Knowledge Graph

A knowledge graph stores facts as triples: subject, predicate, object. "Acme Corp employs Alice." "Alice manages the API team." "The API team owns the billing service." The triples form a graph that can be queried with structured queries (Cypher, SPARQL) or traversed by an agent to answer multi-hop questions. Knowledge graphs are stronger than vector RAG for questions that follow relationship paths ("who manages the team that owns the service that handles refunds?") and weaker for fuzzy semantic search. The two are often combined: vector RAG for unstructured retrieval, graph queries for relationship traversal.

Production knowledge graphs in agent systems usually take one of three shapes. First, hand-curated (highest quality, expensive to maintain): a domain expert builds the graph manually. Second, LLM-extracted (lower quality, scalable): run an LLM over your documents to extract triples automatically, accept some noise. Third, hybrid: extract candidates with LLM, validate with human review, store the validated subset. Neo4j and Memgraph dominate the property-graph market in 2026; many teams run knowledge graphs on PostgreSQL with the Apache AGE extension to avoid managing a second database.

Add a knowledge graph when your queries follow relationship paths and answer quality matters more than retrieval recall. Skip for unstructured Q&A workloads.

• Neo4j: GraphRAG docs
• LangChain: Knowledge graph integration

Related Terms