Running summarizing tools on a new article

Abstract: Large language models (LLMs) frequently hallucinate, producing fluent yet factually incorrect information. Retrieval-Augmented Generation (RAG) mitigates this, but state-of-the-art systems require GPU infrastructure, embedding models exceeding 100 million parameters, and supervised finetuning-prerequisites unavailable in resource-constrained settings. This paper presents LightweightRAG, a training-free hybrid retrieval framework combining BM25 sparse retrieval with the compact all-MiniLM-L6-v2 model (22.7M parameters) via Reciprocal Rank Fusion, augmented with evidence consistency verification and conditional abstention, operating exclusively on CPU infrastructure. Evaluation on SQuAD v2.0$(\mathrm{n}=500$, real generation;$\mathrm{n}={300}$, oracle extraction) and TriviaQA$(\mathrm{n}={200})$on Google Colab demonstrates statistically significant hallucination reduction, driven primarily by hybrid retrieval. On SQuAD v2.0, hallucination drops from 98.0% to 41.5% (57.7% relative reduction,$p<0.001)$, and on TriviaQA from 92.0% to 67.7% (26.4% relative reduction,$p<0.001)$, with the majority of the improvement attributable to hybrid retrieval rather than verification alone.