Neural Semantic Graph

The Neural Semantic Graph is K-AI's proprietary implementation of the Semantic Document Layer (Layer 3 of the 5-layer architecture). It is what makes K-AI different from a Data Catalog (metadata-oriented) and from a classical retrieval-augmented-generation (RAG) stack (vector-oriented).

The contrast above is the only place the term "RAG" is used across this documentation. K-AI is not a RAG product, and the rest of this site describes the platform on its own terms — through the Neural Semantic Graph, the role model, and the Govern · Clean · Activate disciplines.

Four capabilities

Unified semantic graph. The graph links documents, concepts, subjects, actors, and dependencies across the entire estate. Every document is positioned relative to every other document by explicit relations — not by vector proximity alone. The graph is the primary data structure of the platform: every other capability (audit, retrieval, observability) reads from it.

Contradiction detection. The graph surfaces inconsistencies between documents that do not reference each other directly. This is the case classical retrieval misses — vector search returns documents semantically close to a query, not documents that disagree with one another. K-AI Audit operates on top of this capability to produce the conflicts list a Steward acts on every day.

Missing-subject clustering. The graph identifies expected subjects that the estate does not yet cover. The signal comes from triangulation: emerging query patterns from Consumers, neighbouring subjects covered in the estate, and gaps in the topological coverage of the graph itself. This is the input to the missing-subjects workflow in K-AI Audit.

Contextual query understanding. Query understanding is grounded in the explicit relations of the graph rather than in vector proximity alone. A query about "the validation procedure for industrial sites" is resolved against the graph's representation of the procedure, the sites concept, and the validation dependency — not against a flat similarity score.

How it's exposed

The graph is queryable at two levels.

Direct read access lives at the Instance API — Semantic Graph tag (four operations). This surface is for machine-to-machine integrations that need raw access to the graph structure — typically internal pipelines, custom indexing layers, or Document Catalog publishers.

High-level semantic queries are exposed through K-AI MCP via the retrieval_semantic_nodes_search tool. This is the surface host LLMs use to ground their answers. The tool returns the semantic nodes that match a query, with lineage attached, and the host LLM orchestrates the rest of the answer client-side.

What it isn't

Not a Data Catalog. A Data Catalog inventories metadata — owners, schemas, lineage, tags. The Neural Semantic Graph encodes the semantic content of the documents themselves and the relations between them. A Catalog can sit on top of the graph (and in many deployments, it does); the graph cannot be reduced to a Catalog view.

Not a vector store alone. A vector store gives proximity. It does not give contradiction. Two documents can have similar embeddings and disagree completely; they can also have distant embeddings and be functionally identical. The Neural Semantic Graph encodes those relations explicitly, which is what makes contradiction detection and missing-subject clustering tractable.

Not an LLM. The graph is built before any inference happens at query time. LLMs participate in building the graph (parsing, entity extraction, concept linking) and in answering queries on top of it, but the graph itself is a persistent, deterministic data structure — not the transient output of a generative model.