Collision to Cognition: Hash-Driven Graph Construction for Efficient RAG

Retrieval-Augmented Generation (RAG) has long been a promising paradigm for enhancing large language models (LLMs) with ex-ternal knowledge. Current embedding-based methods can capture semantic similarity but struggle to establish fine-grained, interpretable logical connections. Recently, GraphRAG has gained increasing popularity for its capability in modeling logical relations. However, it requires substantial API token usage for triple extraction or textual summarization during graph construction, which makes the entire process inefficient and expensive. In this paper, we pro-pose MeshRAG, a novel framework that M ines E fficient S tructures via H ashing for improved RAG. We jointly model chunk-level interactions and community organizations, with the global graph structure naturally emerges from locality-sensitive hash collisions. By replacing neural embedding search with lightweight bit-wise operations, MeshRAG automates a simple and rapid graph construction process. Furthermore, the hash collision mechanism provides transparent evidence for logical connections and retrieval decisions. Experimental results show that MeshRAG outperforms existing base-lines, while its graph construction requires no GPU resources or API budget and can structure over ten thousand chunks within a few minutes.