DBpedia

DBpedia turns the semi-structured parts of Wikipedia into a proper knowledge graph. An extraction pipeline reads infoboxes, categories, article links, and other markup, then publishes the facts as RDF triples: small statements of the form subject, predicate, object, such as "Berlin, capital of, Germany". Across all languages this adds up to billions of facts about people, places, organisations, works, and events, each tied to a stable identifier you can link to and from.

There are two main ways to work with it. For exploration and small live lookups, DBpedia runs a public SPARQL endpoint, where SPARQL is the query language designed for searching knowledge graphs, so you can ask for, say, every city in a country with its population in one query. For anything at scale you will want the bulk dumps (downloadable per language and per dataset), which you load into your own triple store or graph database so you are not depending on, or overloading, the shared public endpoint.

For RAG, DBpedia is most useful as a layer of structured facts and entity relationships sitting beside your text. You can use it to resolve which "Mercury" a query means, pull a tidy set of attributes for an entity to ground an answer, or expand a question with related entities before you retrieve. Because it mirrors Wikipedia's structure across many languages, coverage is broad and general rather than deep in any one field.

Two caveats. The data is only as consistent as Wikipedia's infoboxes, so expect gaps, the occasional stale value, and some messy typing. On licensing, DBpedia is CC BY-SA 3.0: commercial use is fine, but it is share-alike, meaning a derivative dataset you distribute must carry the same licence, and you must attribute the source.

DBpedia predates Wikidata and overlaps with it heavily. Wikidata is now the more actively curated of the two, but DBpedia's linked-data identifiers and language coverage mean the two are still often used together.