A knowledge graph is a structured representation of information in which real-world entities, such as people, places, organizations, or concepts, are stored as nodes and the relationships between them are stored as labeled edges, forming a network that can be queried, traversed, and reasoned over. Knowledge graphs sit within the broader tradition of Symbolic AI representation and have become central infrastructure for search, question answering, and grounding modern generative AI systems.
Structure
A knowledge graph typically encodes facts as triples of subject, predicate, and object, for example "Anthropic, founded in, 2021", often following formal schemas or ontologies that define allowed entity types and relationships. This structure allows systems to perform logical queries, such as finding all organizations founded in a given year, that would be difficult to extract reliably from unstructured text alone, and to check internal consistency against defined rules.
History
The conceptual roots of knowledge graphs lie in semantic networks developed within symbolic AI research from the 1960s onward, and in later efforts such as Cyc, a project begun in 1984 to hand-encode millions of common-sense facts, and the Semantic Web initiative championed by Tim Berners-Lee beginning in the early 2000s, which proposed standardized formats for machine-readable web data. The term "knowledge graph" itself entered mainstream use after Google launched its Knowledge Graph product in 2012, drawing on the acquired Freebase database and other sources to power the information panels that accompany many web search results.
Google Knowledge Graph and successors
Following Google's adoption of the term, major technology companies built their own large-scale knowledge graphs, including Microsoft's Satori and Bing entity graph, Facebook's Entity Graph, and Amazon's product graph, each tailored to power search, recommendation, or virtual assistant products. Open, community-maintained alternatives include Wikidata, launched in 2012 as a structured companion to Wikipedia, which has become a widely used free knowledge graph resource for research and applications.
Use in AI systems
Knowledge graphs have taken on renewed importance in the era of Large language model systems, which are prone to Hallucination (AI), confidently generating plausible but false statements. Pairing a language model with a knowledge graph provides a mechanism for Grounding (AI) outputs in verified structured facts, either by retrieving relevant subgraphs to include in a prompt, a variant of Retrieval-augmented generation, or by validating generated claims against the graph after the fact. Knowledge graphs are also used in Semantic search systems to disambiguate queries and entities, complementing dense Embedding-based retrieval methods stored in a Vector database, and in enterprise settings to give AI systems access to proprietary, structured organizational knowledge that was never part of a model's training data.
Limitations
Building and maintaining a knowledge graph is labor-intensive: extracting accurate structured facts from unstructured sources requires substantial human curation or error-prone automated extraction, schemas can become outdated as domains evolve, and coverage is inevitably incomplete outside well-documented domains. These limitations have motivated hybrid approaches that combine the precision and auditability of symbolic knowledge graphs with the flexibility and broad coverage of neural methods, echoing a longstanding tension in AI between explicit, structured representations descended from Expert systems and statistical pattern learning from Natural language processing and Machine learning more broadly.