Chapter 20 — Graph Engineering

TL;DR: Graphs structure knowledge so agents can navigate by relationships, not just text similarity. Less hallucination, more precision.

When to use graphs

• Entities with strong relationships (person → company → product)
• Hierarchies (domain → subdomain → feature)
• Causes and effects (bug X → cause Y → solution Z)

interface KnowledgeGraph {
  nodes: { id: string; type: string; properties: Record<string, any> }[];
  edges: { from: string; to: string; type: string }[];
}

async function queryGraph(query: string, graph: KnowledgeGraph) {
  // 1. Find relevant nodes
  const startNodes = await semanticSearch(query, graph.nodes);

  // 2. Expand through edges (relationships)
  const neighborhood = expandNeighborhood(startNodes, graph.edges, depth: 2);

  // 3. LLM reasons over the subgraph
  return await llm.generate({
    context: formatGraph(neighborhood),
    query
  });
}