LLM Knowledge Engineering Ecosystem Map

KNOWLEDGE SOURCES

RAG Pipelines Ch02

Self-RAG, Corrective RAG
Adaptive RAG, Agentic RAG
1,250x cheaper than long context
71% of enterprises use RAG

Knowledge Graphs Ch02

Microsoft GraphRAG
Entity-relationship reasoning
Multi-hop query resolution

Long Context Windows Ch02

1M+ token windows (Claude, Gemini)
Best for static, bounded docs
"Lost in the middle" degradation

feeds into

CONTEXT ENGINEERING

Dynamic Context Assembly Ch03

6-layer context model (Anthropic)
KV-cache hit rate = key metric
100:1 input-to-output ratio

System Prompts + Instructions Ch03

Rules, constraints, persona
Routing tables for skill dispatch

Tool Definitions + Schemas Ch07

MCP protocol standard
97M+ monthly SDK downloads
Meta-tool pattern for efficiency

Conversation + Memory Ch06

Episodic / Semantic / Procedural
Cross-session persistence
Zettelkasten-inspired (A-MEM)

shapes

HARNESS ENGINEERING

Skill Systems + Routing Ch05

700K+ community skills
Skill Graphs (arscontexta)
Progressive disclosure routing
85-95% token reduction

Agent Architecture Ch04

Planner / Generator / Evaluator
Sprint contracts for QA
6x performance gap (Meta-Harness)

Safety + Control Ch04

Guides (feedforward) vs Sensors (feedback)
Computational vs Inferential controls
Multi-layer content guards

MCP / Tool Layer Ch07

Universal protocol standard
Linux Foundation governance
Auth, sandboxing, audit trails

The model is the CPU. Context is RAM. The harness is the OS.
Each generation contains the last. Harness engineering includes context engineering, which includes prompt engineering.

awesome-llm-knowledge-systems -- The first unified guide connecting every piece of the AI knowledge ecosystem.

The Map Everyone's Missing