A new approach to retrieval-augmented generation (RAG) systems aims to improve factuality by making evidence explicit. This method requires answers to cite specific supporting passages and allows the system to confidently state when it lacks sufficient information. By maintaining provenance with each indexed passage and implementing an 'abstention contract' in the prompt, the system can avoid fabricating answers or misrepresenting evidence. This workflow includes rigorous testing with curated datasets to ensure reliable abstention on unanswerable questions. AI
IMPACT Enhances the reliability of AI systems by ensuring they cite evidence and abstain from answering when information is insufficient.
RANK_REASON The item describes a new workflow for retrieval-augmented generation systems, which is a type of AI tool.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →