We're working with spanish, and we've seen errors during the call, where the LLM was replying no-sense. After debugging and trying to optimise the prompt at our best, we found that running the EVAL everything works amazingly. So we copy pasted exactly what VAPI spit out from the "transcriber" from the AI (not talking about the user here!!!) and we are able to reproduce the same issue. Then we discover this "modelOutputInMessagesEnabled": true, However if you get the https://api.vapi.ai/call/019e77dd-bf95-744e-86d9-7069f1593ab4 you can literally see that the transation is incorrect. This is what the Bot has been instructed to say: "'Gracias por compartirlo. Antes de empezar, te explicaré rápidamente cómo será la entrevista. Te haré algunas preguntas para conocerte mejor y entender tu experiencia. Durará unos 10 a 15 minutos... ¿Listo para comenzar?" These are the logs from VAPI:

{
                "role": "bot",
                "time": 1780127504455,
                "source": "",
                "endTime": 1780127517485,
                "message": "Gracias por compartirlo. Antes de empezar, explicaré rápidamente cómo será la entrevista. Quería => ("ERROR HERE, it must be Te haré") algunas preguntas para conocerte mejor y entender tu experiencia. Dorará unos 10 a 15 minutos. ¿Listo para comenzar?",
                "duration": 10479.9970703125,
                "secondsFromStart": 22.42
            },

019e77dd-bf95-744e-86d9-7069f1593ab4 Now is the API not outputting exactly what has been sent to the LLM now that we have that setting enabled or what?

Hi there, Thank you for your message. Our team is currently out of the office. We operate Monday through Friday, from 9:00 AM to 8:00 PM Pacific Standard Time (PST). We’ll get back to you as soon as possible during our normal business hours. If your message is urgent, please mark it accordingly or include “URGENT” in the subject line, and we’ll do our best to respond promptly. Warm regards, Vapi Customer Support Team

From the log LLM Request Started, seems like is still sending the transcript vs the output.

Hi @Alberto , Right now modelOutputInMessagesEnabled only works with ElevenLabs. To put the model's exact text into the conversation history, we need per-word timestamps from the TTS so we know precisely what was spoken and when, mainly to handle interruptions correctly (on a barge-in, we only keep the part actually spoken before the cutoff, with the right timing). ElevenLabs has a word-timestamped streaming endpoint that gives us this. The other TTS providers don't expose that yet, so we can't support it reliably with them for now.

Hi @Sahil i am a bit confused as someone in support was saying something different

Please ignore the support message, I looked into the codebase directly and confirmed it. No, there is no other way to do it.

Thanks @Sahil ! Is there a way to get full access to simulations and be one of the first guinea pig? We have the perfect use-case as we do run job iterviews and would be great to use that vs the dev team writing tests 😄

will let the team know about it. Can you share your org_id.

@Sahil 🙂 7456a09b-1f5f-44bc-a5e9-fdd5ba169548 I'm already using it as I "found" the link

thanks!