We Built Graph RAG Without the Graph Database

Introduces 'Vector Graph RAG', a Python library that enables multi-hop reasoning for RAG systems using only Milvus. It details a novel approach to represent graph structures (entities, relations, passages) within Milvus collections using ID references. The post outlines the four-step retrieval flow: seed retrieval, subgraph expansion (one hop by default), LLM reranking of expanded relations, and answer generation. It highlights the performance benefits of avoiding a separate graph database and iterative LLM calls, achieving competitive recall on multi-hop QA benchmarks.