The ontology, BrainDup's frame of truth
The AI proposes, the ontology and the code decide.
A map of facts, not plausible text
BrainDup turns an archive of documents (scanned PDFs, web pages, reports, email) into a queryable knowledge graph. Instead of an assistant guessing at plausible text, it builds a map of the facts actually present in your documents, each one tied to its exact source.
The ontology is the plan of that map, chosen before the first line is drawn.
Before a single document is read, BrainDup asks a simple question: what is the vocabulary of your domain? That vocabulary, the types of entity and the relationships that may exist between them, is written in a file in the standard format of the semantic web: the same OWL that has structured knowledge graphs for twenty years. We call that file a pack.
The mechanism, in three steps
1. The pack is chosen
BrainDup ships ready-made packs for several sectors (cultural heritage, property, healthcare, law, events), adoptable as they are or customisable. You do not start from a blank page.
2. The pack is locked for the whole corpus
This is not a configuration detail: it is the boundary between what the AI may assert and what it may not. The vocabulary stops being negotiable as the documents arrive.
3. The code decides, never the model
On reading a document, the language model proposes entities and relationships. Deterministic code decides whether each proposal enters the graph: it is accepted if it matches a type declared in the pack and attaches to a precise passage of the source document. Otherwise it is discarded.
That mechanism is what makes BrainDup's central promise tenable: every assertion points to a verifiable source, and an answer whose source cannot be established is never padded out with invention; it becomes an honest refusal. The same principle governs the citations assembled by code and underpins our answer to the AI Act.
What the ontology changes in practice
| Without an ontology | With a BrainDup pack | |
|---|---|---|
| Vocabulary | Decided along the way by the model | Declared before ingestion, identical for the whole corpus |
| Unknown entity | Created silently, duplicates pile up | Discarded if it matches no declared type |
| Relationships | Inferred from the proximity of words | Typed; causation requires an explicit statement or human validation |
| Homonyms | “Mercury” merges planet, metal and person | Each sense is a distinct entity, identified without ambiguity |
| Traceability | The source is reconstructed after the fact | A fact with no attachable passage does not enter the graph |
| Consistency | Invisible until somebody rereads | Contradictions are caught against the pack's rules |
Put another way: the ontology does not make the model cleverer, it shrinks the surface on which it can be wrong.
What an ontology precisely is
The word here does not refer to the branch of philosophy, but to something very concrete: the formal, explicit and shared specification of a conceptualisation. A logical dictionary paired with a construction plan, for a given domain. It does four jobs.
It defines the concepts
In a law firm: “Case”, “Party”, “Statute”. In a heritage archive: “Place”, “Person”, “Craft”, “Event”. Each class carries its own characteristics.
It describes typed relationships
Where a hierarchical classification would only say “flu belongs to viral diseases”, the ontology expresses treats, causes, contraindicated with, administered by.
It removes ambiguity
Mercury is a planet, a metal, a Roman god and a singer. A human decides in a fraction of a second; a machine can only do so if each sense is declared as a distinct entity.
It serves as a meeting point
Two databases designed separately, with different field names, align on the same ontology. Each keeps its internal structure, both become readable by the same programs.
A proven inheritance, not a house invention
The idea comes from Tim Berners-Lee. The inventor of the web set out, as early as the 1990s, the ambition of a web whose data machines could understand, and published its best-known formulation in 2001 in Scientific American, with James Hendler and Ora Lassila. His goal fits in one sentence: move the web from a network of documents to a network of knowledge, what he would call in 2007 the Giant Global Graph.
Because a hyperlink expresses no meaning: it says “look over there”, never “this is that person's employer”. The building blocks below were designed to fill that gap. They are W3C standards, in production for more than ten years. BrainDup does not invent a proprietary formalism: it applies that one to your archive.
Give every thing a unique address
The classic web gives pages an identifier. The semantic web gives them to concepts too: a person, a medicine, a municipality. Two systems using the same URI are, with certainty, talking about the same thing.
Describe facts in three words
All information is expressed as a triple: subject, predicate, object. Marie Curie / received / the Nobel Prize in Physics. The subject of one triple becomes the object of another, and the knowledge links itself together.
Write the dictionary and the rules
RDF says what is; OWL says what can be. Classes nested within one another, mandatory or unique relationships, reciprocal links, equivalences between identifiers of different origins.
Ask the graph questions
The equivalent of SQL for the semantic web. “Women composers born in Europe before 1900 with a work performed in Vienna”: no page holds that list, the graph assembles it from scattered facts.
Inference: when the machine deduces without guessing
OWL rests on description logics, rigorous enough to be computed automatically. An inference engine applies the ontology's rules to the data and draws conclusions that were never written down.
The ontology declares that “is parent of” is the inverse of “is child of”, and that the parent of a parent is a grandparent. The database holds two facts: Anne is parent of Julie, Julie is parent of Leo. The engine works out on its own that Anne is Leo's grandmother. Nobody entered it, and yet nothing was invented: the deduction is demonstrable.
Reasoning also acts as a guard rail. If the ontology states that a person has only one date of birth and the database holds two, the engine flags the inconsistency. The system detects its own errors, a property no amount of plausible text provides.
Where these technologies already run
Search engines
The summary card beside your results comes from a knowledge graph. The schema.org vocabulary, launched jointly by the major engines in 2011, lets sites make themselves understood that way.
Healthcare and life sciences
The major medical terminologies are ontologies. They guarantee that a diagnosis entered in a Finnish hospital means the same thing in a Portuguese study.
Research and open data
The FAIR principles, which structure the publication of scientific data, rest on these standards to make datasets findable and reusable by programs.
Heritage and libraries
Catalogues, archives and museums publish their collections as linked data: you move from an author to their works, then to the associated places and periods. That is the ground our heritage packs cover.
What ontologies cost
Honesty requires saying it: the vision of a worldwide semantic web, set out in 2001, did not materialise at that scale. Building an ontology is expensive, demands domain experts, and presumes a consensus that takes long to reach. The semantic web therefore came about in islands: very solid where rigour is vital (healthcare, finance, industry, law, research), looser elsewhere.
That is precisely what the packs address: rather than a blank page on every project, an existing sector vocabulary is adopted, refined and reused. And the ambition changes scale: not a global graph, but a private, sovereign and verifiable graph, built for your organisation.
What is the vocabulary of your domain?
That is the first question we ask, and the answer determines everything else. Thirty minutes is enough to know whether an existing pack covers your archive or needs extending.