https://vapi.ai/ logo
Join Discord
Powered by
modelOutputInMessagesEnabled not working
# support
t
Call ID: https://dashboard.vapi.ai/calls/570433bc-a173-4ef4-9ca3-6f39b9973f60 Timestamp: all along the conversation Response expected: I am expecting bot transcript to be the model message, not the transcript from my transcriber Response actually received: the bot transcript was the transcript from transcriber. I tried 2 different transcribers (deepgram and talkscriber). Also for the voice I am using Cartesia All settings are for e conversation in French (transcriber and voice) In my agent, I have used the API to explicitely enable modelOutputInMessagesEnabled: true
p
Call ID: [570433bc-a173-4ef4-9ca3-6f39b9973f60](https://dashboard.vapi.ai/calls/570433bc-a173-4ef4-9ca3-6f39b9973f60) @[TheNewPenguin](https://app.plain.com/workspace/w_01HYEQ5FKPH8CPX2MS608B6S51/customers/c_01JMT4QHTNV3829709AJFRTWWA/) I do see in the call logs at few parts of it was different from model output such as: \[1\] 🔵 19:25:55:944 Voice Input Formatted: " Pouvez-vous me confirmer votre adresse email ?", Original: " Pouvez-vous me confirmer votre adresse email ?" Actual: pouvez-vous me confirmer votre adresse mail ? Parfait. \[2\] 🔵 19:25:55:891 Voice Input Formatted: "Très bien, je vais commencer par confirmer votre email.", Original: "Très bien, je vais commencer par confirmer votre email." Actual: Très bien, je vais commencer par confirmer votre mail. https://cdn.discordapp.com/attachments/1343304419210694656/1343442519077289984/Screenshot_2025-02-24_at_10.08.14_AM.png?ex=67bd49d8&is=67bbf858&hm=f2c2dc16a3b17bbe10f6bca518dcd29c92fea7a894d9d46d79dbf9594f20684c&
t
Hey there. Thanks for your reply. Is it an issue? How can I fix it or is it on your end?
Do you need me to do more tests? Happy to do them in english if that helps you
s
@TheNewPenguin If you have a few more call IDs, please share them. It would be helpful to reproduce them and resolve the issue more quickly.
r
i am seeing the same problem with my assistant. Call ID 
235d6fe9-6e74-46f5-8002-e0c7f4f2311c
has an example where the assistant message in the conversation history says something like "Email from Air Force Marathon about the two thousand twenty five Museum Mile" when the model output is "Email from Air Force Marathon about the 2025 Museum Mile". I have set modelOutputInMessagesEnabled via the API. Assistant ID 
48b68590-564a-44b2-94d4-7a7a649e7c53
i deleted this assistant and re-created an identical one (assistant ID 
49d99b8b-d351-4686-8b5d-c3fd48f8eb08
) and the problem has gone away 🤔
t
Hey @Shubham Bajaj
Hey @Shubham Bajaj Here is a few 0a003df0-987c-46fa-a6ff-04d2ff101806 0d1b01a7-534e-483c-99ef-aae74de1a6fc (this one is in english)
My assistant was created by API
r
I mentioned that deleting / re-creating the assistant fixed it but that was premature... it worked at first but no longer
s
@TheNewPenguin @Rafael Garcia looking into it.
@TheNewPenguin for this call I do see model output in messsage array. Anything between * and * is removed because of voice formatting plan.
Copy code
0d1b01a7-534e-483c-99ef-aae74de1a6fc (this one is in english)
https://cdn.discordapp.com/attachments/1343304419210694656/1347196992920485898/Screenshot_2025-03-06_at_6.49.05_PM.png?ex=67caf278&is=67c9a0f8&hm=b7f3fe268cfeb741ba05857baab144c84b3901105ba7440c2306c3697b138578&
@TheNewPenguin let me know your thoughts on this. "transcript": "AI: Hello?\nUser: Yes. Is this John?\nAI: Yes. Speaking. Who's calling?\nUser: Yeah. It's Trevor from o 2. Don't know if you'll remind me.\nAI: You said Trevor from where? And what is this regarding?\nUser: Yeah. It's up to we are helping companies to grow their business. With, uh, training sales training program and recruitment. Does that ring a bell? Maybe it's on your Agenda for the coming months.\nAI: Look, Trevor. Get a lot of calls about sales training. I'm quite busy right now. What exactly is your reason for calling?\nUser: Uh, I would like to book a demo with you maybe next Thursday at 2 PM if you are available.\nAI: No. That's not gonna work. I don't schedule meetings without understanding the value proposition first. What specifically makes your sales training different from the dozens of other providers out there? S top 123. The conversation ends here. Because the salesperson, 1, Wasn't clear about their purpose. 2, jump straight to booking without establishing value. 3. Didn't address my needs or concerns. 4,\n",
@Rafael Garcia I was looking at you call logs and yes for this model output in messages enabled didn't worked. I am tracking this issue why sometimes it doesn't works. Sorry for the bad experience.
r
Ok thanks for looking into it. I can share more examples when I test it some more later today
g
Hello @Shubham Bajaj , another example: 1a670edf-1e66-476e-97df-a0940d32961f At the end of the conversation, the LLM output was "Voice input: Oui bien sûr, c'est le zéro six cinquante deux quarante-huit trente-neuf vingt et un." and transcript was "Oui bien sûr, c'est le zéro 6 cinquante-deux quarante-huit trente-neuf vingt et un. " which says the transcript is not the output of the LLM but a Speech to text
s
@GHP thanks for bringing in this issue, can you please create a new #1211483291191083018 ticket.
PreviousNext
Powered by