Structured and semi-structured knowledge sources covering broad, general-purpose factual knowledge. The backbone of most RAG systems.
8sources
Encyclopaedic sources are the backbone of most RAG systems, and for good reason. They give broad, general-purpose factual coverage in prose that is already written to be read, which is exactly what you want when a language model needs a reliable passage to ground its answer. If you are building a general assistant and can only add one knowledge base, this is usually the category to reach for first.
The main choice here is between free text and structured data. Encyclopaedia articles give you clean, self-contained prose that chunks well into passages and drops straight into a vector database, the store that lets you search text by meaning rather than exact words. Structured knowledge bases instead give you facts as entities and relationships, which are precise and easy to filter but read awkwardly when handed to a model as-is. Many strong systems use both: prose for explanation, structured facts for exact lookups like dates, populations, or identifiers.
Two things deserve care. First, freshness: some dumps are snapshots taken months ago, so check the date if your users ask about current events. Second, licensing looks simple but is not always. Much of this content is CC BY-SA, meaning you must credit the source and share any adaptations under the same terms, which can matter for a commercial product. Structured extracts sometimes carry different terms again.
The sources below range from full encyclopaedias to the structured knowledge bases built alongside them, so you can pick the shape that fits your pipeline.
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.
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.
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 rendered as pre-parsed, machine-readable JSON: abstracts, short descriptions, infoboxes, sections, parsed tables and references, with links to Wikidata entities. A beta from Wikimedia Enterprise covering nine languages, also mirrored on Hugging Face. The section-segmented shape a RAG pipeline actually wants.
A fully open synthetic corpus of amplified multilingual encyclopaedic text with built-in reasoning traces and exercises covering RAG, information extraction and QA. Released by PleIAs with the AI Alliance, it targets training and evaluating small, grounded, citeable reasoning models.
A free, collaborative knowledge base that holds structured facts for Wikipedia and the other Wikimedia projects. More than 100M items, each one machine-readable with statements, relationships, and identifiers, so you can pull clean facts instead of parsing article prose. Released into the public domain under CC0.
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.
A semantic knowledge base that combines facts from Wikipedia, WordNet, and GeoNames into a single, high-accuracy collection of statements about entities. It adds when and where each fact holds, so you get temporal and spatial detail alongside the plain relationships.