A knowledge graph is a data structure that stores entities as nodes and their relationships as labeled edges, so connected facts can be queried directly. In commercial real estate it links tenants, leases, properties, and borrowers across documents, letting a system answer a question that spans many documents by traversing relationships rather than reading text.
How Does a Knowledge Graph Work?
A knowledge graph works by storing facts as triples, each a subject, a predicate, and an object, such as "Tenant ABC" "leases" "Suite 210." Entities become nodes and predicates become the labeled edges between them. To answer a question the system traverses edges, following relationships from node to node, instead of scanning unstructured text for a match.
The triple is the atomic unit, written as (subject, predicate, object) or (h, r, t), as knowledge-graph references describe it. Chaining triples builds a graph: if "ABC Coffee" "leases" "Suite 210" and "Suite 210" "is in" "Riverside Plaza," a query can hop from tenant to property in two steps. This structure captures relationships that flat records lose.
Component | Function |
Node (entity) | A tenant, property, lease, or borrower |
Edge (relationship) | A labeled link such as "leases" or "guarantees" |
Triple | Subject, predicate, object; the atomic fact |
Ontology | The schema of allowed entity and relationship types |
Traversal | Following edges to answer multi-hop questions |
Why a Knowledge Graph Matters
A knowledge graph matters because the hardest CRE questions span documents, and flat storage cannot follow the links. Questions like "which properties does this borrower guarantee across the portfolio" require hopping from borrower to loans to properties. A graph traverses those edges directly, while a keyword or vector search returns passages a person must still stitch together by hand.
The reasoning gain is measurable. In Microsoft's GraphRAG study, Edge et al. 2024, a knowledge-graph-backed system won LLM-judged comprehensiveness comparisons against vector-based retrieval-augmented generation roughly 72 to 83 percent of the time on broad sensemaking questions. The working principle is direct: when the answer lives in the relationships between documents, a graph beats a search that only ranks passages.
Example
A knowledge graph earns its place when one borrower connects to many assets. A lender asks, "What is our total exposure to guarantor Jane Roe across the loan book?" The answer requires linking a guarantor to loans to properties, a multi-hop path no single document holds.
Triple | Relationship |
Jane Roe guarantees Loan 1 | guarantor to loan |
Jane Roe guarantees Loan 2 | guarantor to loan |
Loan 1 is secured by Riverside Plaza | loan to property |
Loan 2 is secured by Oak Center | loan to property |
The graph traverses from the Jane Roe node across two "guarantees" edges to Loan 1 and Loan 2, then across two "secured by" edges to Riverside Plaza and Oak Center. If Loan 1 is 8,000,000 dollars and Loan 2 is 5,000,000 dollars, the traversal sums total guaranteed exposure to 13,000,000 dollars across two properties. A flat document search would return four separate files the analyst then has to reconcile by hand.
Variations and Edge Cases
A knowledge graph is a category with several designs, and the right one depends on schema strictness and query style. Some enforce a formal ontology, others grow loosely from extracted text. The variants below trade rigor against flexibility.
Variant | Treatment |
Property graph | Nodes and edges carry key-value properties; common in practice |
RDF triplestore | Standards-based triples queried with SPARQL |
Ontology-driven | A fixed schema constrains valid entities and relationships |
GraphRAG | A graph feeds an LLM for multi-hop, cross-document answers |
Extracted graph | Built automatically from documents; needs entity resolution |
Knowledge Graph vs Vector Database
A knowledge graph is often confused with a vector database, and the difference is what they store and search. A knowledge graph stores explicit entities and labeled relationships, and answers by traversing those links. A vector database stores numeric embeddings and answers by finding passages closest in meaning, with no explicit relationships between them.
A vector database answers "which clauses are about early termination," a similarity search over text. A knowledge graph answers "which tenants share a common guarantor," a traversal over relationships. Many CRE systems use both: the vector database finds relevant passages, and the knowledge graph connects the entities those passages name into a queryable structure.
Frequently Asked Questions
What is a knowledge graph in commercial real estate?A knowledge graph is a data structure that stores entities like tenants, leases, and properties as nodes and their relationships as labeled edges. In CRE it links parties and assets across documents so a system can answer questions that span a whole portfolio by traversing relationships rather than reading text.
What is the difference between a knowledge graph and a vector database?A knowledge graph stores explicit entities and labeled relationships and answers by traversing links. A vector database stores embeddings and answers by finding passages closest in meaning. A graph handles multi-hop relationship questions; a vector database handles meaning-based passage search.
What is a triple in a knowledge graph?A triple is the atomic fact in a knowledge graph, a subject, a predicate, and an object, such as "Tenant ABC leases Suite 210." Chaining triples builds the graph, and following the labeled edges between them lets a system answer questions that connect many entities.
Related Terms
Entity Resolution
Vector Database
Named Entity Recognition
Retrieval-Augmented Generation
Semantic Search