Who popularised the term knowledge graph?

Google, on 16 May 2012, announcing its Knowledge Graph with the line "things, not strings". The underlying concept is older: it extends the semantic networks of the 1960s and the semantic web described by Tim Berners-Lee in 2001.

How is a knowledge graph different from a relational database?

In a relational database the relationship is encoded in the schema and needs a join to traverse. In a graph the relationship is a first-class object, typed and directly traversable. The difference becomes obvious as soon as you query chains several hops long.

Is the Obsidian graph view a knowledge graph?

Only partly. It is a document graph with untyped links: it tells you two notes are connected without saying how. A knowledge graph in the strict sense qualifies the relationship — "contradicts", "is an instance of", "cites" — which is what makes it queryable.

What is a knowledge graph?

A knowledge graph represents information as entities connected by named, typed relationships. Unlike a conventional database where links are implicit, a graph makes the relationship itself explicit and directly queryable.

How does it actually work?

A knowledge graph represents information as a network of entities connected by named relationships. Its atomic structure is the triple: a subject, a predicate, an object — "Niklas Luhmann", "taught at", "Bielefeld University". Any node can belong to several triples, which produces a network rather than a tree.

The payoff comes down to one thing: the relationship is explicit and typed. Saying that two documents are "related" tells you almost nothing; saying that one "refutes" the other lets you ask a question like "which of my sources contradict this claim" and get an answer.

Where does the concept come from?

The term took hold on 16 May 2012, when Google announced its Knowledge Graph with the line "things, not strings" — meaning the system understands that "Paris" names a city, a person or a film, rather than handling a character string. The graph has powered the information panels beside search results ever since.

The foundations are much older. Semantic networks appear in artificial intelligence in the 1960s with Ross Quillian's work. In 2001, Tim Berners-Lee, with James Hendler and Ora Lassila, set out the vision of the semantic web and the RDF/OWL/SPARQL stack, still the technical base of open graphs — Wikidata, launched in 2012 by the Wikimedia Foundation, is today its most consulted representative.

How does it differ from other models?

Model	Unit	Relationship	Strength
Relational database	Table row	Foreign key, via join	Integrity, aggregates
Document store	Document	Nested	Schema flexibility
Document graph	Note or page	Untyped link	Serendipity, browsing
Knowledge graph	Entity	Typed, first-class	Multi-hop querying

The line separating the last two is the one that matters for note-taking. The Obsidian or Logseq graph is a document graph: pleasant to look at, useful to browse, but unable to answer a question because it does not know what its edges mean.

What is a knowledge graph good for in practice?

Disambiguation. Attaching a mention to an identified entity resolves name clashes — Paris the city, Paris the character, Paris the Texas town.
Multi-hop questions. "Which authors cited in my PKM sources also work on attention" requires traversing two relationships; that is impossible without typing.
Spotting contradictions. Two sources asserting incompatible things about the same entity can only be compared if the entity is identified as such on both sides.
Giving a language model context. This is the use that revived the field from 2023 under the name GraphRAG: handing a model not isolated passages but the structured neighbourhood of an entity.

What are the limits?

Construction cost is the first obstacle. Extracting entities, connecting them, maintaining a coherent ontology: it is continuous work, and a poorly maintained graph produces wrong answers with exactly the same confidence as a correct one.

The second obstacle is subtler: a graph does not tell you what matters. It says two notions are connected without ranking them. On a personal graph of a few thousand nodes everything ends up connected to everything, and the visualisation becomes decorative string art — a well-known pattern among note-taking users, where the graph is admired at first and then never reopened.

The third is the entry point. A graph is traversed from somewhere. If the user cannot name the entity they are looking for — the usual case when the memory is vague — the graph stays out of reach despite everything in it.

How do the graph and search complement each other?

The graph solves connection; it does not solve entry. Search supplies the first node, and that is where most personal systems break down: you remember an idea, not the words you used to write it.

The numbers make the point. On a measured corpus of 948 syntheses embedded with bge-m3, the query "feudalism", backed by five genuinely relevant documents, peaks at 0.510 similarity, against 0.424 for a query matching nothing at all — nine hundredths apart. A fixed threshold at 0.65 returned nothing to a user who did in fact hold all five sources.

Synthiz combines the two. Concepts extracted from sources link to one another in a thematic "memory", with their sourced quotations attached; and the search that opens that memory merges semantic and lexical signals, judging relevance by how far a document sits from the distribution of its own query rather than by an arbitrary threshold. The web of links keeps the serendipity; search guarantees you can trace every idea back to its exact source.