ConceptNet

ConceptNet joins words and short phrases with edges that carry a meaning, for example "is a", "used for", "causes", or "part of". Instead of just knowing that two words appear near each other, you get an explicit statement of how they relate, which is the kind of common-sense link people use without thinking. The graph is large and multilingual, built from crowdsourcing, expert-made resources such as WordNet and Wiktionary, and games designed to gather knowledge as people play.

You can work with it three ways. Download the full graph as a database dump for offline use, query the live REST API at api.conceptnet.io when you want to look up a single concept's neighbours, or use ConceptNet Numberbatch, the associated word embeddings (numeric representations of meaning that place related words close together). Numberbatch is often the most practical piece for retrieval, because it blends ConceptNet's relationships with distributional embeddings and tends to beat plain word vectors on word-similarity tasks.

In a RAG pipeline ConceptNet earns its place at query time rather than in the index. Use it to expand a user's query with related terms and synonyms so retrieval catches documents that phrase things differently, or to add a layer of everyday common sense (that a kettle is used for boiling water, say) that a purely text-based system tends to miss. It also helps normalise and connect entity mentions across languages.

Be realistic about what it is. ConceptNet captures loose everyday associations, not precise, verifiable facts, and because much of it is crowdsourced the edges can be noisy, uneven, or occasionally wrong, with coverage thinning outside the best-served languages. Do not treat it as a source of truth. On licensing, CC BY-SA 4.0 permits commercial use with attribution, but the share-alike term means a database you build on top of it and redistribute must carry the same licence.

For precise lexical relations in English, WordNet is a tighter companion, and for hard facts about named entities, Wikidata is the better tool. ConceptNet fills the common-sense gap between them.