Evaluating RAG with LLM as a Judge

Mistral AI has introduced a methodology for evaluating Retrieval-Augmented Generation (RAG) systems using an "LLM as a Judge" approach. This framework allows developers to scale evaluation by using a judge LLM to grade the outputs of a generator LLM based on specific, machine-readable criteria.

LLM as a Judge for Scalable Evaluation

LLM as a Judge is a solution for evaluating LLM systems at scale when human evaluators are unavailable. In this setup, a "judge LLM" is instructed to grade the response of a "generator LLM" using a predefined scale, which can be numerical (e.g., 0-3 or 1-10), binary (True/False), or qualitative (e.g., "Excellent" to "Bad").

To implement this, developers create custom instruction prompts that define the grading criteria, such as whether a response is accurate, relevant, or grounded. Mistral's models can serve as both the generator and the judge components of these systems.

The RAG Triad Evaluation Framework

To move beyond simple coherence checks, Mistral recommends the RAG Triad, a holistic framework (introduced by TruLens and similar to RAGAS) that evaluates the reliability and contextual integrity of RAG systems. The RAG Triad assesses three specific metrics:

1. Context Relevance

Context relevance measures whether the retrieved documents align with the user's query. This ensures that the information provided to the generator LLM is appropriate and sufficient to produce a useful answer.

2. Groundedness

Groundedness verifies that the generated response is based accurately on the retrieved context. This metric is critical for reducing hallucinations by ensuring the output is factual and faithful to the source data.

3. Answer Relevance

Answer relevance assesses how well the final response addresses the user's original query, ensuring the output aligns with the user's intent and provides valuable insights.

Implementing Evaluation with Structured Outputs

Mistral AI leverages its "structured outputs" API feature to make the LLM-as-a-judge process more reliable and machine-readable. By enforcing a specific schema for the judge's response, developers can ensure that the evaluation results are consistent and easy to process programmatically.

Using Pydantic models, Mistral demonstrates how to define a RAGEvaluation schema that requires the judge LLM to provide both a step-by-step reasoning explanation and a score for each of the three Triad metrics: context relevance, answer relevance, and groundedness.

Conclusion

Combining the RAG Triad framework with structured outputs allows for a robust, end-to-end evaluation of RAG systems. This approach assesses both the retrieval and generation steps, ensuring that the final interaction between the human, the AI, and the knowledge base is accurate and trustworthy.

Sources

Related

  • Dispatch
    Mistral Agentic Search Release

    Mistral AI has introduced Agentic Search, a multi-step retrieval loop that enables AI models to navigate complex, dense documents and improve accuracy on financial and government data benchmarks.

  • Project
    hhy-huang/HiRAG

    HiRAG is a Retrieval-Augmented Generation framework that uses hierarchical knowledge to improve the quality and comprehensiveness of LLM responses compared to standard RAG.

  • Project
    GraphRAG-Bench/GraphRAG-Benchmark

    A comprehensive benchmark for evaluating Graph Retrieval-Augmented Generation (GraphRAG) models to determine when they outperform traditional RAG across various reasoning and retrieval tasks.

  • Project
    ray-project/llm-applications

    A comprehensive guide and implementation for building, scaling, and productionizing RAG-based LLM applications using the Ray framework.

  • Project
    NovaSearch-Team/RAG-Retrieval

    A comprehensive toolkit for training, fine-tuning, and distilling RAG retrieval models, including embedding models, late interaction models, and rerankers.