I’m trying to reduce latency in a LangGraph-powered chatbot that uses 14 REACT-style agents. The RAG layer pulls external knowledge from a vector DB. The issue: The REACT reasoning loop triggers several thought + action cycles RAG retrieval can be slow Combined, the system takes 6–8 seconds to return a response I’m trying to identify whether the bottleneck is the multi-step REACT chain, the agent orchestration, or the vector retrieval. If anyone has optimized multi-agent REACT architectures, reduced reasoning depth, or improved RAG latency, I’d appreciate guidance — caching strategies, agent pruning, batching, or structural improvements. Continue reading...
