RAG or Long Context ?? This paper finds, when resourced sufficiently, Long Context consistently outperforms RAG in terms of average performance.🤯 However, RAG’s significantly lower cost remains a distinct advantage. To solve this they propose a new 'Hybrid' method ----- Key Insights 💡: • Long Context consistently outperforms RAG when resourced sufficiently • RAG remains relevant due to significantly lower computational cost • Over 60% of queries yield identical predictions from Long Context and RAG Solution in this Paper 🔧: • SELF-ROUTE: A hybrid approach combining RAG and Long Context • Uses model self-reflection to route queries • RAG-and-Route step: Provides query and retrieved chunks to LLM • Long-context prediction step: Uses full context for unanswerable queries • Leverages strengths of both RAG and Long Context Results 📊: • SELF-ROUTE achieves comparable performance to Long Context at lower cost • Cost reduction: 65% for Gemini-1.5-Pro, 39% for GPT-4O • Performance: Slight drop for GPT-4O (-0.2%) and Gemini-1.5-Pro (-2.2%) • Improvement for GPT-3.5-Turbo (+1.7%)