Talk to knowledge graph – sparql-tool
Use LLM.: Hallucinations (sometimes good for creativity), Outdated knowledge, no access to your data ( trained on public knowledge).
Solutions:
Fine tuning – retrain the model on domain data
RAG –
Search: look things up in real time. Claude code looks at searched doc.
Vector embeddings: semantic similarity search over documents.
Knowledge graphs: structure queryable machine readable knowledge
Two main flavors:
Property graphsQ: nodes and edges carry key value properties
Nodes and edges carry properties – Neo4q, TigerGraph
RDF graphs: everything is triple – Subject – predicate – object
Engines
Datasets:
Everthing is URI
So for example for age – do not put the exact age – put the date of birth and then compute the age.
The RDF ecosystem:
RDF
OWL (ontology language
Triple stores – databases for RDF
Ontologies (formal schemas
SPARQL – language for RDF
Problem with RDF
Low traction for academia
steep learning curve
SPRQL is tedious by hand
Ontologies are complex
LLM models have this formal knowledge baked in (claude etc)
They understand RDF, OWL, SPARQL, ontologies
How to use them effectively.
Sparql-tool
Democratizes RDF
Three components : skill, agent, CLI tool
CLI does not need LLM
This is replacing MCP.
Websearch sometimes does not work – because there was semantic loss on what was being searched for
By the way, people are complaining because people aren’t going to websites, they are going to LLM’s.
DBPedia KG – they extract wikipedia very often…
Try Kevin Bacon algorithm and show hops is one way to test things.
Biotech : Uniprot
Uniprot KG. (they publish about proteins)
What proteins are associated with Alzhiemer’s
So then the better question is tho what protein intearctis with what proteins ..Takes long.
Another one is used – IntAct…
Have to know where the dataset came from ?
Pokemon: Knowledge graph.
Community graphs – has 100k entries. Used that as such.
