Wikidata
Wikidata stores knowledge as data rather than text. Every item, from a person to a chemical element to a film, has a stable identifier and a set of statements that link it to other items and to outside databases. Because the facts are structured, you can ask precise questions like "which cities are capitals of European countries" and get exact answers, which is harder to do against free-flowing article text.
You can download the full dumps in JSON and RDF, a data format for describing things as subject, predicate, and object triples. You can also query the live data through a SPARQL endpoint, a query language built for knowledge graphs, or read individual items through the API. The data is multilingual, so labels and descriptions come in hundreds of languages.
For RAG, Wikidata is most useful as a fact layer that sits alongside your text corpus: resolve an entity mentioned in a document to its Wikidata item, then pull reliable attributes and links. It also underpins many other projects, since it absorbed much of the old Freebase data and now acts as a shared hub of identifiers across the open data world.
Related sources
DBpedia
A knowledge graph built by pulling the structured parts of Wikipedia, mainly the infoboxes, into machine-readable data. It holds billions of facts about people, places, organisations, and more as RDF triples, small subject, predicate, object statements, which you can search with the SPARQL query language.
Freebase
A collaborative knowledge base once run by Google, now retired but still available as downloadable data dumps. It holds structured facts about millions of entities, and much of its content has since moved into Wikidata.
OpenCyc
The open release of Cyc, one of the oldest attempts to hand-build common-sense knowledge for machines. It holds hundreds of thousands of concepts and millions of assertions about how the everyday world fits together. Now archived, but the data is still available.
Wikipedia
Complete database dumps of Wikipedia and the other Wikimedia projects, in every language, refreshed roughly twice a month. The most widely used knowledge source for RAG systems, and an easy first corpus to build a retrieval pipeline on.