Lots of people talk about "knowledge graphs" these days and this often gets associated with GraphRAG and graph DBs like Neo4j.

Graph databases sound amazing but from experience they're really hard to get right with unstructured data and introduce more baggage than they're worth for RAG.

I outline the architecture we've adopted instead here. You can get get really good results and mileage with MongoDB + any search index that can handle hybrid (e.g. vector + bm25) search. The fancy solution often isn't worth it.