How to prevent hallucinations using OpenAI's gpt-3.5-turbo API?

I am attempting to make a personal recruiting assistant tool to help me automate the process of answering questions for job applications like "Why do you want to work here?" or "Write a cover letter saying how your experience relates to this position.".

I am using my resume and a Q&A section as the context for the API call and the question as the prompt. Over all, the system works well but it hallucinates my experience often. For example, it will state I have more years of experience than is actually outlined in my resume or it will say I am an expert in technologies that I have never used or listed in my context.

I have tried explicitly prompting the model through this code but it has not been successful:

def prepare_context(self):
        context = (
            "Your role in this instance is "
            "to act as a recruiting assistant and help answer interview questions. "
            "Your responses should be based on the information provided in the resume and "
            "previous Q&As without hallucinating additional information or experience.\n\n"
            "My Resume:\n\n"
            f"{self.resume}\n\n"
            "My Previously answered questions:\n"
        )

        for q, a in self.qa_list:
            context += f"Q: {q}\nA: {a}\n"

        context += (
            "\nWhen answering the following questions, remember to answer "
            "in the first person as if you were the job applicant. The responses "
            "should be concise, truthful, and based solely on the given information. "
            "Do not create or infer any additional experiences or skills that are "
            "not explicitly mentioned in the resume or previous Q&As. "
            "Do not generate any information that is not included in the resume or previous questions and answers. "
            "Remember, you are answering as me "
            "so answer all further questions from my perspective.\n"
        )

        return context

You can’t

Unless I missed a headline there’s currently no way to prevent hallucinations with any large language model. If you’re going to do this you either need to be ok with random errors no matter what they are, or you’ll need to double check all of the AI’s output every time. You could create a second-stage automated error checker, but no matter how you implement it some errors will still get through, leaving you with the choice of either accepting the errors or reviewing everything.

In general I would only use generative AI to automate low risk tasks. Applying for jobs isn’t something I’d consider low-risk as “hallucinations” could have severe consequences for your professional reputation and future employability. Just my two cents.

There is a way to eliminate LLM hallucination with a Retrieval Augmented Generation process. Pinecone and langchain are not sufficient on their own. Pre-processing of content is necessary.EyelevelAI has develop a solution called GroundX.You can request API access from GroundX.ai