Key takeaways
- A knowledge graph stores facts as entity, relationship, entity: Jane, reports to, the CFO.
- A schema or ontology on top defines which entities and relationships are valid.
- Semantic network is the older name for a graph of concepts; knowledge graphs carry the idea at larger scale.
- Graphs suit questions about connections, tables suit records and vector search suits similar meaning.
- Account mapping is a natural graph problem, even when it lives in a spreadsheet.
How a knowledge graph works
A knowledge graph works by storing each fact as a small statement, two entities joined by a named relationship, and letting statements that share an entity connect into a network. 'Company A, acquired, Company B' and 'Company B, uses, Product X' together let a program answer a question neither fact answers alone.
On top of the stored facts sit two kinds of technique, which research surveys of the field describe as deductive and inductive. Deductive methods apply rules and ontologies to derive new facts: if A owns B and B owns C, then A indirectly owns C. Inductive methods learn patterns from the graph, for example to suggest links that are probably missing.
Semantic network: the older name
A semantic network is the older name for a graph of concepts joined by labelled links such as 'is a' and 'part of', and it is the direct ancestor of the knowledge graph. Early AI researchers used semantic networks to model how ideas relate in memory and to let programs answer simple questions by following links.
Knowledge graph is the term that took over as the same idea was applied to large collections of real-world facts, with standard formats, query languages and schemas so many people and programs can share one graph.
Knowledge graph vs database and vector database
A knowledge graph differs from a relational database by treating relationships as first-class facts, and from a vector database by storing exact, named links rather than closeness in meaning. The three are complementary, and many systems use more than one.
Knowledge graph examples in go-to-market
Account mapping is the clearest knowledge graph example in go-to-market work. The champion reports to the economic buyer, the parent company owns three subsidiaries, and a former customer's buyer now works at the prospect: each is an entity, relationship, entity fact.
Held as a graph, those facts answer questions a list cannot, such as which open deals include someone who used to work at a customer, or which subsidiaries sit under a parent you already sell to. The example is illustrative; most teams keep this knowledge in notes and heads, which is why it is lost when a rep leaves.
Knowledge graphs and large language models
Knowledge graphs and large language models are increasingly used together, with the graph supplying explicit facts and the model supplying fluent language. In graph-based retrieval, a system fetches the part of the graph relevant to a question and gives it to the model as context, so answers about relationships rest on stated facts rather than on what the model happens to recall.
The limits are practical. A graph takes effort to design, fill and keep current; facts extracted automatically from text carry errors; and a graph cannot answer anything its schema did not anticipate. It is worth building where the questions are about connections and the cost of a wrong answer is high.
Knowledge graph, relational database and vector database
How the three ways of storing knowledge compare, in general terms.
| Stores | Good at | Less suited to | |
|---|---|---|---|
| Knowledge graph | Entities and named relationships | Following chains of connections and inferring links | Bulk records with few relationships |
| Relational database | Rows in tables, linked by keys | Counting, filtering and reporting on records | Deep, many-step relationship questions |
| Vector database | Embeddings: numbers that capture meaning | Finding text or items similar in meaning | Exact facts and named relationships |
A general comparison of approaches, not of any named product.
Where Pebbles Ai fits
Common questions
Is a CRM a knowledge graph?
Not usually. A CRM stores records with some links between them, such as contacts on accounts, but it rarely treats relationships as typed facts a program can reason over and chain together.
Do you need a graph database to build a knowledge graph?
No. A graph database makes relationship queries faster and easier to write, but the model of entities and named links can sit in other stores too. The modelling matters more than the engine.
What is a triple in a knowledge graph?
A triple is one fact in subject, relationship, object form, such as Company A, acquired, Company B. A knowledge graph is a large set of triples that share entities.
About this page. Written by the Pebbles Ai team for people who run sales, marketing and growth. Last updated .
