Vector Graph RAG: Multi-Hop RAG Without a Graph Database

Standard RAG falls apart when the answer isn't in one chunk. Ask "What side effects should I watch for with the first-line diabetes medication?" and the system needs to first figure out that metformin is the first-line drug, then look up metformin's side effects. The query never mentions "metformin" — it's a bridge entity the system has to discover on its own. Naive vector search can't do this. The industry answer has been knowledge graphs plus graph databases. That works, but it means deploying Neo4j or similar, learning a graph query language, and operating two separate storage systems. The complexity doubles for what's essentially one feature: following entity chains across passages. I built Vector Graph RAG to get multi-hop reasoning without any of that overhead. The entire graph structure lives inside Milvus — entities, relations, and passages stored as three collections with ID cross-references. No graph database, no Cypher queries, just vector search and metadata lookups. Buildi