Knowledge Graphs: Answering the Questions Vector Search Can't

A knowledge graph (KG) stores real facts as a network: things are nodes, and relationships are the lines between them. Each fact is a triple: (subject, relationship, object).

If the ontology is the house blueprint, the KG is the actual house with people living in it. It works like a family tree: "Who is my uncle?" is easy to answer by following the lines.

Knowledge graphs you use every day

• Search "Sundar Pichai" on Google. The info box (CEO of → Google, born in → Madurai) comes from Google's Knowledge Graph.
• LinkedIn's "People you may know" follows links like you → worked at → TCS ← works at ← friend.

Facts in our HR assistant

(Rahul,     worksIn,    Sales)
(Rahul,     reportsTo,  Priya)
(Full-time, governedBy, Leave Policy)

Now ask: "Which policies apply to Anita, and who is her department head?" The graph simply follows the links:

• Anita → Intern → Intern Policy
• Anita → Engineering → Arjun

GraphRAG: using both

Use a vector database for fuzzy text questions ("what's the rule on remote work?") and a knowledge graph for exact relationship questions ("who approves Rahul's leave?"). Combining both is called GraphRAG.

I use a vector DB for fuzzy text questions and a knowledge graph for exact relationship questions like 'who approves Rahul's leave'. Combining both is called GraphRAG.

- The short version

Next in the GenAI series: SHACL: Validation Rules That Keep LLM-Extracted Data Honest

Start from the top: GenAI Quick Answers and Series Guide