@vivantel/virage-core
    Preparing search index...

    Interface EmbeddingMetrics

    interface EmbeddingMetrics {
        avgCosineSimRandomPairs: number;
        intrinsicDimension: number;
        outlierFraction: number;
    }
    Index
    avgCosineSimRandomPairs: number

    Mean cosine similarity between randomly sampled pairs of embeddings. Should be close to 0 for a well-spread embedding space.

    intrinsicDimension: number

    Estimated intrinsic dimensionality via the TWO-NN estimator. A value much lower than dimensions suggests the embedding space is under-utilised for this corpus.

    outlierFraction: number

    Fraction of embeddings whose L2-norm z-score exceeds 2.5. Non-zero values may indicate degenerate or truncated inputs.