional providers can be added by implementing the Client / Session interfaces.

Next Steps

• Installation — build the jar and set up a backend.
• Quick Start — a first run on a minimal ontology.
• Examples — three complete, ready-to-run projects.

License

OWL-SDA is distributed under the GNU General Public License v3.0. See License for details and the full license text.