CASE STUDY / 01
Make an institute's research queryable, without letting the model make answers up.
A chatbot that answers with the institute's own publications, researchers and patents — never at random.
- ROLE
- AI design & reasoning engineering
- SCOPE
- Ontology modeling · RAG · Graph-constrained reasoning
- STATUS
- In development
- YEAR
- 2026
CONTEXT
The problem
The institute's research output — publications, researchers, patents — is hard to explore unless you already know where to look. A plain LLM chatbot would answer fast, but with no guarantee the answer is true or traceable back to its source.
RESPONSE
The response
I modeled an ontology of the institute's research, then built graph-constrained reasoning: the model only answers by grounding itself in entities and relations that actually exist in the graph, retrieved through semantic search.
FLOW
From question to answer
- 01
Ask
The user asks a question in natural language.
- 02
Retrieve
Semantic search across the institute's knowledge graph.
- 03
Constrain
Reasoning is constrained to the retrieved entities and relations.
- 04
Answer
A reliable answer, traceable back to its sources.
SYSTEM
Reasoning pipeline
Ingestion, agents, memory, response — reasoning stays graph-constrained at every step.
TRADE-OFFS
Design decisions
Ontology before prompting
Structure the domain in a graph rather than leaving everything to prompt engineering.
Constrained reasoning
Deliberately narrow the answer space to reduce hallucinations.
EVIDENCE
What it changes
- Answers grounded in real sources, not freely generated.
- Institutional research finally searchable in natural language.
NEXT PROJECT
E-OPS