Calculate Distance Between Vectors

This topic describes how to calculate distance between vectors with Milvus.

Milvus searches most similar vectors based on the distance calculation of vectors. Vice versa, you can use Milvus to calculate the distance between vectors using distance metrics that suit specific scenario. See Similarity Metrics for more information.

The following example simulates the scenarios when you want to calculate the distance between vectors in the collection and some other vectors.

Prepare vectors

Prepare the vectors used for calculation.

Vectors to be calculated must agree in vector type and dimension.

Python Java GO Node.js CLI Curl

vectors_left = {
    "ids": [0, 1], 
    "collection": "book", 
    "partition": "_default", 
    "field": "book_intro"
}
import random
external_vectors = [[random.random() for _ in range(2)] for _ in range(4)]
vectors_right = {"float_vectors": external_vectors}

// Node User Guide will be ready soon.

// GO User Guide will be ready soon.

// Java User Guide will be ready soon.

// CLI User Guide will be ready soon.

vectors_left='{
  "dim": 2,
  "ids": {
    "id_array": [1,2],
    "collection_name": "book",
    "partition_names": ["_default"],
    "field_name": "book_intro"
  }
}'
vectors_right='{
  "dim": 2,
  "vectors": [1,2,3,4,5,6,7,8] # The numbers in the list will be automatically split into four vectors. 
}'

Parameter	Description
`vectors_left` and `vectors_right`	Vectors on the left and right side of the operator. Dict type that can be represented as `{"ids": [primary_key_1, primary_key_2, ... primary_key_n], "collection": "collection_name", "partition": "partition_name", "field": "vector_field_name"}`, `{"float_vectors": [[1.0, 2.0], [3.0, 4.0], ... [9.0, 10.0]]}`, or `{"bin_vectors": [b'', b'N', ... b'Ê']}`.
`ids`	List of primary key of entities that in the collection.
`collection`	Name of the collection that holds the entities.
`partition`	Name of the partition that holds the entities.
`field`	Name of the vector field in the collection.
`float_vectors` or `bin_vectors`	Type of the vectors.

Parameter	Description	Option
`dim`	Dimension of the vector.	N/A
`id_array`	List of the primary keys of entities in the collection.	N/A
`collection_name`	Name of the collection that holds the entities.	N/A
`partition_names`	Names of the partitions that hold the entities.	N/A
`field_name`	Name of the vector field in the collection.	N/A
`vectors`	Temporarily only floating-point vectors are supported.	N/A

Prepare calculation parameters

Specify the parameters used for the calculation.

Python Java GO Node.js CLI Curl

params = {
    "metric": "IP", 
    "dim": 2
}

// Node User Guide will be ready soon.

// GO User Guide will be ready soon.

// Java User Guide will be ready soon.

// CLI User Guide will be ready soon.

params='[
  {"key": "metric", "value": "IP"}
]'

Parameter Description Option

params Calculation parameters. N/A

Parameter	Description	Option
`params`	Calculation parameters.	N/A
`metric`	Metric types used for calculation.	For floating-point vectors: `L2` (Euclidean distance) `IP` (Inner product) For binary vectors: `JACCARD` (Jaccard distance) `TANIMOTO` (Tanimoto distance) `HAMMING` (Hamming distance) `SUPERSTRUCTURE` (Superstructure) `SUBSTRUCTURE` (Substructure)
`dim`	Dimension of the vector.	N/A

metric

Metric types used for calculation.

For floating-point vectors:

L2 (Euclidean distance)
IP (Inner product)

For binary vectors:

JACCARD (Jaccard distance)
TANIMOTO (Tanimoto distance)
HAMMING (Hamming distance)
SUPERSTRUCTURE (Superstructure)
SUBSTRUCTURE (Substructure)

dim Dimension of the vector. N/A

Parameter	Description	Option
`metric`	Metric types used for calculation.	For floating-point vectors: `L2` (Euclidean distance) `IP` (Inner product)

(Optional) Load collection

If you calculate with the vectors in a collection in Milvus, you must load the collection to memory first.

Python Java GO Node.js CLI Curl

from pymilvus import Collection
collection = Collection("book")      # Get an existing collection.
collection.load()

await milvusClient.collectionManager.loadCollection({
  collection_name: "book",
});

err := milvusClient.LoadCollection(
  context.Background(),   // ctx
  "book",                 // CollectionName
  false                   // async
)
if err != nil {
  log.Fatal("failed to load collection:", err.Error())
}

milvusClient.loadCollection(
  LoadCollectionParam.newBuilder()
    .withCollectionName("book")
    .build()
);

load -c book

curl -X 'POST' \
  'http://localhost:9091/api/v1/collection/load' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "collection_name": "book"
  }'

Calculate vector distance

Calculate the distance between vectors based on the vectors and parameters provided.

from pymilvus import utility
results = utility.calc_distance(
    vectors_left=vectors_left, 
    vectors_right=vectors_right, 
    params=params
)
print(results)

// Node User Guide will be ready soon.

// GO User Guide will be ready soon.

// Java User Guide will be ready soon.

// CLI User Guide will be ready soon.

curl -X 'GET' \
  'http://localhost:9091/api/v1/distance' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d "{
    \"op_left\": $vectors_left,
    \"op_right\": $vectors_right,
    \"params\": $params
  }"

Output:

{"status":{},"Array":{"FloatDist":{"data":[3,7,11,15,4,10,16,22]}}}

What's next

Learn more basic operations of Milvus: