Some time ago, I asked if VAPI cached anything, and the answer was no. However, I am pretty sure it is caching something. I have experienced the following type of calls: 1. Hi, can you transfer me to Daniel? 2. Hang up and call again. 3. Hi, can you transfer me to any agent? 4. Bot: You want me to transfer you to Daniel, right? It seems like the bot remembers conversations from past calls. Is this correct? How can I remove this behavior? I don’t want this cache. Another example is in this call: 8869b245-4618-4c88-a2a5-90c8b4276145. When the caller requested a transfer to someone specific, the bot didn’t trigger the transfer_call tool as expected because it already knew from a past call that we didn’t have that agent’s name. Is there a way to remove this behavior?

🔵 11:17:55:542 OpenAIRequest Request. Messages:
[
  {
    "role": "system",
    "content": "\n  # OBJECTIVES\n\n  Your name is Taylor, an AI assi..."
  },
  {
    "role": "assistant",
    "content": "Hi, Jason. I see you are calling back. Would you like to be transferred to the agent"
  },
  {
    "role": "user",
    "content": "Connect me with noise."
  }
]

🔵 11:18:23:990 OpenAIRequest Request. Messages:
[
  {
    "role": "system",
    "content": "\n  # OBJECTIVES\n\n  Your name is Taylor, an AI assi..."
  },
  {
    "role": "assistant",
    "content": "Hi, Jason. I see you are calling back. Would you like to be transferred to the agent"
  },
  {
    "role": "user",
    "content": "Connect me with noise."
  },
  {
    "role": "assistant",
    "content": "I don't have an agent named Noise. Available. Can you describe what you're calling about so I can route you to the appropriate agent?"
  },
  {
    "role": "user",
    "content": "No. I want to talk to Nardis. About my It shouldn't be just"
  },
  {
    "role": "assistant",
    "content": "I don't have an agent named Nardis available. You please describe what you're calling about so I can route you to the appropriate agent?"
  },
  {"role": "user", "content": "Thank you."}
]

Thanks for the response, then why it didn't trigger the transfer_call tool when requested to connect/speak with someone ? I've noticed that if we already asked for an agent name in a past call and we ask again in another call, if the agent is the same sometimes it won't trigger the transfer_call tool, seems like it saves the past call result, is this cache or I am doing something wrong ?

@alep your system prompt token length is 8k which is not recommended instead keep it under 3k tokens length. https://cdn.discordapp.com/attachments/1343768614209323018/1345968206547910747/Screenshot_2025-03-03_at_9.23.17_AM.png?ex=67c67a13&is=67c52893&hm=9b9f9799ab3cac602aba93dbc4d6150d03c6786c12797ae8849fbceb8ef0df46&

Yes, we are passing some big info of previous calls. I will reduce it, in any case besides reducing it any recommendation for this type of data, should we add it in the knowledge base ?. Also the prompt size is relevant to the model we use ? For example, if we now change to model 4.5, will we be able to use a bigger prompt ?

How to check this

@alep bigger prompt may not produce good results instead keep it under 3k limit for now. For previous conversations history you can add it as another message with role type system in the messages array.