A Structure-Aware Knowledge Retrieval Framework for Hybrid Table–Text Aeronautical Information Publications in Intelligent Air Traffic Management

Abstract

Aeronautical Information Publication (AIP) documents provide authoritative operational information for flight planning, airport operations, and air traffic management. However, their hybrid table–text layouts make reliable knowledge retrieval difficult, particularly when operational questions require field-level table interpretation, cross-section evidence aggregation, or verification of safety-critical constraints. Existing large language model (LLM) based retrieval-augmented generation (RAG) methods usually treat AIP documents as plain text, which can destroy row–column semantics and weaken the traceability of generated answers. To address this problem, this paper proposes a structure-aware RAG framework for AIP knowledge retrieval in intelligent air traffic management. The framework uses separate Schema and Cell retrieval channels to locate field definitions and concrete values, constructs a compact virtual DataFrame from retrieved evidence, and employs a ReAct-based code execution agent for deterministic reasoning. Experiments on a real-world AIP question-answering benchmark covering 11 major civil airports and 500 operational queries show that the proposed framework improves both answer accuracy and traceability compared with unstructured retrieval baselines. Additional analyses of retrieval fusion, model choice, ablation settings, and error types show that preserving table structure is important for aviation-oriented transportation information services.