codejunkie99/graph-engineering

Purpose and Distribution

This repository is an agent-consumable “Graph Engineering” skill, not an executable application. It teaches and guides two connected practices: knowledge-graph construction and agent task-graph design. The source skill is graph-engineering/SKILL.md; the distributable artifact is dist/graph-engineering.skill.

Primary installation copies the skill directory into a skill-compatible agent harness. The expected entry interaction is a natural-language request to build a knowledge graph or teach graph engineering.

Skill Orchestrator

graph-engineering/SKILL.md is the central control document. It defines the agent’s operating modes, likely including build guidance and interactive teaching, and directs the agent through the knowledge-graph pipeline and task-graph principles.

WORKFLOWS.md provides alternate prompt-level entry points: an end-to-end tutor workflow plus focused commands that can be chained across a build. This is the user-facing orchestration layer for environments that do not invoke the skill directly.

Knowledge-Graph Pipeline

The core production model is a nine-stage, ordered pipeline:

1. scope

2. representation

3. ontology

4. entity extraction

5. relation extraction

6. event modeling/extraction

7. quality gate

8. fusion

9. LLM-facing serving

The key boundary is “model before extraction”: scope, representation, and ontology establish constraints before source facts are produced. Quality gates validate intermediate outputs; fusion resolves and reconciles records before graph storage or retrieval serving.

The conceptual pipeline is defined by graph-engineering/SKILL.md and decomposed into reusable workflow prompts in WORKFLOWS.md.

Domain Modeling and Extraction

Ontology and representation design are the schema boundary between an unstructured domain and graph data. graph-engineering/references/modeling.md supplies the modeling guidance, while graph-engineering/references/extraction.md covers deriving entities, relations, and likely source-grounded facts.

Extraction is treated as a staged transformation rather than direct triple generation: entities, relations, and events are separate concerns. This makes provenance, temporal context, validation, and later entity resolution possible.

Fusion and LLM Serving

Fusion is a required consolidation phase between extraction and storage/serving. It resolves duplicate or conflicting entities and facts, preserving graph quality rather than allowing raw extracted claims to become canonical graph records.

graph-engineering/references/fusion-and-llm.md defines the integration boundary with GraphRAG and LLM applications: the graph is a structured memory/retrieval substrate, not merely a collection of extracted triples.

Task-Graph Design

The second architectural half models how AI-agent work executes. graph-engineering/references/task-graphs.md defines task nodes as jobs and edges as actual execution dependencies.

The central execution pattern is: split independent work, run workers in parallel, verify outputs in separate contexts, then merge through a single accountable owner. This distinguishes genuine data/control dependencies from “fake edges” that add coordination without transferring work.

A stop rule prevents unnecessary multi-agent decomposition for sequential work, and a human gate is placed at costly-to-reverse decisions.

Reference Base and External Inputs

The repository packages distilled instructional material rather than runtime integrations. The reference corpus is organized across graph-engineering/references/curriculum.md, graph-engineering/references/modeling.md, graph-engineering/references/extraction.md, graph-engineering/references/fusion-and-llm.md, and graph-engineering/references/task-graphs.md.

Its external intellectual sources are Southeast University’s Knowledge Graph course, GraphRAG/LLM practices, and published multi-agent research from Google DeepMind, MIT, and Anthropic. No database, cloud service, web server, or application runtime is present in this repository.