Although large language models (LLMs) are widely used in natural language processing, they still face limitations in long-context comprehension, especially in specialized domains. Retrieval augmented generation (RAG) can improve the accuracy and reliability of question-answering systems. However, traditional RAG approaches often rely on a single retrieval path, which can cause semantic bias and limited retrieval coverage when processing complex queries. This study proposes an improved RAG intelligent question-answering system based on the Haystack framework. Its core innovations include: first, a multidimensional query generalization strategy based on prompt engineering is adopted to semantically expand and reconstruct original queries, generating diverse query variants to broaden retrieval coverage; second, a parallel dual-path recall mechanism is designed, combining the strengths of dense retrieval and sparse retrieval to capture query relevance at both semantic and lexical levels, thus improving retrieval precision and recall; third, a fusion and reranking module selects high-quality document fragments, which are then fed into a large language model to generate accurate answers. Experimental results on the PubMedQA medical dataset demonstrate that the proposed method significantly improves the retrieval effectiveness and generation quality of the question-answering system, validating its effectiveness and superiority in specialized-domain question-answering tasks.