A financial services team runs a RAG-based chat application in Azure AI Foundry. Users report that responses take 8–12 seconds during peak hours, but the deployed model's benchmarked latency is under 1 second. The team enables tracing in Foundry to diagnose the bottleneck. They observe spans for the full request lifecycle. Which tracing signal should they examine FIRST to isolate whether the delay originates from retrieval rather than model inference?