SafeLLM: Extraction as a Hallucination-Resistant Alternative to Rewriting in Safety-Critical Settings

The paper introduces SafeLLM, which evaluates extraction methods as a more hallucination-resistant alternative to rewriting in retrieval-augmented generation (RAG) systems for safety-critical applications. It compares various prompting strategies, notably line-number-based source selection, which achieves up to 95% term recall and outperforms other methods across different model scales and document types. This research is significant for practitioners as it highlights effective strategies for enhancing precision and safety in LLM applications, particularly in regulatory and compliance contexts.