We’re seeing an increased rate of silence-timeouts since switching to Gemini 2.5 Flash (Through custom-llm) Example: Call ID

70c43c6f-5045-41e4-908a-4503214f17d1

from yesterday. The agent did not play any configured silence fillers (e.g. “I’m having a hard time hearing you…”). Instead, the call auto-timed-out during a long silence. End-of-call report confirms silence fillers are configured. Vapi logs show the LLM responded with “Great, What’s your first name?”, but this was never sent on the call, leading to a timeout. I also ran test calls today, and silence fillers did work, so this seems intermittent. Could you help investigate why the LLM response wasn’t sent and why silence fillers didn’t trigger in this call?

Hey, Our team will look into this and get back to you soon. We appreciate your patience.

Still waiting 🙂

With how large the prompt is, the assistant most likely had difficulty retrieving the next set of instructions within the large context. If you reduce your prompt size by offloading your menu to a knowledge base text file, you will see an improvement in consistency in your assistant performance.