ContextRAG: LLM-Free Graph Construction for RAG Systems

ContextRAG is a retrieval-augmented generation (RAG) system that utilizes a graph-based structure without depending on large language models (LLMs) for extracting entities or relationships. It employs residual-quantization k-means and Formal Concept Analysis using Lukasiewicz residuated logic to create a fuzzy concept graph from chunk embeddings. Context nodes, derived from bridge-like and meet operations, are generated through soft fuzzy joins. In a 130-task UltraDomain subset, ContextRAG only needed 30 LLM calls and 22,073 tokens to build its index, in contrast to a local HiRAG variant that required 870 calls and 3.54 million tokens for a 20-task subset. This method greatly lowers both token and time costs, enhancing scalability for extensive corpora.