MetricType

This enum specifies the distance or similarity metric used by vector indexes and searches. Pass a MetricType value to IndexParam when building an index, or to a search request. The valid choices depend on the vector field’s data type.

pub enum MetricType {
    Default,
    L2,
    Ip,
    Cosine,
    Hamming,
    Jaccard,
    MhJaccard,
    Bm25,
    MaxSimCosine,
    MaxSimIp,
    MaxSimL2,
    MaxSimJaccard,
    MaxSimHamming,
}

VARIANTS:

• Default

  Lets the Milvus server determine the metric type. This is the default.

• L2

  Euclidean distance; Milvus returns the squared value sum((a_i - b_i)^2). Smaller is more similar.

• Ip

  Inner product, sum(a_i * b_i). Larger is more similar.

• Cosine

  Cosine similarity, cos(a, b). Larger is more similar.

• Hamming

  Hamming distance, the number of positions at which two binary vectors differ. Smaller is more similar.

• Jaccard

  Jaccard distance, 1 - |A ∩ B| / |A ∪ B|. Smaller is more similar.

• MhJaccard

  MinHash Jaccard distance over binary MinHash signatures, 1 - estimated similarity. Smaller is more similar.

• Bm25

  BM25 relevance score for full-text search over sparse vectors.

• MaxSimCosine

  MaxSim Cosine, used to search embedding lists stored in array-of-structs vector fields.

• MaxSimIp

  MaxSim Inner Product, used to search embedding lists stored in array-of-structs vector fields.

• MaxSimL2

  MaxSim L2, used to search embedding lists stored in array-of-structs vector fields.

• MaxSimJaccard

  MaxSim Jaccard, used to search embedding lists stored in array-of-structs vector fields.

• MaxSimHamming

  MaxSim Hamming, used to search embedding lists stored in array-of-structs vector fields.

Example

let index = IndexParam::new()
    .field_name("embedding")
    .index_type(IndexType::Hnsw)
    .metric_type(MetricType::Cosine)
    .extra_params(HashMap::from([("M".to_owned(), "16".to_owned())]));