An arXiv Paper Question-Answering System Based on Qwen and RAG

This paper explores combining large language models and retrieval-augmented generation (RAG) techniques to build a question-answering system for academic papers. The system is built upon Qwen2.5 models and open-source tools like LlamaIndex, utilizing the arXiv dataset to implement an end-to-end pipeline from user queries to relevant answers through models including query routing, hybrid retrieval, and answer generation. In the retrieval phase, the system leverages both the BGE-M3 embedding model and the BGE-Reranker-V2-M3 reranking model. The generation stage integrates knowledge from multiple sources using a fine-tuned Qwen2.5 model. Experiments show significant improvements in retrieval accuracy and answer coherence by incorporating Reranker and RAG-Fusion. This study highlights the potential of LLM and RAG technologies in academic applications and offers insights for enhancing paper question-answering systems.