We recently had our Assistant share incorrect information to a customer which was our first intentions to avoid here. It will also seem impossible for us to monitor this kind of this on a large scale but sifting through every call to ensure it provides correct information. We want to keep our Assistant conversational and "less robotic" and from our understanding, dropping the Model Temperature to 0 will cause make it less conversational and sound more robotic. Is there a way to ensure Assistant only shares specific information strictly from KB and Website only but keep it conversational? CALL ID: eeb275bc-a84d-4110-9ceb-b46901a16720 What went wrong: 00:09 Assistant said there are complimentary wheelchairs available (when there is no information provided about this in KB or website) What we expect to happen: Assistant to never share incorrect information and to only share information that is provided on KB and website.

I understand your concern about keeping your Assistant both accurate and conversational. Here's how you can ensure it only shares information from your KB and website while maintaining a natural tone: ### Ensuring Information Accuracy 1. Configure Your Knowledge Base Properly - Upload your files (Markdown, PDF, text, Word) to the Vapi Knowledge Base [through the platform or API](https://docs.vapi.ai/knowledge-base) - Create a custom knowledge base with appropriate search settings:

json
     {
       "searchPlan": {
         "searchType": "hybrid",
         "scoreThreshold": 0.3
       }
     }

2. Use a Clear System Prompt - Add this instruction to your assistant's system message:

You are a smart assistant who responds to user queries using the information from your KnowledgeBase. If the answer is not in the KnowledgeBase say 'I do not know'.

- This explicitly tells the model to only use KB information and admit when it doesn't know something ### Maintaining Conversational Tone 1. Balanced Temperature Setting - Set temperature between 0.1-0.3 (not 0) for a balance of consistency and natural conversation - This range helps maintain predictability while allowing some conversational flexibility 2. Utilize Filler Injection - Vapi's filler injection model adds natural speech elements like "umm" and "ahh" to make responses sound more human 3. Adjust Interaction Timing - Configure

numWordsToInterruptAssistant

and

maxDurationSeconds

(e.g., 5 words and 10 seconds) to create natural conversation flow ### Monitoring and Improvement - Use tools like Langfuse to monitor your assistant's responses and identify inaccuracies - Regularly review and update your KB to ensure comprehensive coverage of topics your users might ask about These settings should help your Assistant stay strictly within the bounds of your knowledge base while maintaining a natural, conversational tone. Sources:

@firefly_0206 The message.txt contains the content collected from KB and the model’s generated output. This should clarify its origin. Let me know if you need further assistance or have more questions. https://cdn.discordapp.com/attachments/1347536243931480156/1347736555162304522/message.txt?ex=67cce8fa&is=67cb977a&hm=1cf87516902e152cc231b57fe67feebdd9969362e8df23f1c5508ca96f92e742& https://cdn.discordapp.com/attachments/1347536243931480156/1347736555703238747/Screenshot_2025-03-08_at_6.33.44_AM.png?ex=67cce8fa&is=67cb977a&hm=50073d5fc58792f9e17bb53b59cea4bb3db4a3954f90c3f0c1fdcc154944258c&