This may be helpful to anyone having challenges with using both a knowledge base and tool calling. I've been struggling with getting Vapi to call tools reliably, such as "endCall" when the user has no further questions. After banging my head against the wall for a while, I decided to run the LLM through my own proxy to OpenAI to see the raw messages and it turns out that it inserts additional context into all user messages, whether they actually have a question or not, which ends up confusing the LLM. Example dialogue: * Vapi: Hi, do you have any questions? * User: Yes, here's my question... (Vapi inserts context from KB) * Vapi: Here's your answer, any more questions? * User: No (Vapi again inserts context from KB) * Vapi: Here's a continuation of previous answer (instead of calling endCall) Not sure what my solution will be yet, but it seems I'll have to setup my own KB as a tool. Vapi may also want to reconsider their own implementation of the KB search if my findings are correct. https://cdn.discordapp.com/attachments/1211484481848873020/1294368147893125211/image.png?ex=670ac1c6&is=67097046&hm=542e6e34d9e09558eeb0893ba5503c8fb234cdc65340cefac60907fec4e97590&