Why does our chat UI feel slow even though responses are fast?
measure_ttftCompleted
Time to first token across the last 24 hours of sessions.
Assistant
Perceived speed is set by time-to-first-token1 and paint behaviour, not total generation time2. Stream tokens into a shell that has already painted, keep the transcript pinned while the reader is at the bottom, and reserve layout space for tool cards so nothing jumps as states resolve.
Sources
1Streaming SSR with React Server Components react.devSelective hydration lets the shell paint immediately while slower subtrees stream in as they resolve.
2Measuring Time to First Token in production vercel.comTTFT is dominated by queueing and prefill; chunked transfer keeps perceived latency flat.