Agent Skills for Creating and Using Ontologies with LLMs

---
name: ontology-builder-assistant
description: >
  derive a minimal reusable ontology from purpose statements, competency questions,
  sample data, reusable vocabularies, supporting semantic evidence, and implementation
  constraints. use when you're asked to draft an ontology for a specific use case.
  Especially useful for ontology bootstrapping, information extraction schemas,
  mapping-oriented semantic models, and defensible cq-to-model traceability with
  final serialization.
---

## Inclusion test
Every included element must pass both tests:
1. requirement test: needed by an explicit requirement or competency question.
2. evidence test: supported by representative data.

## Depth test
- maximum taxonomy depth: 3
- prefer properties over subclasses
- add a third level only when a competency question or mapping task requires it

## Grounding test
- choose a small top-level grounding scheme
- keep top-level classes mutually disjoint
- assign each included class to one grounding branch

## Serialization test
- include only final included classes and properties
- assert top-level disjointness
- avoid speculative axioms
- add domain/range only when stable and useful

## Definition test
- every class gets an informative definition
- use Aristotelian form when possible
- avoid circular or generic wording

# run_kg_pipeline.py
from neo4j_graphrag.experimental.components.schema import GraphSchema
from neo4j_graphrag.experimental.pipeline.kg_builder import SimpleKGPipeline
from neo4j_graphrag.llm import OpenAILLM
from neo4j_graphrag.embeddings import OpenAIEmbeddings
import neo4j

async def main(schema_file: str) -> None:
    schema = GraphSchema.from_file(str(ONTOLOGIES_DIR / schema_file))

    llm = OpenAILLM(
        model_name="gpt-4o",
        model_params={
            "response_format": {"type": "json_object"},
            "temperature": 0,
        },
    )
    embedder = OpenAIEmbeddings(model="text-embedding-3-small")
    driver = neo4j.GraphDatabase.driver(
        NEO4J_URI, auth=(NEO4J_USER, NEO4J_PASSWORD), database="medical"
    )

    pipeline = SimpleKGPipeline(
        llm=llm,
        driver=driver,
        embedder=embedder,
        schema=schema,
        from_pdf=False,
        perform_entity_resolution=True,
    )

    for doc_path in sorted(DATA_DIR.glob("*.txt")):
        text = doc_path.read_text(encoding="utf-8")
        print(f"Processing {doc_path.name} ({len(text)} chars) …")
        result = await pipeline.run_async(text=text)
        print(f"  result: {result}")

@prefix ccs: <https://w3id.org/goingmeta/ccs#> .
@prefix owl: <http://www.w3.org/2002/07/owl#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .

ccs:Party a owl:Class ;
    rdfs:label "Party" ;
    rdfs:comment "An agent or person that participates in clinical events and interactions." .

ccs:Event a owl:Class ;
    rdfs:label "Event" ;
    rdfs:comment "A temporal occurrence in a clinical context." .

ccs:ClinicalFinding a owl:Class ;
    rdfs:label "ClinicalFinding" ;
    rdfs:comment "An observation, measurement, assessment, or finding made in a clinical context about a patient." .

# Mutual disjointness of top-level classes
[] a owl:AllDisjointClasses ;
    owl:members ( ccs:Party ccs:Event ccs:ClinicalFinding ) .

ccs:Patient a owl:Class ;
    rdfs:subClassOf ccs:Party ;
    rdfs:label "Patient" ;
    rdfs:comment "A Patient is a Party who is the subject of clinical care and documentation in a case sheet." .

ccs:Encounter a owl:Class ;
    rdfs:subClassOf ccs:Event ;
    rdfs:label "Encounter" ;
    rdfs:comment "An Encounter is an Event that represents a clinical contact between a patient and a healthcare setting." .