What Is a Knowledge Graph

Google Knowledge Graph - 500 billion facts. Wikidata - 100M+ entities. Neo4j in NASA for mission management. Microsoft Azure KG powering Copilot enterprise RAG. LLM + KG = neuro-symbolic AI: llamaindex KnowledgeGraphIndex makes GPT-4 reason over structured facts instead of hallucinating. This is not research - it's production 2024.

• **Google Knowledge Panel**: direct answers from 500B facts - no page click required
• **Microsoft Copilot enterprise**: Azure Knowledge Graph as structured memory for RAG, accuracy +40% vs plain embeddings
• **NASA missions**: Neo4j component graph - instant failure cascade analysis for any spacecraft part
• **Drug Discovery**: BioKG / OpenTargets link genes, proteins, diseases; $2B+ saved by filtering out dead-end molecules early
• **llamaindex KnowledgeGraphIndex**: LLM reads the graph, not raw text - 3-5x fewer factual hallucinations

Nodes (Entities)

Google Knowledge Graph stores 500 billion facts about 8 billion entities. When the search engine returns a "Christopher Nolan - director of Inception" card without requiring a page click, it's not a fancy database table - it's a **knowledge graph**: a network of connected entities. Microsoft Copilot enterprise RAG and llamaindex KnowledgeGraphIndex (neuro-symbolic AI) run on the same principle. Each entity is a **node** (vertex) in the graph.

Google Knowledge Graph contains more than **800 billion** facts about 8 billion entities. Wikidata - the largest open knowledge graph - holds more than 100 million entities, each with a unique identifier (for example, Q42 = Douglas Adams).

Entities in a Knowledge Graph (KG) are not just text strings. Each entity has a **unique identifier** (URI), a **type** (class), and **properties**. "Moscow" in a KG is a specific entity distinct from "Moscow Mule" (cocktail) or "Moscow" (film). Wikidata holds 100M+ such unambiguous entities - which is precisely why it's valuable for training data annotation in LLMs.

Entities are grouped into **classes** (types): Person, Place, Organization, Event. This creates a hierarchy: Scientist is a subclass of Person, University is a subclass of Organization. This typing enables meaningful queries: "All scientists born in Germany".

Edges (Relationships)

Entities without relationships are just a list. The power of a Knowledge Graph lies in its **edges** - typed relationships between entities. Every edge has a **direction** and a **type**: "Einstein → born_in → Ulm" is not the same as "Ulm → born_in → Einstein".

Relationship types in a KG are standardized through **ontologies** (schemas). Wikidata defines thousands of relationship types (properties): P19 (place of birth), P69 (educated at), P106 (occupation). Standardization lets different systems exchange knowledge without ambiguity.

Edges in a KG can carry **qualifiers** - metadata on the edges themselves. For example: "Einstein → citizenship → Germany [start_time: 1879, end_time: 1896]". This is impossible in a simple graph - it requires a **property graph** or **RDF reification**.

Relationships create **multi-hop paths**: "Einstein → worked_at → Princeton → located_in → New Jersey → part_of → USA". These paths enable answers to complex questions - "In which country is the university where Einstein worked?" - without storing that fact directly. NASA uses Neo4j on this same principle: a graph of spacecraft component relationships lets engineers instantly trace the failure cascade of any single part.

Triples

Every fact in a Knowledge Graph is recorded as a **triple**: (subject, predicate, object). This is the atomic unit of knowledge. The entire graph is a collection of such triples. "Moscow - capital - Russia" = one triple. "Earth - orbits - Sun" = another.

The object can be one of two things: an **entity** (another node in the graph) or a **literal** (a value: string, number, date). "Einstein → born_in → Ulm" - object is an entity. "Einstein → birth_date → 1879-03-14" - object is a literal. Literals are the leaves of the graph.

Wikidata contains more than **1.7 billion** triples. DBpedia (extracted from Wikipedia) holds around 3 billion. YAGO - 120 million. Triples are a universal format: any fact in the world can be expressed as (subject, predicate, object).

The power of triples lies in **composability**. Complex questions decompose into chains: "In which country was the creator of the theory of relativity born?" = (?, created, Relativity) → (Einstein, born_in, ?) → (Ulm, country, ?) → Germany. Each step is one triple.

RDF

Triples are a great model, but a standard format is needed. **RDF** (Resource Description Framework) is the W3C standard for representing triples. In RDF, everything is identified by a **URI** (Uniform Resource Identifier), which allows data from different sources to be merged without naming conflicts.

URIs in RDF solve the **disambiguation** problem: http://wikidata.org/entity/Q937 refers to a specific Einstein, not a namesake. URIs guarantee that two different sources referring to Q937 are talking about the same person.

**Wikidata**, **DBpedia**, and **Schema.org** are the largest open KGs built on RDF. Google's Knowledge Panel (the info card in search results) is powered by a proprietary KG enriched with Schema.org markup from web pages. JSON-LD is the format through which websites pass structured data to search engines.

Key Takeaways

• **Nodes** - entities with a unique URI, type, and properties; 100M+ in Wikidata, 500B+ facts in Google KG
• **Edges** - typed, directed relationships; multi-hop paths answer questions without storing every fact
• **Triples** (subject, predicate, object) - the atomic unit of knowledge; Wikidata has 1.7B+ triples
• **RDF** - W3C standard with global URIs; JSON-LD delivers KG data from millions of websites to search engines
• **LLM + KG** - neuro-symbolic AI: llamaindex KnowledgeGraphIndex cuts hallucinations 3-5x on factual queries

Related Topics

Knowledge Graphs combine ideas from databases, the semantic web, and AI:

• RDF and the Semantic Web — The next lesson covers RDF formats, OWL ontologies, and the SPARQL query language in depth
• Property Graphs: Neo4j — An alternative graph model with native properties on edges

Вопросы для размышления

• Pick any public scientist and write five triples about them in (subject, predicate, object) format. Which predicates had to be invented because they don't exist in Wikidata?
• Why can't Google simply store all the world's facts in a single SQL table?
• What is the advantage of URIs over plain strings when merging data from Wikipedia, IMDb, and academic papers?

Связанные уроки

• db-01-intro — Relational and graph models are two alternative ways to store connected facts
• nlp-01 — Knowledge graphs are used in NLP for entity linking, fact-checking, and knowledge-augmented LLMs
• ds-01-arrays — Graph data structures (adjacency list, matrix) are the in-memory representation of a knowledge graph
• st-01-feedback-loops — Ontologies and systems models both describe networks of interacting concepts
• dm-01