Natural language queries hide technical complexity beneath apparent simplicity. A question may contain multiple entities, implicit relationships, temporal constraints, and domain-specific knowledge that users express naturally but which require structured processing.

Text retrieval-based systems would search for documents mentioning the relevant terms. They would likely find scattered pages about the concepts involved. But connecting these pieces of information to answer the specific question requires structured reasoning, not word matching.

Intent type detection determines whether the user is looking to locate a specific entity, describe a concept, or list examples. This initial classification guides all subsequent processing and defines the expected format of the response.

Element classification identifies each component of the query and its nature: what represents the type of entity being sought, what constitutes a temporal constraint, what expresses a semantic relationship. The parser separates nouns (entities), adjectives (properties), and verbs (relationships), mapping them to ontological categories.

Natural language is transformed into structured logic. The generated SPARQL queries encode precisely the semantics of the original question, but in a form that the triple store can process deterministically.

This translation is not literal but semantic. Each linguistic element is mapped to formal constructs: temporal constraints become date filters with precise boundaries, abstract concepts are translated into typed relationships in the ontology, and multiple conditions are combined through explicit logical operators.

The system queries verifiable knowledge structures, not documents. This difference represents the fundamental leap between neurosymbolic architectures and systems based on textual statistics. We do not search document indices that mention relevant concepts; we query knowledge formalised as a network of explicit relationships.

Each fact is represented as a triple: subject-predicate-object. The graph consists of RDF triples where subjects and objects are resources (nodes) connected through predicates (properties) that define the exact nature of each relationship.

Nodes represent entities with unique identifiers. Edges represent typed relationships connecting those entities. Each node has properties (intrinsic attributes) and relationships (connections to other nodes). This structure enables precise queries: either the triple exists in the graph, or it does not. There is no probabilistic interpretation.

A traditional RAG system would retrieve documents by semantic similarity and rely on the LLM to infer which information actually answers the question, with a high probability of error or hallucination. Lokut queries directly over formal relationships. The response is deterministic.

The engine is not limited to literal matching. It applies ontological inferences defined in the graph schema. If the ontology establishes that one class is a subclass of another, the reasoner automatically infers membership of the parent class even if that specific triple does not exist explicitly.

Transitive, symmetric, or inverse relationships are applied during query execution. The system completes implicit information without adding external knowledge, simply by applying the logical rules encoded in the ontology.

In this phase, generative agents do not generate knowledge — they generate humanised language. They summarise and structure the knowledge already retrieved from the graph. Their function is organisational: they connect the relationships found, adjust the linguistic register, add natural connectives, and structure information according to principles of discursive coherence.

The agents work exclusively with the propositions derived from the graph. They add no external information. They do not speculate. If the graph does not contain a piece of data, the response does not include it. This architectural constraint eliminates hallucinations by design.