Performance

Your indexing pipeline tokenizes 10M docs in 18 hours. What single change gives the biggest speedup?

In a production RAG system with a 2-second end to end latency budget for non-streaming responses, which step typically dominates and where should the optimization effort go?