|
step-level retention
Placement
- Chunks carry their producing turn: StoredChunk.turnId (backend [email protected]) for
committed rows (persisted in the IDB cache; old cached rows are upgraded on
sync) and the live event's turnId for provisional ones (the optimistic echo is
tagged on turn-start).
- interleaveTurnMetrics attributes every group to its turn (turnId, else the
turn owning its stepId) and puts the turn total after the turn's LAST visible
group. No user-message counting: sending, streaming, steering, images, errored
turns and chat-limit unloading can no longer shift metrics between turns — at
any point during a turn, not just after it ends. A turn cut in half by the
chat limit keeps its total under its visible tail.
- Legacy rows (no turnId) fall back to an order-preserving, count-guarded match
that leaves a turn blank rather than show another turn's numbers.
- A turn that errors re-syncs metrics (the backend now persists them).
TPS
- Step TPS = outputTokens / decodeMs only (no genTotalMs fallback — it includes
time-to-first-token). Turn TPS sums tokens and decode time over timed steps
only. Tool execution is excluded by the backend.
Display
- Retention = this turn's first-step cache reads / previous turn's last-step
prompt (step-level; turn sums made it pin at 100%).
- Turn row shows output tokens + input processed (with cached), instead of a
misleading 'total tok' that summed every step's full prompt.
|
