I’ve been working through the idea of context engineering lately (as well as the related context sovereignty) lately, and had this thought, which may be a way for others to help contextualise the implications of this relatively new approach to interacting with language models.
GraphRAG and knowledge graphs are to context engineering, what RAG and vector databases are to prompt engineering.
Prompt engineering
Prompt engineering = retrieval augmented generation (RAG) + vector databases
- Manual process: Crafting prompts by manually researching and including relevant information
- Automated process: RAG automatically finds and retrieves relevant text chunks from vector databases based on semantic similarity
- Core function: Augmenting prompts with relevant information to improve LLM responses
- Information architecture (i.e. vector database): Flat, similarity-based retrieval of text snippets
- Reasoning capability (RAG): “Find similar information”
- Problem space: RAG addresses the question, “How do I get relevant information into my prompt?”
Context engineering
Context engineering = GraphRAG + knowledge graphs
- Manual process: Manually structuring information, defining relationships, and organising contextual frameworks
- Automated process: GraphRAG automatically constructs and queries knowledge graphs to provide structured, relationship-aware context
- Core function: Engineering sophisticated contextual understanding through structured relationships
- Information architecture (i.e. knowledge graph): Hierarchical, relationship-based retrieval of interconnected concepts
- Reasoning capability (GraphRAG): “Find related concepts, understand relationships, reason across connections”
- Problem space: GraphRAG addresses the question, “How do I create sophisticated understanding of complex, interconnected information?”
Differences between prompt and context engineering
Prompt engineering involves:
- Crafting effective inputs
- Managing context windows
- Optimising for specific outputs
- Manual iteration and refinement
Context engineering involves:
- Designing information architectures
- Structuring knowledge relationships
- Creating reasoning pathways
- Systematic knowledge organisation
Implications
- Context engineering is a distinct system-level discipline
- Information architecture is as important as information retrieval
- Relationship-aware AI replaces similarity-based AI
- Automated reasoning over complex domains becomes the norm