A domain-specific RAG framework for aviation intelligence: integrating localized LLMs with hybrid reasoning

Abstract

Driven by the strict requirements of the "Intelligent APP" project—completely offline local deployment and acceptable performance under constrained hardware resources—this study proposes a localized intelligent question-answering system integrating Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) technology to address highquality knowledge acquisition and data security risks in aviation R&D. General-purpose models often suffer from domainspecific knowledge gaps and hallucinations. To overcome this within strict hardware limits, we integrate 7B-parameter general and deep-thinking models, along with a lightweight text-embedding model, to efficiently vectorize data into a local database. By employing prompt-tuning strategies, the system dynamically retrieves relevant knowledge fragments via vector similarity matching to guide response generation, thereby effectively suppressing hallucinations. The proposed framework establishes a completely internet-independent, closed-loop workflow encompassing query preprocessing, precision local retrieval, and answer optimization. Quantitative evaluation using the RAGAS framework demonstrates robust system performance. Measured on a scale of 0 to 1, the Vector RAG system achieved a context precision of 0.99, a context recall of 0.81, and a faithfulness score of 0.87. The answer correctness reached up to 0.73, significantly approaching the performance boundary of cloud-based massive models. The results indicate that this fully localized architecture significantly enhances the accuracy and security of knowledge retrieval for aviation professionals, providing a replicable, resource-efficient, and secure technical solution for LLM deployment in specialized vertical industries.