Structured Output Confidence Score

API
1

I am using Structured Output and passing a JSON schema to the Chat Completion API. I am looking for the LLM to provide a “confidence score” for each object in an array returned in the JSON. I have defined this element this way:

                        gen.writeFieldName('confidenceScore');
                        gen.writeStartObject();
                            gen.writeStringField('type', 'number');
                            gen.writeStringField('description', 'Your confidence (0-100) that this load data is correct'); 
                        gen.writeEndObject();

I am getting way overconfident values back from the model. Sometimes “100” when I provide gibberish as the user input.

How can I craft a better description or definition in the JSON so I get a better confidence score?

Thanks for your help!!

1 Like
2

We had the same requirement. So, we created a small open source project to calculate confidence score:
NPM - @promptrepo/score - npm
Source - GitHub - ManiDoraisamy/promptrepo-score: Calculate confidence score for structured output generated by LLMs like OpenAI