Facing issues with Realtime API when using it to implement AI Interviewer

API Bugs
1

I am currently developing an AI Interviewer utilizing OpenAI’s Realtime API. The system is designed to conduct structured interviews by adhering to a predefined set of guidelines and questions. Despite providing detailed instructions, I am encountering the following issues:

  1. Non-compliance with Instructions: The AI occasionally deviates from the specified guidelines, leading to inconsistencies in the interview process.

  2. Incomplete Question Coverage: The AI sometimes concludes the interview prematurely, skipping some questions from the provided list.

I am seeking advice on the following:

  • Enhancing Instruction Compliance: Strategies to improve the AI’s adherence to provided guidelines.

  • Ensuring Complete Question Coverage: Methods to prevent the AI from skipping questions and concluding prematurely.

1 Like
2

Hey!

You might have to implement RAG via tool use for this.
Or tool use that first returns a question. On answering, it calls the tool again to return the next question and so on until you’re done with the interview.
Don’t rely on systemprompt alone!

Cheers. :hugs:

3

Hey!! @j.wischnat

Thanks for the reply. Could you provide any resources that would help me implement this solution.

4

Flowise is a pretty cool opensource software that allows you to make low-code LLM applications in no-time. It’s a perfect starting point. They also have an option to host it for you but for testing purposes, I’d recommend installing it locally.

If you want to implement just tool use, check the documentation of the OpenAI API. :hugs: