Evaluating RAG with LLM as a Judge

Large Language Models (LLMs) are rapidly becoming essential tools for creating widely-used applications. But making sure these models perform as expected is much easier said than done. Evaluating LLM systems isn't just about verifying the outputs are coherent, but also about making sure the answers are relevant and meet the necessary requirements.

Retrieval-Augmented Generation (RAG) systems have become a popular way to boost LLM capabilities. By pairing an LLM with a data retrieval system, LLMs can generate responses that are not only coherent but also grounded in relevant and current information. This helps cut down on moments when the model sounds confident but may actually be hallucinating.

However, evaluating whether these RAG systems are performant isn't straightforward. It's not just about whether the output generated by the LLMs sounds correct, it's also about verifying at the source if the retrieved information is relevant and accurate. Traditional methods often miss these nuances, making a comprehensive evaluation framework essential.

Custom structured outputs is a recent feature launched in our API which allows you to control the output format of LLMs and give values in a structured machine readable way so you have greater reliability.

Structured outputs sets a perfect framework to define new and reusable criteria and enforce a format that is then easily machine readable for evaluation.

Interested in more custom work with the Mistral AI team? Contact us for solution support

Taking a step back, there are many domains that face this same problem where there may be a lack of clear quantitative metrics or evaluation data to measure performance against and the measures of success may be more qualitative and nuanced.

In scenarios in which evaluations must be run at scale and there is a lack of human evaluators, “LLM As A Judge” has become a popular solution to evaluate the answers of LLM systems.

The way an “LLM As A Judge” typically works is to create a “judge LLM” that given the answer of a “generator LLM” is instructed to grade that answer based on a given scale. This scale can be numerical (e.g a 1-10 or 0-3 scale), binary (e.g True or False) or qualitative (e.g “Excellent”, “Good”, “Okay”, “Bad”). This grade can then be averaged across all questions of an evaluation dataset for an overall weighted score.