Search engines can tell you that Kathmandu is the capital of Nepal, that Nepal borders India and China, and that Mount Everest lies on the border. Behind such answers there is often a knowledge graph: a structured network of facts in which things are connected by meaningful relationships. Knowledge graphs are increasingly combined with language models to make AI answers more accurate and explainable.
What is a knowledge graph?
A knowledge graph represents information as:
- Entities (nodes): things such as people, places, organisations, products, concepts.
- Relationships (edges): how entities are connected, such as "is located in", "works for", "was born in", "is part of".
- Properties (attributes): details like name, date, population, price.
Facts are commonly stored as triples: subject, predicate, object.
- (Kathmandu, capital_of, Nepal)
- (Nepal, borders, India)
- (Bibek Dhakal, founded, Wake and Tech Nepal)
Linking many triples creates a graph that can be explored and queried: "Which companies were founded by people born in Nepal?"
Why use graphs instead of tables or text?
- Relationships are first-class: questions that follow connections across many hops are natural in graphs but clumsy in tables with many joins.
- Flexible schema: new types of facts can be added without redesigning tables.
- Disambiguation: each entity has a unique identity, so "Mercury" the planet differs from "Mercury" the element.
- Reasoning: rules and ontologies allow inferring new facts, for example if X is part of Y and Y is part of Z, then X is part of Z.
- Explainability: you can show the path of facts that led to an answer.
Building blocks
Ontology / schema
A formal description of entity types and allowed relations, such as Person, Organisation, Place, and relations like employedBy. Standards include RDF, RDFS and OWL in the semantic web world.
Graph models
- RDF graphs use triples with URIs and are queried with SPARQL.
- Property graphs (used by Neo4j and similar) put properties on nodes and edges and are queried with languages like Cypher.
Graph databases
Systems optimised for storing and traversing connections, such as Neo4j, Amazon Neptune, TigerGraph, and RDF stores like Apache Jena or GraphDB.
Data sources
Public knowledge graphs such as Wikidata and DBpedia, enterprise databases, documents, spreadsheets and APIs. Many are created by extracting entities and relations from text with NLP.
A small example in Python
The networkx library lets you model and query small graphs:
import networkx as nx
G = nx.MultiDiGraph()
triples = [
("Kathmandu", "capital_of", "Nepal"),
("Pokhara", "located_in", "Nepal"),
("Nepal", "borders", "India"),
("Nepal", "borders", "China"),
("Mount Everest", "located_in", "Nepal"),
("Mount Everest", "located_in", "China"),
("Wake and Tech Nepal", "headquartered_in", "Nepal"),
]
for s, p, o in triples:
G.add_edge(s, o, relation=p)
# Which things are located in Nepal?
print([s for s, o, d in G.edges(data=True) if o == "Nepal" and d["relation"] == "located_in"])
# Is there a path from Pokhara to India?
print(nx.shortest_path(G.to_undirected(), "Pokhara", "India"))Production systems use databases and query languages rather than in-memory structures, but the thinking is the same.
A SPARQL example
Wikidata can be queried with SPARQL. A query to list some countries bordering Nepal looks like this:
SELECT ?country ?countryLabel WHERE {
wd:Q837 wdt:P47 ?country . # Nepal (Q837) shares border with ...
SERVICE wikibase:label { bd:serviceParam wikibase:language "en". }
}Identifiers like Q837 are Wikidata IDs, which you can look up on the site.
Knowledge graphs and language models
Language models are fluent but can hallucinate; knowledge graphs are precise but rigid. Combining them uses the strengths of each.
Grounding answers
Retrieve facts from a graph and supply them in the prompt, instructing the model to answer from them. The graph provides verified, up-to-date facts, and the model provides natural language.
GraphRAG
Instead of retrieving only similar text chunks, systems build a graph of entities and relations from documents, then retrieve connected subgraphs or community summaries. This helps with questions that need linking information across many documents, such as "How are these suppliers related to these incidents?"
Building graphs with LLMs
Language models can extract entities and relations from text, proposing triples that are then validated and merged. Quality control matters: extraction errors, duplicate entities with different names ("Wake & Tech" versus "Wake and Tech Nepal") and conflicting facts must be resolved by entity resolution and human review.
Natural-language querying
A model can translate a question into a graph query (Cypher or SPARQL), run it, and explain the results. Validate generated queries and restrict them to read-only access.
Use cases
- Search and recommendations: connecting products, categories and customers.
- Fraud detection: finding rings of accounts sharing devices or addresses.
- Healthcare and life sciences: linking drugs, genes, diseases and studies.
- Enterprise knowledge management: connecting people, projects, documents and systems.
- Supply chains: tracing dependencies between suppliers, parts and sites.
- Education: mapping concepts and prerequisites in curricula.
- Chatbots and assistants: more accurate, explainable answers about a domain.
Challenges
- Cost of building and maintaining a high-quality graph; knowledge changes and must be updated.
- Entity resolution: deciding when two records refer to the same thing.
- Schema design: choosing the right level of detail.
- Data quality and provenance: tracking where each fact came from and how reliable it is.
- Scalability for very large graphs.
- Skills: graph modelling and query languages are specialised.
- Not always necessary: if your data fits neatly in tables and questions are simple, a relational database may be enough.
When to consider a knowledge graph
Consider one when your questions involve many relationships, your data comes from diverse sources that must be integrated, you need explainable answers, or you want to ground a language model in verified domain facts. Otherwise start simpler.
Frequently Asked Questions
Is a knowledge graph the same as a graph database?
A graph database is a technology for storing graphs. A knowledge graph is a graph of meaningful facts, which may be stored in a graph database or elsewhere.
What is the difference between a knowledge graph and a vector database?
A vector database finds similar items by meaning using embeddings. A knowledge graph stores explicit facts and relations. They complement each other.
Can a language model build a knowledge graph automatically?
It can help extract candidate facts, but results need validation to be reliable.
Where can I get a free knowledge graph to practise?
Public sources such as Wikidata and DBpedia provide large open graphs.
Conclusion
Knowledge graphs store facts as connected entities and relations, enabling precise, explainable reasoning. Combined with language models, they ground AI answers in verified information. Use them when relationships and trustworthiness matter most.