The RDF road not taken
One road not taken: RDF + SPARQL. The same fact, three ways. As a property-graph relationship: (:Person{name:'Keanu'})-[:ACTED_IN]->(:Movie{title:'The Matrix'}). As RDF triples: :keanu :acted_in :the_matrix . and :keanu rdf:type :Person .. And a SPARQL Basic Graph Pattern ?actor :acted_in ?movie . binds variables exactly like the binding model you built — your executor is, conceptually, a BGP matcher.
When to choose which: RDF/SPARQL for open-world, standardized, reasoning-heavy knowledge graphs (Wikidata, DBpedia, ontologies). Property graphs/Cypher for performance-critical, edge-attribute-rich operational graphs (fraud, recommendations, GraphRAG). The two are converging — GQL is the new ISO standard for property graphs, and RDF-star adds edge properties to triples.
You're building GraphRAG: an LLM answers questions by retrieving facts from a knowledge graph, and you must attach a confidence score and source_url to each MENTIONS edge. Which model fits with the least friction?