Hi, I want to create a squad that acts as a single assistant (from the caller perspective). I thought I could do this by setting the transfer message in the prompt to silently transfer the call, and by setting the firstMessageMode of the 2nd assistant to 'assistant-speaks-first-with-model-generated-message' . However, even with this, the assistant says "Let me transfer you", before transferring to the 2nd assistant also sometimes there is an awkward silence and I have to speak first to continue the conversation with the second assistant. Is there a way to prevent this? call ID - 013d9207-fa8c-4ac6-a600-fd9ef242aa52

@08anas I understand you're looking to create a seamless experience where the transfer between assistants in a squad is invisible to the caller. To achieve this, you'll want to ensure that the transfer message is indeed set to be silent and that the second assistant is configured to initiate the conversation without any prompt from the user. For the silent transfer, you should check the configuration of your squad to make sure that the transfer message is not set or is explicitly configured to be silent. If the assistant is still announcing the transfer, it might be a setting that needs to be adjusted within the squad configuration. Regarding the second assistant initiating the conversation, the

firstMessageMode

should be set to 'assistant-speaks-first-with-model-generated-message' as you've done. However, if there's still an awkward silence, it could be due to a delay in the assistant's response or a misconfiguration. To resolve these issues, please ensure the following: 1. The transfer message is set to be silent in the squad configuration. 2. The

firstMessageMode

for the second assistant is correctly set to 'assistant-speaks-first-with-model-generated-message'. 3. There are no delays in the assistant's response due to network issues or processing time. If you've checked these settings and the issue persists, it might be necessary to look into the specific configurations of your squad and the assistants involved. Since I don't have access to the system settings or the ability to modify configurations, I would recommend reaching out to support with the call ID you provided for a more detailed investigation. Please let me know if there's anything else I can assist you with!

All of these are already there, still it is not working. @Shubham Bajaj , can you please look into this one.

Hey @08anas To help track down this issue, could you share: - What response you expected to get - What response you actually got instead This would really help me figure out what went wrong!

I don't want the message from the assistant such as "I will transfer the call to XYZ assistant" or "Let me transfer the call to the correct assistant". The assistant should silently transfer the call to the other assistant seamlessly w/o giving any information about the assistants transfers.

@08anas What happened is that it generated the prompt using the Knowledge Base, and then when the user said "hello" again, it performed the actual transfer. To troubleshoot this, you can: 1. Remove the Knowledge Base and try another call 2. On the second call, try using Trieve as the Knowledge Base provider Please let me know how it goes. I've included the relevant logs below for your reference.

Hi @Shubham Bajaj , I have removed the Knowledge Base from the Assistant and still facing the same problem. This is the call ID: 68806a5b-9978-4bf4-a413-7de18d04d6af If possible can we connect today and get this thing resolved?

@08anas I have looked into this call ID and it appears to be related to prompting. While your instructions are clear, you're relying on LLM reasoning for transfer and response generation which is causing inconsistent behavior. I suggest reviewing the silent squad transfer approach as an alternative solution. https://www.notion.so/vapi/Silent-Squad-Transfers-14e008c8393f80abbcd6d1a5c10532b8

Already referred to this while creating the prompt and transfer description.

@08anas The difference is in the deterministic approach. In your prompt, you have the LLM decide how to do the transfer, which is correct, but you have to be more specific here. For example, I only want to do the tools calling instead of "Use the appropriate assistant based on membership status," which gives the scope of doing both content generation and tool calling. This is what I was referring to - rework your prompt to make it more deterministic. [YOUR PROMPT]

Step 3: Assist the Caller
Ask, “How can I assist you today?”
Listen without assumptions and ask clarifying questions if needed.
For appointment scheduling:
Use the appropriate assistant based on membership status.
Do not mention membership status or the transfer process.
For other inquiries (e.g., company, pricing):
Use the relevant assistant based on membership status.
Avoid discussing the transfer process.
If out of service area = true, inform the caller and offer to transfer them to human support.
If unresolved, confirm if the caller wants human support and proceed with the transfer.

[EXAMPLE]

[Task]
Greet the customer and ask if they are interested in purchasing widgets.
Wait for the customer's response.
If the customer is interested, ask for their name.
Wait for the customer's response.
Ask how many widgets the customer would like to purchase.
Wait for the customer's response.
Confirm the order details with the customer.
trigger the transferCall tool with Payment HPPA Assistant

@Shubham Bajaj But this is not the exact prompt I have in my assistant. Is this getting generated as soon as the caller is stating about the problem? Also, i just wanted to know how to make it more deterministic if there are 5 assistants to which the call can get transferred based on the certain conditions? I know this query is becoming subjective so that is why I was hoping we can connect and resolve it easily.

@08anas yeah correct , let's connect dropping you link in your DM. https://cdn.discordapp.com/attachments/1326458283531305030/1327621106928058428/Screenshot_2025-01-11_at_18.16.01.png?ex=6783bb01&is=67826981&hm=775bc9783c88c064dfdbff1838d5d77e0c86457c6275b6af8922ca2c60a193b7&

@Shubham Bajaj Done, I have booked the slot for Tuesday evening. For now I will try to update the prompt and keep you posted on it.

@08anas can you join today's discord office hours? we'll take it live over there.

@Shubham Bajaj At what time is it?

I can try joining because I might not be free at that time. But if not, we are connecting tomorrow, right?