What is Knowledge Graph?
Also called Semantic Network, Entity Graph.
A knowledge graph stores information as entities connected by typed relationships, such as a person working at a company that acquired another company. Unlike a document index, it makes connections explicit and queryable, which supports multi step questions that require following links between facts rather than matching text.
The building block is a triple: a subject, a relationship, and an object. Many triples form a graph in which each entity appears once and accumulates edges from every source that mentioned it. Querying traverses edges, so a question about which suppliers a company depends on indirectly becomes a path search rather than a keyword search.
Graphs complement vector retrieval instead of replacing it. Vector search excels at finding passages about a topic, while a graph excels at aggregation and multi hop reasoning, such as counting, comparing, or chaining relationships across documents. Systems combining both are commonly labeled graph retrieval augmented generation, and they route each question to the structure that suits it.
The main cost is construction. Extracting entities and relationships from unstructured text requires a model or hand written rules, and the output needs entity resolution so that repeated mentions of one thing collapse into one node. Errors compound, because a wrong edge silently produces wrong answers for every query that traverses it.
Schema choice is a persistent design tension. A tight ontology gives precise queries and rejects messy input, while a loose schema absorbs anything and yields ambiguous results. Practical systems usually fix a small set of entity and relationship types for the questions they must answer, and leave everything else in documents.
Key points
- Stores entities and typed relationships as a queryable graph.
- Strong at multi hop questions and aggregation.
- Requires extraction and entity resolution to build.
- Wrong edges silently corrupt every query that uses them.
- Often paired with vector retrieval rather than replacing it.
In practice
A graph built from company filings holds nodes for organizations, people, and products, joined by edges for employs, acquired, and supplies. Asking which vendors are owned by the same parent becomes a two hop traversal that returns three vendors, an answer no single document contains, because the ownership links appeared in three separate filings.