Forward-deployed / Learning zone
Knowledge graphsa standalone module
Lesson 01

What is a knowledge graph?

TL;DR

A knowledge graph stores knowledge as things and the relationships between them — (Acme Corp) —supplies→ (Widget X), (Widget X) —part of→ (Product Y) — instead of as rows in disconnected tables or words in disconnected documents. Each fact is a small, explicit statement: a triple of subject → relationship → object. Facts share entities, so they snap together into one connected web you can traverse. From a customer, you can hop to their contracts, to the products in them, to the supplier of a failing part — all in one query. That's the entire trick, and it's enough to change what your product can answer. Google named the strategic shift in 2012: things, not strings. Stop matching text; start knowing what the text refers to. The product test is equally plain. If your most valuable unanswered questions require joining knowledge across three or more hops — or the relationships themselves are the product — you're in knowledge-graph territory. If not, a well-modeled database is honestly fine.

🎯 For the product leader

Why it matters — You will hear "knowledge graph" from vendors, data teams, and AI strategy decks, often meaning wildly different things. If you can't state what one actually is — explicit entities and relationships, queryable by traversal — you can't tell a genuine capability from a rebranded database. You also can't price what building one would take.

What it changes in your decisions — You stop asking "should we have a knowledge graph?" That's an infrastructure question. You start asking "which questions can't we answer today because the knowledge is disconnected?" That's a product question. The questions justify the graph, never the other way around.

Ask yourself — "What are the three most valuable questions my product can't answer today because the answer lives in more than one system?"

Risk if ignored — You fund a two-year platform project with no killer query attached. Or you dismiss the idea entirely, and a competitor answers in one hop what takes your team a week of spreadsheet archaeology.

Things, not strings

Search for "jaguar" in a text index and you get documents containing the string. A knowledge graph instead knows there are three things — an animal, a car brand, a sports team — each with its own identity, properties, and relationships. Google's 2012 Knowledge Graph launch put a name on this shift. The phrase is still the cleanest summary of the whole field: a string is what something is called; a thing is what it is.

Your company runs on strings today. "Acme Corp" in the CRM, "ACME Inc." in billing, "acme-corp-2019" in the contract system — three strings, one thing, and no system that knows it. The knowledge graph move is to mint one identity for the thing, then attach every fact to it:

Things, not strings

One connected view · every arrow is a stored fact

A question unanswerable in any single source system answers in one traversal once the arrows exist.

"Which renewals are at risk because of Supplia's recall?"
Acme Corp
Customer
Contract #4411
renews Nov 2026
Product Y
Widget X
component
Supplia GmbH
Supplier
37 support tickets
(90 days)

One traversal: Supplia → widgets → products → contracts → customers. That's the entire trick — and it's enough to change what your product can answer.

Every arrow is a stored fact. The question on the left is unanswerable in any single source system. Once the arrows exist, it answers in milliseconds. That's the product argument for knowledge graphs, in one picture.

The anatomy: entities, relationships, triples

Strip away vendor vocabulary and there are only four parts:

Two properties of this shape do most of the work. Facts compose. Nobody wrote down "Supplia's recall threatens Acme's renewal" — it emerges from four independently recorded facts. A graph answers questions nobody anticipated when the data was entered. A table schema answers only the questions its designer anticipated. The schema bends. Adding a new relationship type ("is regulated by") doesn't require migrating tables — you just start drawing new arrows. That flexibility is a gift and a hazard, which is why the ontology lesson comes next.

What a knowledge graph is not

The fastest way to sharpen the concept is to name its neighbours:

It's not... Because... Where that shines instead
A relational database Tables can store relationships (join tables), but every hop is a join you designed in advance. Five-hop paths and "any path between A and B" questions get slow and hard to write Transactions, reporting, anything with a fixed, known access pattern
A vector store Embeddings capture similarity ("these two documents feel alike"). That's fuzzy, unexplainable, and unauditable. A graph captures stated facts with provenance Semantic search over text, RAG retrieval
A data warehouse / lake Warehouses centralize records for aggregation. They don't resolve identity across sources or make relationships first-class BI, metrics, large-scale batch analytics
An LLM A model's knowledge is latent — statistical, unattributable, frozen at training, and sometimes wrong with confidence. A graph's knowledge is explicit, current, and citable Language understanding, generation, extraction (lesson 6)
A taxonomy or org chart Trees allow one parent per node. Real knowledge is a web — a product belongs to a category, a supplier, and a compliance regime, all at once Simple classification, navigation menus

The honest summary: a knowledge graph is a complement to all of these, not a replacement for any of them. It typically sits across your systems of record as a connective layer. That's exactly why identity resolution (lesson 3) is the hard part.

The test: do you actually need one?

Three signals say yes. Their absence says no.

  1. Your valuable questions are multi-hop. "Which customers are exposed to this supplier?" "Which features does this regulation touch?" "Who in the org has worked with this account before?" If the answer requires chaining three or more relationships across systems, tables fight you and graphs don't.
  2. The relationships are the product. Recommendations ("people who bought X"), fraud detection (rings of accounts sharing devices), professional networks, supply-chain risk — in these products the edges are the inventory.
  3. Many teams keep re-deriving the same connections. If sales ops, risk, and support each maintain their own spreadsheet mapping customers to products to owners, the company is paying for the same graph three times, badly.

If none of these hold — your questions are one-hop, your joins are known in advance, one system of record covers the domain — a knowledge graph is a solution seeking a problem. The capstone lesson will tell you to spend the money elsewhere.

Failure modes

Practitioner checklist