Retrieval-augmented generation (RAG) is the standard way to equip large-language models (LLMs) with knowledge that lies outside their training data. The classic vector-store recipe works well when a single, tightly focused passage suffices, but it struggles whenever a question demands evidence scattered across many documents. Microsoft’s GraphRAG addresses this by prompting an LLM to extract entity-relation triples from each chunk, then querying a community-partitioned knowledge graph. However, this approach requires expensive LLM calls per chunk, generates thousands of triples, and extends preprocessing time to several hours even for modest-sized corpora. We show that constructing a triple-based knowledge graph can be skipped entirely. Instead, our proposed method, GraphRAG-V, treats each raw chunk as a node, embeds them once, builds a similarity graph, and applies the scalable VLouvain algorithm to discover chunk-level vector communities in embedding space. A single pass then summarizes each community, enabling the LLM to receive both local and global evidence in one prompt. On the 2,556-question MultiHopRAG benchmark, GraphRAG-V indexes the corpus in minutes, answers all questions in under an hour on a single A100 GPU, and improves recall by eleven percentage points over a strong vector-store baseline. It also outperforms Microsoft’s GraphRAG across all metrics, while being orders of magnitude faster. These results suggest that implicit chunk-level structures, not explicit triple-based graphs, are the key to scalable, cost-efficient long-context reasoning.