Traditional Retrieval-Augmented Generation (RAG) methods face two major challenges when processing specialized technical documents in fields such as civil aircraft manufacturing processes due to their fixed segmentation strategies: first, loss of contextual information caused by disrupting the inherent chapter structure of documents; second, lack of cross-document associations resulting from treating documents as isolated information silos. To systematically address these issues, this paper proposes a novel Hyperscale Cross-Document Retrieval Framework (HCRF). This framework aims to fully leverage the intrinsic structure and semantic relationships within documents by constructing document structure trees, performing cross-document topic clustering and abstract generation, and implementing hierarchical retrieval and hybrid ranking strategies. Experiments on a self-built civil aircraft manufacturing Q&A dataset demonstrate that HCRF significantly outperforms existing baseline models in both retrieval accuracy and answer generation quality, validating the framework’s unique advantages in processing structured technical documents.
Zhi et al. (Fri,) studied this question.