Knowledge Graph

General definition

A Knowledge Graph models a domain as a network of entities (people, products, organisations, drugs, documents) connected by named relationships (works at, treats, depends on, cites). Each fact is a triple: subject, predicate, object. Entities and edges can carry attributes, and a schema or ontology defines which types and relationships are allowed. Google popularised the term in 2012 for the entity panel beside search results; the underlying ideas come from the Semantic Web (RDF, OWL, SPARQL) and from property-graph databases such as Neo4j, which use query languages like Cypher.

• Multi-hop questions: “Which of our customers use a component supplied by a vendor under sanctions?” is a walk across three edges, awkward for keyword or similarity search
• Explicit relationships: the graph states that A treats B; it does not have to be inferred from text every time
• Consistency and provenance: one canonical node per entity, with sources attached, avoids the duplicates and contradictions found in raw documents
• Reasoning: rules and ontologies let the system infer new facts (if X is a subtype of Y, X inherits Y’s properties)

The natural comparison in AI systems is with a vector database. Vectors store embeddings of text chunks and return the chunks most similar to a query, which is excellent for “find me the passage about X” and poor at “how are X and Y connected”. A graph is the reverse. Building one is the expensive part: entities and relations must be extracted from documents, either by hand, by classic NLP pipelines or, increasingly, by having a language model read the corpus and propose triples that a reviewer checks.

GraphRAG combines the two. Documents are chunked and embedded as in ordinary retrieval-augmented generation, and an extracted graph of entities, relationships and community summaries sits alongside. A query can then pull both the closest passages and the relevant subgraph, which helps with questions that need the whole picture rather than one paragraph. It costs more to build and maintain than plain RAG, so it earns its place on corpora where relationships carry the meaning: regulatory frameworks, supply chains, clinical pathways, large codebases.

In the Ethora ecosystem

Retrieval in the Ethora AI SDK is RAG over a knowledge base built from a website crawl, uploaded files and per-agent live web indexing; the platform does not ship a graph database. Where a customer already runs a knowledge graph, the fit is through tool calling: an agent can query the graph as a tool, take the returned facts into its context and answer with them, with the graph remaining the system of record.

That division suits regulated deployments. The chat, the agent and the retrieval index run on the customer’s own infrastructure on a dedicated or self-hosted install, and the graph stays wherever the organisation already governs it. Answers that cite graph facts are recorded in the room transcript and the audit trail like any other message, so a compliance reviewer can see which relationships the agent relied on.