Rédiger un PREreview

Retrieval-Augmented Generation (RAG) systems face substantial challenges when navigating large volumes of complex scientific literature while maintaining reliable semantic retrieval, a critical limitation for automated scientific discovery where models must connect multiple research findings and identify genuine knowledge gaps. This study introduces a question-based knowledge encoding method that enhances RAG without fine-tuning to address these challenges. Recognizing the lack of syntactic understanding in major Large Language Models, we generate syntactic and semantic-aligned questions and apply a syntactic reranker without training. Our method improves both single-hop and multi-hop retrieval with Recall@3 to 0.84, representing a 60% gain over standard chunking techniques on scientific papers. On LongBenchQA v1 and 2WikiMultihopQA, which contain 2000 documents each averaging 2k-10k words, the syntactic reranker with LLaMA2-Chat-7B achieves F1 = 0.52, surpassing chunking (0.328) and fine-tuned baselines (0.412). The approach additionally reduces vector storage by 80%, lowers retrieval latency, and enables scalable, question-driven knowledge access for efficient RAG pipelines. To our knowledge, this is the first work to combine question-based knowledge compression with explicit syntactic reranking for RAG systems without requiring fine-tuning, offering a promising path toward reducing hallucinations and improving retrieval reliability across scientific domains.