Knowledge graph
Overview
A knowledge graph (KG) is a graph-structured knowledge base in which nodes represent entities (people, places, concepts, products) and edges represent typed semantic relationships between them (is-a, part-of, founded-by, located-in). Knowledge graphs encode symbolic, explicitly curated knowledge as opposed to the implicit, distributional knowledge stored in language model weights.
Large-scale public knowledge graphs include DBpedia, Freebase (now inactive), Wikidata, and Google's Knowledge Graph (which powers the Knowledge Panel in search results). Organizational knowledge graphs are also built for enterprise use to represent internal entities and their relationships.
In the LLM era, knowledge graphs are used in combination with language models in two primary patterns: as structured retrieval targets in KG-augmented RAG systems, and as sources of structured ground truth for hallucination checking and faithfulness evaluation.
Knowledge graph vs. LLM parametric knowledge
| Dimension | Knowledge graph | LLM parametric knowledge |
|---|---|---|
| Representation | Explicit graph (nodes, typed edges) | Implicit in model weights (distributed vectors) |
| Update mechanism | Add/edit nodes and triples | Requires retraining or fine-tuning |
| Auditability | Fully inspectable and traceable | Opaque; cannot enumerate stored facts |
| Coverage | Curated; may miss long-tail entities | Broad; reflects training corpus diversity |
| Factual reliability | High for curated triples | Variable; subject to hallucination |
| Reasoning | Formal query (SPARQL); path traversal | Approximate; pattern-completion based |
Neither representation is strictly superior: KGs offer precision and traceability; LLMs offer fluency, generalization, and coverage of long-tail knowledge.
Role in GEO and brand entities
Google's Knowledge Graph directly influences Knowledge Panels in search and may influence entity recognition in generative AI systems that use structured data. A brand or person with a verified Knowledge Graph entry has a more stable and correctly attributed entity profile in Google's ecosystem than one represented only through text.
Schema.org markup on web pages encodes entity attributes in a KG-compatible format, providing a crawlable signal that feeds into both KG construction and LLM training data.
See also
- Entity authority / salience
- Brand entity in LLMs
- Retrieval-augmented generation
- Embeddings
- Fundamentals
