Hi Team, I wanted to working of VAPI a bit more. I am working on a project which has GPT4.1 Model, A system Prompt and few tools. This pipeline was first created in the Notebook and then moved to VAPI. Now both the environment is giving different results. Notebook one is more like using markdown, long texts etc.. but VAPI is short and concise and feels like more voice oriented. Question : Is VAPI Internally augmenting my System Prompt to make it more voice like or adding certain instructions which will make it different from notebook environment ?

VAPI does not automatically augment or rewrite your system prompt behind the scenes. According to the documentation, VAPI treats your prompt as "sacred" and avoids modifying it to prevent unintended changes in behavior. However, VAPI does apply real-time orchestration models (like filler injection and backchanneling) to make responses more natural for voice, which can make outputs feel different from a notebook environment. If you want your assistant to sound more voice-friendly, consider explicitly adjusting your prompt for brevity and conversational tone as shown in the [prompting guide](https://docs.vapi.ai/prompting-guide). Sources: - [How Vapi works](https://docs.vapi.ai/how-vapi-works) - [Prompting guide](https://docs.vapi.ai/prompting-guide)

How can i replicate it in Notebook. I want to do some bulk testing and then send the prompt to VAPI, but it seems the results I have seen during the testing would be different from what VAPI will eventually give me.

VAPI doesn’t modify your system prompt but adds real-time voice-friendly features like filler words and backchanneling, making responses shorter and more conversational compared to the notebook environment..

Does it provide these infromation to LLM or only apply this logic once the response has been recevied from the LLM as a post processing logic ?

VAPI does not internally change your system prompt to make it more “voice-oriented.” The prompt you provide is passed directly to the LLM. What actually happens: Pre-LLM processing - No voice-specific augmentation is applied - RAG documents are added only if you’ve set up a knowledge base - The prompt is formatted per the LLM provider’s API (OpenAI, Anthropic, etc.) Post-LLM processing After the model responds, output is lightly reformatted for TTS: - Markdown symbols (

_

,

*

, \`\`\` ) are removed - Numbers converted to words for pronunciation - Phone numbers split into digits - Email addresses spelled out (“user at domain dot com”) - Punctuation adjusted (newlines → periods, colons → periods, etc.) Pipeline: Your system prompt → LLM (unchanged) → Model response → Voice formatting → TTS