Knowledge Graph Construction

Version: 1.0.0 | Last updated: 2026-07-16 | Maturity: Emerging for LLM-assisted construction

Purpose

Construct relationship-centric knowledge with explicit ontology, provenance, and review.

Why

Graphs help multi-hop, entity, and relationship queries, but LLM extraction can create plausible unsupported edges at scale.

How

Start only after relationship-heavy queries beat document retrieval in a pilot. Define competency questions and a versioned ontology; assign stable entity IDs. Extract candidate entities/relations with source spans, confidence, extractor version, tenant/ACL, and valid time. Validate schema deterministically, reconcile entities conservatively, and require human or trusted-rule approval for high-impact edges. Publish immutable graph versions and rollback manifests.

Tradeoffs

Graphs improve explicit traversal and conflict representation but add ontology, entity-resolution, and stewardship cost. Keep source documents authoritative.

Anti-patterns

Creating edges without source-span provenance.
Global fuzzy entity merges.
Building a graph because it appears more advanced than hybrid search.

Enterprise Considerations

Tenant-scope nodes and edges, govern ontology changes, and audit merges/splits. LLM-generated graph construction remains emerging and requires corpus-specific validation.

Checklist

Competency questions justify the graph.
Ontology and entity identity are versioned.
Every edge has source, time, ACL, and extraction provenance.
Merge, rollback, and correction tests pass.

References

Changelog

1.0.0 — 2026-07-16: Initial emerging-practice standard.