Start from questions, not a giant ontology
Write the decisions the system must support. If the goal is incident impact analysis, begin with questions such as “Which services depend on this component?” and “Who owns the affected service?” A graph built without real queries often becomes a collection of labels nobody uses.
The first build sequence
- Define a small set of entity and relationship types.
- Choose trusted sources and record provenance for each fact.
- Extract candidate triples, but validate them deterministically or with a reviewer.
- Store identifiers, dates, ownership, and access rules.
- Test the graph against multi-step questions.
- Add vector retrieval only where text evidence is still needed.
The part AI should not decide alone
A model can suggest that two entities are related, but it should not silently create an authoritative relationship. Keep proposed facts separate from approved facts and retain the source passage that supports each edge.