Context Engineering and GraphRAG (IBM Technology)

Martin Keen on why model intelligence is no longer the bottleneck — context is. Maps out Context Engineering, the four pillars of a contextual system, and the family of RAG variants (RAG, GraphRAG, agentic RAG, context compression).

Key claims

• Context engineering = the system's ability to discover the right data, understand what it means, and apply it correctly in real time, within governance constraints. Not a prompt — an infrastructure problem.
• Example: "prep me for the client meeting tomorrow." A naive model gives a generic template. A context-engineered system pulls support tickets (open issue), deal history (renewal coming up), and excludes internal pricing (role-based access).
• Where context lives: databases, document stores, APIs, SaaS, cloud, on-prem. Some structured, some not. Some role-restricted. Getting it to the model at the right time with the right permissions is the hard part.
• Four pillars of a contextual system:
  1. Connected access — visibility across the data estate; zero-copy federation (query data where it lives, preserve original ACLs)
  2. Knowledge layer — entity resolution, relationships, hierarchies, decision traces, institutional knowledge
  3. Precision retrieval — filter by intent, role, time, policy. "Better context is not more context — it's more precise context."
  4. Runtime governance — enforced live at retrieval time AND response time. Can the agent query this? Should this result be included given who's asking?
• RAG = vector similarity over chunked + embedded docs. Great for simple lookups.
• Agentic RAG = iterative; agent does first pass, decides if it needs more, goes back. Step up from one-shot.
• GraphRAG = graph navigation: what entities are connected to this client, what docs relate to those entities? Vector search fills detail within graph-defined scope.
• Context compression = summarize and rank what reaches the model. Even with large context windows, more noise = worse results. Maximize signal.
• Punchline: model intelligence isn't the bottleneck; access to good context is. "A model is only as good as the context it can access."

Cross-source resonance

• Updates RAG significantly — the previous page treated RAG as the only flavor; this source expands the family.
• Karpathy's LLM Wiki Pattern is one realization of context engineering: the wiki IS the precision-retrieval + knowledge layer. The four pillars also map cleanly to the wiki:
  • Connected access ≈ raw/ + the index
  • Knowledge layer ≈ wiki cross-references and entity pages
  • Precision retrieval ≈ index-first lookup
  • Runtime governance ≈ schema in CLAUDE.md
• Praveen's "data access boundaries" (Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk)) is the same conversation from the enterprise CIO seat — APIs vs MCP vs .md policy files.

Cross-links

• IBM Technology · Context Engineering · RAG · GraphRAG · LLM Wiki Pattern