What is a knowledge graph?

What a knowledge graph is

Google's 2012 announcement, by Amit Singhal, explained the shift in three words: things, not strings. Instead of matching the letters of a search, Google would know that "Taj Mahal" is a monument, a musician, or a casino, and how each relates to other things. The launch graph held more than 500 million objects and more than 3.5 billion facts about them and their relationships.

The W3C's RDF standard describes the same idea in its simplest form: a graph is a set of triples, each a subject, a predicate and an object. "Loop Analytics — employs — Priya" is a triple. "Priya — maintains — CRM connector" is another. Join them and you can answer a question neither states alone: who at Loop Analytics maintains the CRM connector.

Two related words often come along. An ontology is the agreed list of kinds of things and relationships (customer, product, uses, maintains). Stanford's Ontology Development 101 calls an ontology plus its individual instances a knowledge base, which is close to what most people mean by a knowledge graph.

Why a small team should care

• Your knowledge is already a graph, scattered. Who owns which account, which customer uses which feature, which contract has which renewal date: it lives in a CRM, a drive and people's heads.
• AI can draft it. Microsoft's GraphRAG work describes LLM-generated knowledge graphs built straight from documents; you don't need to type every fact.
• It makes AI answers checkable. If every link keeps the quote it came from, a teammate can check the one link a decision rests on.

A worked example

Illustration: a made-up business. Harbor Hiring, a seven-person recruiting agency, has 60 client briefs and placement notes. The partners want to know which clients hire the same kinds of roles and which candidates they have already been shown. They ask an AI to extract entities (client, role, candidate, recruiter) and relationships (hired, interviewed, rejected, owns) from the notes, merging "J. Ortiz" and "Jamie Ortiz" into one person, with a quote and a confidence level for every link.

The table has 410 triples. Sorted by client and role, it shows that four clients each hired backend engineers twice in a year, and that one candidate was sent to two competing clients the same month. The second finding becomes a rule in the agency's process. The first becomes a sales list.

How to build a first one

1. Choose the question the graph must answer, and the 20 to 60 documents that hold the facts.
2. Name five to eight kinds of things and the relationships between them before extracting.
3. Extract triples with a source quote and a confidence level for each, and merge different spellings of the same name.
4. Check the low-confidence links and any link a decision depends on.
5. Draw it or sort it: a table sorted by subject answers most questions; a diagram shows clusters.

Common confusions

• "It needs a graph database." A first graph can be a spreadsheet of triples. Move to a database when the queries get long.
• "It's the same as a glossary." A glossary defines words. A graph records facts about specific things.
• "Once built, it stays true." Facts change. Keep a date on each link, and rebuild from the documents on a schedule. For using a graph to answer AI questions, see GraphRAG.

Questions

What is a knowledge graph in simple terms?

A map of facts: the things you care about (customers, products, people) as points, and how they relate (buys, uses, reports to) as labelled links. Each fact is one link, such as 'Acme uses the Salesforce connector'.

Who coined the term knowledge graph?

The idea is older, but Google made the term widely known in May 2012, when it launched the Knowledge Graph in search with the line 'things, not strings'. At launch it held more than 500 million objects and over 3.5 billion facts about them.

How is a knowledge graph different from a spreadsheet?

A spreadsheet holds one kind of thing per sheet. A knowledge graph holds many kinds of things and the links between them, so you can follow a chain (customer, uses connector, made by partner) without a lookup per sheet.