How to search for images with a similarity greater than 0.9?

[bbhxwl]

How to search for images with a similarity greater than 0.9?
Which is more suitable for searching image similarity between Cosine and L2?

[yhmo]

"Range Search" is exactly the feature for this purpose: https://milvus.io/docs/range-search.md

res = client.search(
    collection_name="my_collection",
    data=[query_vector],
    limit=100,
    search_params={
        "params": {
            "radius": 0.9,
        }
    }
)

For Cosine metric, the distance range is [-1.0, 1.0]. distance=1.0 is the most similar item, you can set "radius=0.9". Formula of Cosine: https://milvus.io/docs/metric.md?tab=floating#Cosine-similarity-
For L2 metric, the distance range is [0.0, +∞]. distance=0.0 is the most similar item, you can set "radius=0.1". Formula of L2: https://milvus.io/docs/metric.md?tab=floating#Euclidean-distance-L2

Both are OK for image similarity search.

[bbhxwl]

https://docs.zilliz.com.cn/reference/restful/search
I am using params radius range_filter which does not work

[bbhxwl]

{
"params":{
"radius":0.9,
"range_filter":1.0
}
}

[yhmo]

Use this script to create a collection:

import random

from pymilvus import (
MilvusClient,DataType,
)


client = MilvusClient(uri="http://localhost:19530")

collection_name = "AAA"
dim = 4

def create_coll():
    client.drop_collection(collection_name=collection_name)

    schema = MilvusClient.create_schema(auto_id=False)
    schema.enable_dynamic_field = False
    schema.add_field(field_name="id", datatype=DataType.INT64, is_primary=True)
    schema.add_field(field_name="vector", datatype=DataType.FLOAT_VECTOR, dim=dim)

    client.create_collection(
        collection_name=collection_name,
        schema=schema,
    )
    index_params = client.prepare_index_params()
    index_params.add_index(field_name="vector", index_type="IVF_FLAT", metric_type="L2")
    client.create_index(collection_name, index_params)
    client.load_collection(collection_name=collection_name)
    print(collection_name, "created")

create_coll()
client.load_collection(collection_name=collection_name)


client.insert(collection_name=collection_name, data={"id":1, "vector": [1,1,1,1]})
client.insert(collection_name=collection_name, data={"id":2, "vector": [2,2,2,2]})
client.insert(collection_name=collection_name, data={"id":3, "vector": [3,3,3,3]})
client.insert(collection_name=collection_name, data={"id":4, "vector": [4,4,4,4]})
client.insert(collection_name=collection_name, data={"id":5, "vector": [5,5,5,5]})
client.flush(collection_name=collection_name)
print("data inserted")


res = client.search(collection_name=collection_name,
                    data=[[2,2,2,2]],
                    limit=10,
                    anns_field="vector",
                    search_params={
                            "params": {"radius": 2}
                        }
                    )
# Expect only one item returned
# {'id': 2, 'distance': 0.0, 'entity': {}}
for r in res:
    for k in r:
        print(k)

Use this restful V1 api to call search:

curl --request POST \
--url "http://127.0.0.1:19530/v1/vector/search" \
--header "Authorization: Bearer ${TOKEN}" \
--header "Content-Type: application/json" \
-d '{
    "collectionName": "AAA",
    "vector": [2,2,2,2],
    "annsField": "vector",
    "limit": 10,
    "params": {
        "radius": 2
    }
}'

Only one item returned, expected:

RESTful V1 will be deprecated.
RESTful V2 is recommended:

curl --request POST \
--url "http://127.0.0.1:19530/v2/vectordb/entities/search" \
--header "Authorization: Bearer ${TOKEN}" \
--header "Content-Type: application/json" \
-d '{
    "collectionName": "AAA",
    "data": [
        [2,2,2,2]
    ],
    "annsField": "vector",
    "limit": 10,
    "searchParams": {
        "params": {
            "radius": 2
        }
    }
}'