This is usually a timing issue in the VAPI pipeline. The filler isn’t triggering early enough because Gemini KB retrieval delays the first token, so VAPI never fires the “start speaking” event in time. By the time the response is ready, the turn state has already moved forward, which is why you see cancellations instead of fallback switching. Fix is to trigger the filler at stream start (first token event), not after tool/KB completion, and ensure Gemini isn’t blocking initial token streaming. I can help with that @Isha Malik