Standard vector Retrieval-Augmented Generation (RAG) relies on top-k cosine similarity over chunked text. While effective for localized semantic queries ("What is our return policy?"), it fails at topological multi-hop synthesis ("Which subsidiary owned the vendor that supplied defective chips to our 2023 EV line?").

The Vector Blindspot

When an answer spans four separate documents that share zero lexical proximity with the root prompt, vector search yields fragmented context. In contrast, GraphRAG navigates explicit paths:

  • Global Summarization: Hierarchical Leiden community detection generates high-level cluster abstractions.
  • Path Traversal: Traverses explicit relation links regardless of physical document boundaries.
  • Deterministic Lineage: Every extracted fact traces back to a verified edge and source entity.