RAG vs. Context-Window in GPT-4: accuracy, cost, & latency