The wiki is a knowledge graph

graph
Every typed edge reifies to an RDF triple, so the wiki is a SPARQL-queryable graph, which buys queries prose and grep cannot answer.
Author

Chris Sweet

Published

May 14, 2026

What a knowledge graph is. A knowledge graph represents information as entities (the nodes) and the relationships between them (the edges), with enough formal meaning attached that software can query and reason over it, not just display it. You already rely on some: Google’s Knowledge Graph is the fact panel beside its search results, and Wikidata is the open one behind Wikipedia. Under the hood there are two common flavors, RDF triple-stores and labeled-property graphs; the atomic unit of the RDF kind is a triple (subject-predicate-object), and its query language is SPARQL. This wiki becomes an RDF knowledge graph almost for free.

Every typed edge in a page’s front matter reifies to (is recorded as) an RDF triple, a subject-predicate-object fact a database can query. So the wiki is not just interlinked markdown; it is a queryable knowledge graph, with a formal ontology (an explicit vocabulary of page and relationship types), SHACL rules that check the structure is valid, and a shipped library of SPARQL queries (SPARQL is the query language for graphs, like SQL for tables) an agent can run against it directly.

That layer buys queries prose and keyword search cannot answer. “Which page holds the wiki together” is a single query over the mentions edges:

SELECT ?page (COUNT(?src) AS ?inbound)
WHERE  { ?src :mentions ?page }
GROUP BY ?page
ORDER BY DESC(?inbound)
LIMIT 1

On a live research wiki that returned the page on Random-Walk-With-Restart (a graph-ranking method, the idea behind PageRank) with 48 inbound edges, the structural hub, which grep could never surface. Other questions have the same shape: walk a criticizes: edge transitively to find every downstream claim a retraction implicates, or lint at commit time for supports: claims with no ingested source. These are graph queries over meaning, not keyword queries.

The distinction matters because a browsable pile of markdown limits an agent to what it can grep. A reified graph lets it query relationships, transitive closures, and structural roles, and it ships with the template, so a project gets it without building graph infrastructure.

One honest caveat is worth keeping. The graph is only as valuable as its consumption, and today an LLM navigating a wiki reads by name-matching, not by walking the graph. So the graph is dark matter until a retrieval step reads it on the LLM’s behalf. Building the graph is necessary; wiring it into the read path is the open work.

Sources