I had a call the other day where the user was asked a question, they clearly answered "yes", the transcriber transcribed as "yes" with a confidence of .99..., but when the LLM call is made the user message (yes) is not appended to the prompt causing the LLM to respond with the same question again. Happened three times in a row for the same call. Screenshot of the call transcript is attached, as is the detailed log showing the transcription result and accuracy along with the next log where the LLM is called WITHOUT the user message that was accurately transcribed. This feels like a system issue, not a configuration issue, please advise. Call ID: 019a788c-6cbe-700b-aa12-2232fe45353d https://cdn.discordapp.com/attachments/1438910575449997343/1438910575966027918/Screenshot_2025-11-14_at_9.08.36_AM.png?ex=6942c966&is=694177e6&hm=90cdf3221a4f421492869a63a311f0c90d6f6fd257a5cf4542fc7b276d20ab94& https://cdn.discordapp.com/attachments/1438910575449997343/1438910576452571157/Screenshot_2025-11-14_at_9.13.50_AM.png?ex=6942c967&is=694177e7&hm=dc246156fc2755b6d3700a07f55781f9b1538a603ece665ad21e955520820698& https://cdn.discordapp.com/attachments/1438910575449997343/1438910576758886501/Screenshot_2025-11-14_at_9.14.39_AM.png?ex=6942c967&is=694177e7&hm=7b2e88e5429c5df7bb57fb249845b8105f33419f9523e5c28cb90c811601e8bb&

Suggestion: Since the Vapi system is designed to stream transcribed user input directly into the LLM prompt in real time, a failure in this step (as you described) points to a possible system-level orchestration issue. The best next step is to report this behavior to Vapi support with your call logs and transcript details for further investigation, as this is not expected and cannot be resolved via configuration. For more on how the message flow should work, see the [Vapi quickstart documentation](https://docs.vapi.ai/quickstart). Source: - [Vapi Quickstart Documentation](https://docs.vapi.ai/quickstart)

How do I report this to Vapi support?

To report this system issue to Vapi support: - Email support@vapi.ai with your call logs, transcript screenshots, and a description of the problem (including the Call ID). This ensures your report reaches the Vapi team directly for prompt investigation. Reference: [Vapi Support Documentation](https://docs.vapi.ai/support) Source: - [Vapi Support Documentation](https://docs.vapi.ai/support)

Message marked as helpful by @Ernest! 🎉

REOPEN

The issue was caused by weak conversation logic. The assistant ignored the caller’s initial request for a payoff/guarantee letter, then kept repeating the same “resolve or manage an account” question. Because it didn’t handle ambiguous answers like “Yes,” it got stuck in a loop instead of rephrasing or using the context it already had. It also delayed transferring to a representative even after multiple requests. The assistant needs better context use, clearer follow-up questions, and faster human handoff when requested.