A knowledge graph is a machine-readable map of your organisation’s entities (customers, products, contracts, machines) and the relationships between them. It stores facts as connections instead of rows. That structure lets people and AI systems answer questions spanning dozens of systems at once, with a traceable path back to every answer.
Knowledge graphs are not new: Google shipped one in 2012. Generative AI made them urgent, because a model that reasons well but cannot look anything up is only as good as the context you hand it. Gartner predicts 40% of enterprises will have leveraged GraphRAG techniques by 2029 to improve the factual accuracy of large language models.
What is a knowledge graph?
A knowledge graph stores information as a network of nodes and edges. A node is a thing: a customer, an invoice, a pump, a clinical trial. An edge is a labelled, directed relationship, for example Customer 4412 – signed – Contract 88. Both carry properties, so a machine can walk the model one relationship at a time.
Two ingredients separate a knowledge graph from any old graph. The first is an ontology: an explicit definition of which types of thing exist and which relationships are allowed. The second is identity: every real-world entity gets one stable identifier, so the customer in your CRM and the customer in your billing system are provably the same. Without both you have a network diagram, not knowledge.
Two standard families exist and both are mature. The W3C stack uses RDF for data, OWL for the ontology, SPARQL for queries and SHACL for validation, with the formal semantics that regulated or shared data needs. The labeled property graph model puts properties on nodes and edges, and became an ISO standard in April 2024 as GQL, the first new ISO database language since SQL.
Why are knowledge graphs important?
Knowledge graphs matter because they fix what stops most enterprise AI projects: the model has no reliable, connected view of the business. Gartner reported in February 2025 that through 2026 organisations will abandon 60% of AI projects not supported by AI-ready data.
The effect on accuracy is measurable. In a benchmark on a 199-table insurance data model, GPT-4 answered business questions correctly 16.7% of the time writing SQL straight against the schema, and 54.2% of the time when the same questions ran through a knowledge graph. On complex questions against a complex schema, direct SQL scored 0% and the graph scored 38.7%.
Production results follow. LinkedIn’s customer service team published a system that parses historical tickets into a knowledge graph before retrieval. It improved mean reciprocal rank by 77.6% and, after roughly six months in production, cut median per-issue resolution time by 28.6%.
The governance argument is separate but real. A graph answer carries its own path: which record, which source system, which relationship. That matters when an auditor or the EU AI Act asks why your system said what it said. Gartner expects universal semantic layers to be critical infrastructure by 2030.
How does a knowledge graph work?
A knowledge graph works by turning source records into resolved entities and typed relationships, validating them against an ontology, and serving the result as a queryable layer. Six steps.
- Model the domain. Pick the entity types and relationships you need. Ten to twenty is plenty at first.
- Ingest and map. Pull from CRM, ERP, ticketing and documents, mapping each field into the ontology.
- Resolve entities. Merge records that describe the same real-world thing under one identifier.
- Validate. Apply constraints, SHACL shapes or schema rules, so broken facts are rejected at the door.
- Serve. Expose the graph through a query language, and as tools an AI agent can call.
- Maintain. Source systems change. An unrefreshed graph becomes a confident liar within a quarter.
From scattered source systems to an answer that carries its own evidence trail.
Step three is where projects die. Entity resolution is quadratic by nature: every record must be compared with every other unless you add blocking and filtering. A 2023 survey of knowledge graph construction pipelines found it the least supported task in available tooling.
How AI reads the graph: GraphRAG
GraphRAG is retrieval-augmented generation where retrieval traverses a graph instead of, or alongside, a vector index. Microsoft’s 2024 study built an entity graph from a corpus, clustered it into communities and summarised each. Against a vector RAG baseline it won on comprehensiveness in 72% to 83% of comparisons and on diversity in 62% to 82%.
The cost profile surprises people. Answering from top-level community summaries used over 97% fewer context tokens than summarising the source text. Microsoft’s LazyGraphRAG goes further: indexing costs 0.1% of full GraphRAG, at roughly 4% of its query cost.
Example: how to use a knowledge graph
The clearest use case is a question your stack cannot answer because the answer lives in the joins. Take a manufacturer with a CRM, an ERP, a ticket system and twenty years of PDF manuals. The board asks: which customers run equipment containing a component that has failed twice in the field, and what is our contract exposure?
- Define eight entity types. Customer, Site, Machine, Component, Supplier, Ticket, Contract, Manual section. Enough to answer the question, small enough to finish.
- Load the spine. CRM and ERP supply customers, sites, machines, bills of materials and contracts, with their existing keys.
- Resolve identities. “Acme B.V.”, “ACME BV” and “Acme Group, Rotterdam” become one Customer node. Keep a human in the loop for ambiguous matches.
- Enrich from text. Run tickets and manuals through an LLM extraction step to attach failure modes to the right Component nodes. Check a sample by hand.
- Serve it to the assistant. Give the model graph queries as tools, so it retrieves facts instead of paraphrasing documents.
The board’s question becomes one traversal: component, to machines containing it, to sites, to customers, to active contracts. The answer arrives with the ticket IDs and contract numbers behind it.
The pattern is documented at scale. AstraZeneca’s Biological Insights Knowledge Graph holds 10.9 million nodes and over 118 million edges; trained on pre-2015 data, its pipeline recovered 16 of the 95 relevant new drug targets found between 2015 and 2020. NASA’s chief knowledge architect described the Orion programme finding an Apollo-era lesson in fifty years of records, saving two years of work and a million dollars.
What does a knowledge graph cost?
Hosting a knowledge graph is cheap. Building one is not. Platform pricing is published and predictable, while ontology design, integration and entity resolution absorb most of a first-year budget.
| Option | Entry price | What you get | Notes |
|---|---|---|---|
| Neo4j AuraDB Free | $0 | 200,000 nodes, 400,000 relationships | Deleted after 30 days without activity; no SLA |
| Neo4j Community, self-hosted | $0 licence | Full single-instance database | GPL v3; you pay for your own infrastructure and operations |
| Neo4j AuraDB Professional | From $65 per GB per month | A 1 GB instance is $65.70 per month | Single zone, no uptime SLA, billed even while paused |
| Neo4j AuraDB Business Critical | From $146 per GB per month | 2 GB minimum, so $292 per month | Multi-zone, 99.95% SLA, 24/7 support |
| Amazon Neptune Serverless (Ireland) | $0.1774 per NCU-hour | The 1 NCU floor running continuously is about $130 per month | Plus $0.10 per GB-month storage and $0.22 per million I/O requests |
| TigerGraph Savanna | $45 per GB per month | Free trial workspace available | Dublin and Frankfurt carry a 1.125x regional factor |
| Memgraph, Stardog, Ontotext GraphDB | Free edition, paid tiers on request | Varies | None of the three publishes list prices for paid tiers |
Now the number that decides your budget. For a realistic first production build, assume one ontologist, two data engineers and part-time domain experts for six months. At European consultancy rates of €800 to €1,200 per day that lands between roughly €250,000 and €450,000. Against that, a full year of an 8 to 16 GB managed instance at $525 to $1,051 per month costs $6,300 to $12,600: low single digits as a share of the build. That ratio, not the licence, is what you plan around. Our own model from the rate cards above, not an analyst statistic.
For the upside case there is one substantial study, and it is vendor-commissioned. A Forrester Total Economic Impact analysis for Neo4j from October 2021 modelled a composite organisation at 417% ROI and $4.18 million NPV over three years, against $1.0 million in present-value licence and administration costs. Read it as directional: its cost model covers only licences and half an FTE of administration.
You can also start at zero: the free Aura tier holds 200,000 nodes and 400,000 relationships, enough to prove or kill an idea in a fortnight.
Alternatives and comparison
A knowledge graph is one of five reasonable answers to “our AI cannot find anything”, and most organisations end up combining two of them.
| Approach | Best at | Falls down on | Entry cost |
|---|---|---|---|
| Knowledge graph | Questions that span records, governance, answers with provenance | Needs an ontology and entity resolution; months to build | $0 self-hosted, from $65 per GB per month managed |
| Vector database and RAG | Finding the passage that answers a question; fast to stand up | Multi-hop reasoning, counting, aggregation | Free tiers; Pinecone from $20 per month, Weaviate from $45 per month |
| Relational database or warehouse | Aggregation, reporting and transactions on a known schema | Open questions across many tables; GPT-4 scored 16.7% on a 199-table schema | Your existing licence |
| Data catalogue or semantic layer | Shared definitions, lineage, access governance | Reasons over metadata, not over the data itself | Often bundled with your data platform |
| Full-text and document search | Finding a document you already know exists | Answering anything that spans two documents | Low, usually already in place |
The honest comparison is not graph versus vectors, it is which questions you ask. Benchmarks are blunt. In the HippoRAG 2 evaluation at ICML 2025, a strong vector retriever averaged 57.0 F1 across seven question-answering datasets while Microsoft’s community GraphRAG averaged 49.6, worse than plain vector RAG on straightforward factual questions. The 2026 GraphRAG-Bench study agrees: basic RAG matches or beats GraphRAG on simple fact retrieval, and GraphRAG pulls ahead on complex reasoning and summarisation across a corpus. Graph indexing is also 40 to 57 times slower to build.
So the rule is simple. Use vectors for “find me the passage”. Use a graph for “work out the answer across records”. If most of your traffic is the first kind, buy a vector database and stop reading. If your expensive questions are the second kind, and wrong answers carry regulatory or financial weight, the graph earns its build cost. On mixed traffic, a hybrid beats either alone.
Conclusion
A knowledge graph is not a database upgrade. It is a decision to write down what your organisation knows, in a form a machine can traverse and an auditor can follow. The technology is stable and ISO-standardised, the platform cost is a rounding error, and the failure mode is entity resolution, not query performance. Start with one question your stack cannot answer, and prove the value before you buy the platform.
Wondering whether your data is ready for this? An AI assessment maps what you have and what is missing, and AI consultancy covers the ontology and integration work.
Frequently asked questions (FAQ)
What is the difference between a knowledge graph and a graph database?
A graph database is storage technology. A knowledge graph is what you put in it: an ontology defining which types of thing exist, plus resolved identities so each real-world entity appears exactly once.
Do I still need a knowledge graph if I already have RAG?
Only if your questions span records. Vector search is strong at finding the passage that answers a question, and weak at joining facts, counting or following a chain of relationships. Benchmarks show it beating graph approaches on simple fact retrieval and losing on multi-hop reasoning.
How long does it take to build a knowledge graph?
A proof of value on one question takes two to six weeks. A production graph across three or four source systems typically takes four to nine months, most of it spent on entity resolution and data quality.
What is GraphRAG?
GraphRAG is retrieval-augmented generation where retrieval traverses a knowledge graph instead of, or alongside, a vector index. Microsoft’s 2024 study reported comprehensiveness win rates of 72% to 83% against a vector RAG baseline. Gartner predicts 40% of enterprises will use it by 2029.
Should I choose RDF or a property graph?
Choose RDF if you need formal semantics, reasoning, shared vocabularies or exchange between organisations. Choose a labeled property graph if you mainly need fast traversal with rich attributes. The property graph model became an ISO standard in April 2024 as GQL.
What is the hardest part of a knowledge graph project?
Entity resolution: proving that the customer in your CRM, your ERP and your ticket system is one customer. It is quadratic by nature, needs human judgement, and a 2023 survey found it the least supported step in the tooling.
