TLD-HyperRAG: adaptive retrieval and faithful generation via theme-logic-detail hypergraphs

Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm for enhancing large language models in knowledge-intensive and domain-specific tasks. However, existing RAG methods often rely on fixed retrieval strategies and insufficiently coordinate knowledge indexing, query processing, and evidence-grounded answer generation. To address these limitations, this paper proposes Theme-Logic-Detail Hypergraph-based Retrieval-Augmented Generation (TLD-HyperRAG), a multi-layer hypergraph framework designed for query-adaptive and evidence-faithful RAG. The proposed framework constructs a Theme-Logic-Detail hypergraph index to organize domain knowledge across multiple semantic levels, including theme navigation, logical linkage, and detailed evidence. Based on query complexity classification, TLD-HyperRAG dynamically selects different retrieval strategies for simple, moderate, and complex queries, thereby balancing retrieval efficiency, evidence coverage, and reasoning depth. In addition, a faithfulness-checked conditional rewriting mechanism is introduced to assess the consistency between generated answers and retrieved evidence and to revise answers when the evidence support is insufficient. Extensive experiments on eight specialized domains, including UltraDomain, BioASQ, and FinanceBench, show that TLD-HyperRAG achieves the best performance in the large majority of evaluated domains and metrics in terms of answer accuracy, retrieval quality, generation quality, and evidential faithfulness. These results indicate that coordinating multi-level knowledge organization, query-adaptive retrieval, and faithfulness-controlled generation within a unified hypergraph framework is an effective step toward reliable retrieval-augmented generation in specialized domains.