rageval - v0.1.1
    Preparing search index...

    Variable contextRelevanceConst

    contextRelevance: Metric = ...

    Context Relevance — measures whether the retrieved context is relevant to the question being asked (retriever quality signal).

    Score 1.0 = all retrieved chunks are highly relevant and directly useful. Score 0.0 = retrieved chunks are entirely off-topic / irrelevant to the question.

    When to use: Use contextRelevance to diagnose retrieval quality issues. A low score typically indicates that the embedding model, chunking strategy, or similarity threshold is not filtering out irrelevant chunks well enough.

    Difference from contextPrecision: Both measure retrieval quality, but from different angles. contextRelevance makes a holistic judgment ("is this context useful overall?") while contextPrecision computes an explicit ratio ("what fraction of chunks are relevant?"). Use both for a complete retrieval picture.

    Score interpretation (5-point scale):

    • 1.0: All chunks are directly relevant — excellent retriever precision
    • 0.75: Most chunks are relevant; one or two contain minor tangential content
    • 0.5: Mixed — roughly half the retrieved content is relevant to the question
    • 0.25: Most retrieved content is off-topic; only minor relevant signals
    • 0.0: Entirely irrelevant — retriever is fetching the wrong documents completely

    Uses LLM-as-judge pattern — see arXiv:2306.05685 (RAGAS paper).