A Substring Extraction-Based RAG Method for Minimising Hallucinations in Aircraft Maintenance Question Answering
Quentin Signé, Mohand Boughanem, José G. Moreno, Thiziri Belkacem · 2025
Hallucination occurs when a language model generates plausible yet nonfactual information. In particular, faithfulness hallucinations (inconsistency with a given context) cannot be tolerated in critical domains such as aircraft maintenance due to the potentially severe consequences. To mitigate this issue, Retrieval Augmented Generation (RAG) methods have been introduced. These approaches are relevant for reducing the risks of hallucination but do not eliminate them, as the generator may still produce content unfaithful to the retrieved context. This paper proposes a novel RAG approach that leverages a substring extraction tool from retrieved documents to minimise hallucinations. Experiments performed on real aircraft maintenance documentation revealed that, despite the lower accuracy of the answers compared to traditional RAG methods, the proposed approach demonstrates an improved control over hallucination risks. This highlights the potential of our method in highly technical use cases where accuracy and reliability are key.