Machine-readable knowledge graphs that encode entities, relationships, and facts in structured formats like RDF and JSON-LD.
4sources
Knowledge graphs store facts as a web of entities and the relationships between them, rather than as paragraphs of text. Think “Marie Curie, won, Nobel Prize in Physics, 1903” expressed in a machine-readable form. For RAG, they solve a problem that plain text handles badly: precise, unambiguous lookups. When a user asks for an exact date, a capital city, or every drug that treats a condition, a graph can return the fact directly instead of hoping a relevant passage was retrieved.
Choosing within this category is mostly about how you query. Many graphs are published as RDF, a standard way of writing facts as subject-predicate-object triples, and queried with SPARQL, a language built for searching that structure. That is powerful but has a real learning curve. Others expose a friendlier API or downloadable tables. The other axis is breadth versus depth: some graphs cover everything shallowly, while domain graphs go deep on one field like biology or geography. General-purpose graphs are a good default; reach for a specialist one when accuracy in a single domain matters most.
The main things to watch are completeness and freshness. Graphs are often crowd-built, so coverage is uneven and some facts are missing or out of date. Structured data can also be surprisingly hard to feed to a language model, which reads prose more naturally than triples, so many teams convert facts into short sentences before retrieval.
The sources below range from broad, general-purpose graphs to focused domain ontologies, so you can match structure and coverage to your needs.
A multilingual common-sense knowledge graph that links words and phrases with labelled connections, such as "a cat is a pet" or "rain causes wet". It captures the everyday relationships between ideas that plain text rarely spells out.
An API into Google's Knowledge Graph, the store of billions of facts about entities that powers the info panels you see beside search results. You send a name or query and get back matching entities with descriptions, types, and links. Free to use within rate limits.
A machine learning system from Carnegie Mellon that has been reading the web since 2010 and building a knowledge base as it goes. It extracts beliefs, entities, and the relationships between them from text, and keeps refining them over time.
A lexical database of English that groups nouns, verbs, adjectives, and adverbs into sets of synonyms called synsets, then links those sets by meaning. It maps how words relate, which sense means what, what is a kind of what, so software can work with meaning rather than just spelling.