Vector Graph RAG: Milvus retrieval and multi-hop demo

Vector Graph RAG stores knowledge graph entities and relationships as vectors in Milvus. The presenter explains how vector search handles graph traversal and expansion within the same database, avoiding a separate graph database such as Neo4j. The video introduces the approach and demonstrates its retrieval process through a frontend that viewers can deploy locally.

The demo uses a Harry Potter dataset and asks a question linking Voldemort's teacher to the protection of the Philosopher's Stone. Initial retrieval finds four entities and 20 relationships but misses Dumbledore, the connecting entity. Subgraph expansion increases the candidate set to 36 entities and 83 relationships. An LLM reranking call selects five relationships, which then go to an LLM with their source text to generate the answer. The graph contains two structurally separate Dumbledore entities; the presenter says this duplication does not affect the final answer.

The speaker reports 87.8% average recall across three multi-hop question-answering benchmarks, 19.6 points above standard RAG, and performance comparable to HippoRAG 2. These are the project's reported results. Installation is introduced through pip, and the frontend exposes intermediate retrieval states for inspection. The local interface demonstration does not establish that the LLM calls run offline.