Ontologies, Knowledge Graphs, and AI — Live Demo (Mysore, 2026)
vishal-mysore builds GraphBaby — a browser-based demo using WebLLM (an in-browser LLM) — to clarify three concepts the field tends to conflate:
Ontology: a formal, explicit specification of the concepts in a domain and the rules governing how they relate. The schema, not the data.
Knowledge graph: a populated dataset of concrete entities and their factual connections, as subject-predicate-object triples. The data, built on top of a schema.
AI’s role: LLMs now automate both — drafting the ontology schema from a prose description, then populating the knowledge graph from raw text. What previously required a knowledge-engineering specialist collapses into a two-pass LLM pipeline.
The key finding: schema as guardrail
Constraining the LLM’s second pass (graph population) with the ontology schema from its first pass (schema drafting) reduces hallucination. The schema limits what assertions the model can make — only entities and relationships the ontology permits can appear in the extracted graph. The ontology becomes a structural constraint rather than just documentation.
This maps onto the agent-guardrails discipline (bound model output by what it can reach, not by whether it looks safe) applied to knowledge extraction rather than to agent actions.
GraphBaby
A browser-based demo; runs entirely in-browser via WebLLM. The in-browser execution is a practical demo choice, not a production pattern.
Tier note (T3)
Medium article by a practitioner (Vishal Mysore). Content is concrete — actual demo, specific finding about schema constraints — but self-published and not peer-reviewed. A research paper on ontology-constrained LLM extraction would be the T1 upgrade.
Related
ontology · knowledge-graph · retrieval-augmented-generation · vishal-mysore