Image by Steve Buissinne from Pixabay

This glossary translates the vocabulary on graphrag.info into plain language, but also pairs each term with a rough Java developer equivalent.

How to read this glossary

Each entry has three parts: what the term means in plain words, a rough developer equivalent, and where that equivalent breaks. Treat the equivalents as handles, not definitions. A developer who stops at the analogy will build the wrong thing.

One idea runs under every entry. Most enterprise data tells a machine how data is stored. Semantic data also tells a machine what the data means, in a form software can check. Meaning is essential to sharing and reusability.

A terminology note: My blogsite graphrag.info uses semantic graph and knowledge graph almost interchangeably, and so does this glossary, except where an entry says otherwise.

TermPlain EnglishDeveloper equivalentWhere the analogy breaks
FAIR dataData prepared so people and software can find it, get it, combine it with other data, and reuse it. The letters FAIR stand for Findable, Accessible, Interoperable, Reusable. The 2016 paper that coined the term FAIR stresses machines, not just human readers.A well-run internal API: stable IDs, documented schema, versioned, discoverable without emailing the owner.FAIR is a set of principles, not a file format or a test suite. Accessible does not mean open: access can require a login.
MetadataData about data. A label on a jar, not the jam.Column names and types, headers, docstrings, a README.Semantic metadata also records meaning and relationships, not just format and size. It’s reusable. Plain labels just aren’t.
Persistent identifier (PID)A permanent name for one thing that never changes or gets reused.A UUID or DOI. A permalink you promise never to break.A good PID also resolves: following it returns a description of the thing.

Internationalized Resource Identifiers (IRIs) are globally unique and multilingual.
Knowledge graphFacts stored as things and the links between them, plus a model of what those things mean.A database where relationships (action verbs) are first-class data, like a join table promoted to the main event, with a published type system on top.Not every graph is a knowledge graph, and not every knowledge graph reasons. Many only store and link.
Semantic graphA knowledge graph built on open standards, so any compliant tool can read its meaning. graphrag.info uses this term interchangeably with knowledge graph.A graph database whose schema follows an open spec instead of a vendor format.The openness is the point. A graph with a proprietary schema layer is a graph, but not a semantic graph in this sense.
OntologyThe agreed definitions of the concepts in a domain and the rules linking them. Example: a customer is a person or company, an order has exactly one buyer, a buyer is a customer.A domain model or type system, written so machines and every application share it.An ontology can infer new facts from its rules. A Java class hierarchy cannot. It also assumes that a missing fact is unknown, not false.
TaxonomyThe approved vocabulary: preferred names, synonyms, and broader and narrower terms.An enum, a lookup table, a tag registry.A taxonomy handles synonyms and hierarchy (poodle is narrower than dog). A plain enum does not.
Data graphThe actual facts: this customer, that order, this document.The rows. The instance data, as opposed to the schema.Documents and annotations live in knowledge graphs too, not just records.
TripleOne fact as subject, predicate, object. Example: Acme, supplies, Part 42.A row in an edge table: source, relation, target. A key-value pair with a labeled link.A triple is self-describing. Each of its three parts has a global identifier. With disambiguation, action verbs and global IDs, facts from different systems can merge.
Data siloData trapped inside one application.A microservice database nobody else can query.Silos hide meaning as well as data, because each app encodes its own definitions.

Core idea

TermPlain EnglishDeveloper equivalentWhere the analogy breaks
FAIR dataData prepared so people and software can find it, get it, combine it with other data, and reuse it. The letters FAIR stand for Findable, Accessible, Interoperable, Reusable. The 2016 paper that coined the term FAIR stresses machines, not just human readers.A well-run internal API: stable IDs, documented schema, versioned, discoverable without emailing the owner.FAIR is a set of principles, not a file format or a test suite. Accessible does not mean open: access can require a login.
MetadataData about data. A label on a jar, not the jam.Column names and types, headers, docstrings, a README.Semantic metadata also records meaning and relationships, not just format and size. It’s reusable. Plain labels just aren’t.
Persistent identifier (PID)A permanent name for one thing that never changes or gets reused.A UUID or DOI. A permalink you promise never to break.A good PID also resolves: following it returns a description of the thing.

Internationalized Resource Identifiers (IRIs) are globally unique and multilingual.
Knowledge graphFacts stored as things and the links between them, plus a model of what those things mean.A database where relationships (action verbs) are first-class data, like a join table promoted to the main event, with a published type system on top.Not every graph is a knowledge graph, and not every knowledge graph reasons. Many only store and link.
Semantic graphA knowledge graph built on open standards, so any compliant tool can read its meaning. graphrag.info uses this term interchangeably with knowledge graph.A graph database whose schema follows an open spec instead of a vendor format.The openness is the point. A graph with a proprietary schema layer is a graph, but not a semantic graph in this sense.
OntologyThe agreed definitions of the concepts in a domain and the rules linking them. Example: a customer is a person or company, an order has exactly one buyer, a buyer is a customer.A domain model or type system, written so machines and every application share it.An ontology can infer new facts from its rules. A Java class hierarchy cannot. It also assumes that a missing fact is unknown, not false.
TaxonomyThe approved vocabulary: preferred names, synonyms, and broader and narrower terms.An enum, a lookup table, a tag registry.A taxonomy handles synonyms and hierarchy (poodle is narrower than dog). A plain enum does not.
Data graphThe actual facts: this customer, that order, this document.The rows. The instance data, as opposed to the schema.Documents and annotations live in knowledge graphs too, not just records.
TripleOne fact as subject, predicate, object. Example: Acme, supplies, Part 42.A row in an edge table: source, relation, target. A key-value pair with a labeled link.A triple is self-describing. Each of its three parts has a global identifier. With disambiguation, action verbs and global IDs, facts from different systems can merge.
Data siloData trapped inside one application.A microservice database nobody else can query.Silos hide meaning as well as data, because each app encodes its own definitions.

The standards stack

graphrag.info names five open standards from the W3C, the body behind web standards such as HTML. Each does one job. Together they form a layered stack, like HTTP, TLS and JSON do for web APIs.

TermPlain EnglishDeveloper equivalentWhere the analogy breaks
W3C open standardsPublic specifications any vendor can implement.SQL or POSIX versus a vendor-only API.Open spec does not guarantee good tooling. It guarantees you can leave. graphrag.info calls this the defense against vendor lock-in.
RDF (Resource Description Framework)The data model. Every fact is a triple, and every thing has a global identifier.A universal edge list. Think of every JSON field turned into a labeled link.RDF is a model, not a file format. It has several serializations, such as Turtle and JSON-LD.
RDFS and OWL (Web Ontology Language)Vocabulary for defining classes and relationships. OWL adds logic, so a system can derive facts nobody entered.Schema plus type hierarchy plus a rules engine that runs over your data.OWL reasons under an open-world assumption: a missing fact means unknown. SQL assumes missing means false.
SHACL (Shapes and Constraints Language)Rules a graph must satisfy. Example: every order links to a valid customer. The site compares it to a smart contract for data.JSON Schema, java bean validation, or unit tests for data.SHACL checks what is in the graph. OWL infers what follows from it. Developers often confuse the two.
SPARQLThe query language for RDF graphs.SQL for graphs.You describe a pattern of linked facts instead of writing joins. Cypher is the equivalent for property graphs, which use a different data model.
SKOS (Simple Knowledge Organization System)A standard for publishing taxonomies: preferred terms, synonyms, broader and narrower terms.A schema for tag hierarchies and controlled vocabularies.Built for vocabularies, not for modeling business rules. That is the ontology’s job.
YAML-LDOrdinary YAML plus a small mapping, called a context, that turns the file into RDF.Annotating a config file or data class so a tool can map it automatically, like serializer annotations in Java.It removes the syntax barrier. It does not do the modeling work for you.
RDFization or triplificationConverting existing data into subject-verb-object triples.An ETL job or an object-relational mapping step.The hard part is agreeing on meaning, not moving bytes.

How AI uses semantic graphs

The site’s central claim: an LLM guesses fluently, and a graph answers precisely, so pair them. The language model handles the conversation. The graph supplies the facts.

TermPlain EnglishDeveloper equivalentWhere the analogy breaks
Probabilistic AILearns patterns from data and predicts likely answers. It cannot explain how it got one.A trained model: a function you call but cannot read.Temperature zero makes output repeatable. It does not make it explainable.
Deterministic or symbolic AIFollows explicit facts and rules, so every answer traces back to them.A rules engine, a pure function, a Prolog or Datalog program.Rules need people to write and maintain them.
Neurosymbolic or hybrid AIA language model and a symbolic reasoner working together.An LLM that writes the query and a database that executes it.The two halves fail differently. The seam between them is where errors hide.
Vector RAGRetrieval-augmented generation that finds text by similarity, then pastes it into the prompt.Nearest-neighbor search over embeddings, then prompt stuffing.Embeddings compress text into approximations. They can place synonyms and antonyms close together.
GraphRAGThe chatbot retrieves answers from a graph database instead of guessing.An LLM tool call that hits a typed, queryable store instead of a text index.Quality depends on the graph. A sloppy graph feeds the model sloppy facts.
HallucinationA confident answer with no basis in fact.Unvalidated output returned as if it passed checks.Fluent text looks right, which is why it slips through.
GroundingTying an answer to facts a system can verify.Validating against a source of truth and returning the citation.Grounding fails if the source itself is wrong or stale.
Reasoning or inferenceDeriving a new fact from existing ones. Example: this part contains an allergen, and this product uses the part, so this product contains the allergen. (AAAAI: “An allergen is a usually harmless substance (such as pollen) capable of triggering a response that starts in the immune system and results in an allergic reaction.”)A rules engine or query planner deducing a result.graphrag.info argues chain-of-thought prompting walks through steps but does not deduce.
ProvenanceWhere a fact came from and how it changed.Git history, an audit log, data lineage.In a graph, provenance travels with the fact itself.
AgentSoftware that acts on its own to pursue a goal, using an LLM to decide.A service with an LLM in its control loop and permission to call tools.Agents act, so mistakes cost more than a wrong chat answer.
MCP (Model Context Protocol)A general tool-use protocol that lets an agent call outs
ide tools.
An RPC or plugin interface for agents.MCP is not a query language. A separate step still turns the question into SPARQL.
Front-door problemTurning a natural-language question into a formal graph query is not deterministic.The parser at the edge of your system.A misparsed query returns an answer that is authorized, auditable, and wrong.

Enterprise architecture terms

These terms describe where the pieces sit in a production system.

TermPlain EnglishDeveloper equivalentWhere the analogy breaks
Digital twin (of a business)A machine-readable model of how a company operates: customers, products, contracts, and how they relate.A canonical domain model that every service and agent reads from.Here it means a data model, not a 3D simulation. The site lists sensor feeds and geometry as optional extras.
Semantic backboneThe shared RDF foundation that gives every system the same definitions.An N-dimensional, universal data model plus a service bus, for meaning instead of messages.Human first create the backbone. People with machine assist encode the knowledge and maintain it. It does not emerge from the data.
Semantic layerThe layer that models business concepts and constraints independent of where the data is stored.An abstraction layer like an ORM or a metrics layer, with logic attached.Unlike a typical metrics layer, it can validate and infer.
Knowledge engineeringThe human work of building and maintaining the model.Schema design and domain-driven design.It includes agreeing on definitions across departments, which is a negotiation, not a coding task.
Knowledge APIA gateway that takes an agent’s question, turns it into a graph query, filters by what the user may see, and returns a bounded answer with its proof.An API gateway plus a policy-enforcing data-access layer.It filters what the model may know before it reasons, not after it answers.
Control planeThe part of a system that decides what is allowed. The site argues a knowledge graph can serve as one for agents.A Kubernetes-style control plane or a policy engine.Standard agent harnesses watch execution downstream. The semantics community defines and governs knowledge upstream.
Entitlements and RBACRules for who may see or do what. RBAC means role-based access control.IAM roles and row-level security.Same idea. The difference is that the rules sit in the graph as checkable facts.
Proof chainThe trail of facts and rules behind an answer.A query plan plus data lineage.It is only as good as the rules recorded in the graph.
Context graphProcedural and situational knowledge layered over a knowledge graph’s static facts. The term comes from Graphwise executive Andreas Blumauer, as the site reports.Runtime policy and workflow state on top of a database.Ordinarily a vendor-associated term, not a settled standard. Graphwise defines and conforms “context graph” to its standards here.
AI-ready and audit-ready dataThe site argues one structure serves both: self-describing, machine-readable data lets software scan every record and lets auditors trace each fact to its source.Data with schema validation, lineage, and tests, queryable end to end.This is the site’s argument, not a certification. Good structure helps, but content still has to be right.
Data sovereigntyKeeping your data, knowledge, and rules under your own control.Self-hosting your data store instead of pasting data into a third-party prompt.The site argues a graph queried through a scoped interface lets only narrow answers cross the boundary.
Library and information science (LIS)The 150-year-old discipline of classification, metadata, and controlled vocabularies.Information architecture and data governance.GraphRAG.info claims this field supplies the missing discipline for agents.

One example in five views

The example follows the hidden-allergen scenario the site uses: a buyer cannot see an allergen several tiers upstream, so the system infers it.

A Java developer would store the fact like this:

record Contains(String parent, String child) {}
var f = new Contains(“Product12”, “Part42”);

The same fact in RDF, written in Turtle. Each line is one triple. The ex: prefix is a shorthand for a namespace, so every name is globally unique:

@prefix ex:  <https://example.com/&gt; .
@prefix owl: <http://www.w3.org/2002/07/owl#&gt; .

ex:Product12 ex:contains ex:Part42 .
ex:Part42    ex:contains ex:Peanut .
ex:Peanut    a ex:Allergen .

One OWL line gives the graph its inference rule. Declaring contains transitive means that if A contains B and B contains C, the graph treats A as containing C:

ex:contains a owl:TransitiveProperty .

A SHACL shape validates the data. This one requires every part to name at least one supplier:

@prefix sh: <http://www.w3.org/ns/shacl#&gt; .

ex:PartShape a sh:NodeShape ;
sh:targetClass ex:Part ;
sh:property [ sh:path ex:suppliedBy ; sh:minCount 1 ] .

A SPARQL query asks which products contain an allergen. With the transitive rule applied, Product12 matches even though no one entered that fact:

PREFIX ex: <https://example.com/&gt;
SELECT ?product ?allergen WHERE {
?product  ex:contains ?allergen .
?allergen a ex:Allergen .
}

The takeaway for developers: the Java record stores a fact. The graph stores the fact, its definitions, its validation rules, and the logic that derives new facts, all in one open, queryable form. The query returns the answer to the AI agent in a form it can verify, instead of a guess.

For more Information:

The definitions above follow how graphrag.info uses each term, checked against the sources below. Graphwise.ai sponsors graphrag.info.

Leave a Reply

Trending