Building Confidence: A Case Study in How to Create Confidence Scores for GenAI Applications

This post details the evaluation and implementation of confidence scoring mechanisms for GenAI applications in a financial automation use case (invoice parsing). It compares three approaches: calibrator models, logprobs, and majority voting. Majority voting, particularly with weighted models and Platt scaling calibration, demonstrated a strong positive correlation with accuracy and was chosen for production. Challenges with long text fields were addressed by breaking them down, and granularity was improved using a permutation approach with multiple prompts per model. The post provides concrete examples of accuracy vs. confidence plots for different methods and fields, and discusses the trade-offs of ensemble size and weighting strategies.