milvus-logo
LFAI
Home
  • Benutzerhandbuch

Single-Vector Search

Once you have inserted your data, the next step is to perform similarity searches on your collection in Milvus.

Milvus allows you to conduct two types of searches, depending on the number of vector fields in your collection:

  • Single-vector search: If your collection has only one vector field, use the search() method to find the most similar entities. This method compares your query vector with the existing vectors in your collection and returns the IDs of the closest matches along with the distances between them. Optionally, it can also return the vector values and metadata of the results.
  • Hybrid search: For collections with two or more vector fields, use the hybrid_search() method. This method performs multiple Approximate Nearest Neighbor (ANN) search requests and combines the results to return the most relevant matches after reranking.

This guide focuses on how to perform a single-vector search in Milvus. For details on hybrid search, refer to Hybrid search.

Overview

There are a variety of search types to meet different requirements:

  • Basic search: Includes single-vector search, bulk-vector search, partition search, and search with specified output fields.

  • Filtered search: Applies filtering criteria based on scalar fields to refine search results.

  • Range search: Finds vectors within a specific distance range from the query vector.

  • Grouping search: Groups search results based on a specific field to ensure diversity in the results.

Preparations

The code snippet below repurposes the existing code to establish a connection to Milvus and quickly set up a collection.

Python Java Node.js 
from pymilvus import MilvusClient
import random

# 1. Set up a Milvus client
client = MilvusClient(
    uri="http://localhost:19530"
)

# 2. Create a collection
client.create_collection(
    collection_name="quick_setup",
    dimension=5,
    metric_type="IP"
)

# 3. Insert randomly generated vectors 
colors = ["green", "blue", "yellow", "red", "black", "white", "purple", "pink", "orange", "brown", "grey"]
data = []

for i in range(1000):
    current_color = random.choice(colors)
    data.append({
        "id": i,
        "vector": [ random.uniform(-1, 1) for _ in range(5) ],
        "color": current_color,
        "color_tag": f"{current_color}_{str(random.randint(1000, 9999))}"
    })

res = client.insert(
    collection_name="quick_setup",
    data=data
)

print(res)

# Output
#
# {
#     "insert_count": 1000,
#     "ids": [
#         0,
#         1,
#         2,
#         3,
#         4,
#         5,
#         6,
#         7,
#         8,
#         9,
#         "(990 more items hidden)"
#     ]
# }

# 6.1 Create partitions 
client.create_partition(
    collection_name="quick_setup",
    partition_name="red"
)

client.create_partition(
    collection_name="quick_setup",
    partition_name="blue"
)

# 6.1 Insert data into partitions
red_data = [ {"id": i, "vector": [ random.uniform(-1, 1) for _ in range(5) ], "color": "red", "color_tag": f"red_{str(random.randint(1000, 9999))}" } for i in range(500) ]
blue_data = [ {"id": i, "vector": [ random.uniform(-1, 1) for _ in range(5) ], "color": "blue", "color_tag": f"blue_{str(random.randint(1000, 9999))}" } for i in range(500) ]

res = client.insert(
    collection_name="quick_setup",
    data=red_data,
    partition_name="red"
)

print(res)

# Output
#
# {
#     "insert_count": 500,
#     "ids": [
#         0,
#         1,
#         2,
#         3,
#         4,
#         5,
#         6,
#         7,
#         8,
#         9,
#         "(490 more items hidden)"
#     ]
# }

res = client.insert(
    collection_name="quick_setup",
    data=blue_data,
    partition_name="blue"
)

print(res)

# Output
#
# {
#     "insert_count": 500,
#     "ids": [
#         0,
#         1,
#         2,
#         3,
#         4,
#         5,
#         6,
#         7,
#         8,
#         9,
#         "(490 more items hidden)"
#     ]
# }

import com.google.gson.Gson;
import com.google.gson.JsonObject;
import io.milvus.v2.client.MilvusClientV2;
import io.milvus.v2.client.ConnectConfig;
import io.milvus.v2.service.collection.request.CreateCollectionReq;
import io.milvus.v2.service.collection.request.GetLoadStateReq;
import io.milvus.v2.service.partition.request.CreatePartitionReq;
import io.milvus.v2.service.vector.request.InsertReq;
import io.milvus.v2.service.vector.response.InsertResp;

import java.util.*;

String CLUSTER_ENDPOINT = "http://localhost:19530";

// 1. Connect to Milvus server
ConnectConfig connectConfig = ConnectConfig.builder()
    .uri(CLUSTER_ENDPOINT)
    .build();

MilvusClientV2 client = new MilvusClientV2(connectConfig);  

// 2. Create a collection in quick setup mode
CreateCollectionReq quickSetupReq = CreateCollectionReq.builder()
        .collectionName("quick_setup")
        .dimension(5)
        .metricType("IP")
        .build();

client.createCollection(quickSetupReq);

GetLoadStateReq loadStateReq = GetLoadStateReq.builder()
        .collectionName("quick_setup")
        .build();

boolean state = client.getLoadState(loadStateReq);

System.out.println(state);

// Output:
// true

// 3. Insert randomly generated vectors into the collection
List<String> colors = Arrays.asList("green", "blue", "yellow", "red", "black", "white", "purple", "pink", "orange", "brown", "grey");
List<JsonObject> data = new ArrayList<>();
Gson gson = new Gson();
for (int i=0; i<1000; i++) {
    Random rand = new Random();
    String current_color = colors.get(rand.nextInt(colors.size()-1));
    JsonObject row = new JsonObject();
    row.addProperty("id", (long) i);
    row.add("vector", gson.toJsonTree(Arrays.asList(rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat())));
    row.addProperty("color_tag", current_color + "_" + String.valueOf(rand.nextInt(8999) + 1000));
    data.add(row);
}

InsertReq insertReq = InsertReq.builder()
        .collectionName("quick_setup")
        .data(data)
        .build();

InsertResp insertResp = client.insert(insertReq);

System.out.println(insertResp.getInsertCnt());

// Output:
// 1000

// 6.1. Create a partition
CreatePartitionReq partitionReq = CreatePartitionReq.builder()
        .collectionName("quick_setup")
        .partitionName("red")
        .build();

client.createPartition(partitionReq);

partitionReq = CreatePartitionReq.builder()
        .collectionName("quick_setup")
        .partitionName("blue")
        .build();

client.createPartition(partitionReq);

// 6.2 Insert data into the partition
data = new ArrayList<>();

for (int i=1000; i<1500; i++) {
    Random rand = new Random();
    String current_color = "red";
    JsonObject row = new JsonObject();
    row.addProperty("id", (long) i);
    row.add("vector", gson.toJsonTree(Arrays.asList(rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat())));
    row.addProperty("color", current_color);
    row.addProperty("color_tag", current_color + "_" + String.valueOf(rand.nextInt(8999) + 1000));
    data.add(row);
}

insertReq = InsertReq.builder()
        .collectionName("quick_setup")
        .data(data)
        .partitionName("red")
        .build();

insertResp = client.insert(insertReq);

System.out.println(insertResp.getInsertCnt());

// Output:
// 500

data = new ArrayList<>();

for (int i=1500; i<2000; i++) {
    Random rand = new Random();
    String current_color = "blue";
    JsonObject row = new JsonObject();
    row.addProperty("id", (long) i);
    row.add("vector", gson.toJsonTree(Arrays.asList(rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat(), rand.nextFloat())));
    row.addProperty("color", current_color);
    row.addProperty("color_tag", current_color + "_" + String.valueOf(rand.nextInt(8999) + 1000));
    data.add(row);
}

insertReq = InsertReq.builder()
        .collectionName("quick_setup")
        .data(data)
        .partitionName("blue")
        .build();

insertResp = client.insert(insertReq);

System.out.println(insertResp.getInsertCnt());

// Output:
// 500

const { MilvusClient, DataType, sleep } = require("@zilliz/milvus2-sdk-node")

const address = "http://localhost:19530"

// 1. Set up a Milvus Client
client = new MilvusClient({address});

// 2. Create a collection in quick setup mode
await client.createCollection({
    collection_name: "quick_setup",
    dimension: 5,
    metric_type: "IP"
});  

// 3. Insert randomly generated vectors
const colors = ["green", "blue", "yellow", "red", "black", "white", "purple", "pink", "orange", "brown", "grey"]
data = []

for (let i = 0; i < 1000; i++) {
    current_color = colors[Math.floor(Math.random() * colors.length)]
    data.push({
        id: i,
        vector: [Math.random(), Math.random(), Math.random(), Math.random(), Math.random()],
        color: current_color,
        color_tag: `${current_color}_${Math.floor(Math.random() * 8999) + 1000}`
    })
}

var res = await client.insert({
    collection_name: "quick_setup",
    data: data
})

console.log(res.insert_cnt)

// Output
// 
// 1000
// 

await client.createPartition({
    collection_name: "quick_setup",
    partition_name: "red"
})

await client.createPartition({
    collection_name: "quick_setup",
    partition_name: "blue"
})

// 6.1 Insert data into partitions
var red_data = []
var blue_data = []

for (let i = 1000; i < 1500; i++) {
    red_data.push({
        id: i,
        vector: [Math.random(), Math.random(), Math.random(), Math.random(), Math.random()],
        color: "red",
        color_tag: `red_${Math.floor(Math.random() * 8999) + 1000}`
    })
}

for (let i = 1500; i < 2000; i++) {
    blue_data.push({
        id: i,
        vector: [Math.random(), Math.random(), Math.random(), Math.random(), Math.random()],
        color: "blue",
        color_tag: `blue_${Math.floor(Math.random() * 8999) + 1000}`
    })
}

res = await client.insert({
    collection_name: "quick_setup",
    data: red_data,
    partition_name: "red"
})

console.log(res.insert_cnt)

// Output
// 
// 500
// 

res = await client.insert({
    collection_name: "quick_setup",
    data: blue_data,
    partition_name: "blue"
})

console.log(res.insert_cnt)

// Output
// 
// 500
//

When sending a search request, you can provide one or more vector values representing your query embeddings and a limit value indicating the number of results to return.

Depending on your data and your query vector, you may get fewer than limit results. This happens when limit is larger than the number of possible matching vectors for your query.

Single-vector search is the simplest form of search operations in Milvus, designed to find the most similar vectors to a given query vector.

To perform a single-vector search, specify the target collection name, the query vector, and the desired number of results (limit). This operation returns a result set comprising the most similar vectors, their IDs, and distances from the query vector.

Here is an example of searching for the top 5 entities that are most similar to the query vector:

Python Java Node.js 
# Single vector search
res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    # Replace with your query vector
    data=[[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]],
    limit=5, # Max. number of search results to return
    search_params={"metric_type": "IP", "params": {}} # Search parameters
)

# Convert the output to a formatted JSON string
result = json.dumps(res, indent=4)
print(result)

// 4. Single vector search
List<List<Float>> query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

SearchReq searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .topK(3) // The number of results to return
    .build();

SearchResp searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 4. Single vector search
var query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592],

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    limit: 3, // The number of results to return
})

console.log(res.results)
Parameter Description
collection_name The name of an existing collection.
data A list of vector embeddings.
Milvus searches for the most similar vector embeddings to the specified ones.
limit The total number of entities to return.
You can use this parameter in combination with offset in param to enable pagination.
The sum of this value and offset in param should be less than 16,384.
search_params The parameter settings specific to this operation.
  • metric_type: The metric type applied to this operation. This should be the same as the one used when you index the vector field specified above. Possible values are L2, IP, COSINE, JACCARD, HAMMING.
  • params: Additional parameters. For details, refer to search().
Parameter Description
collectionName The name of an existing collection.
data A list of vector embeddings.
Milvus searches for the most similar vector embeddings to the specified ones.
topK The number of records to return in the search result. This parameter uses the same syntax as the limit parameter, so you should only set one of them.
You can use this parameter in combination with offset in param to enable pagination.
The sum of this value and offset in param should be less than 16,384.
Parameter Description
collection_name The name of an existing collection.
data A list of vector embeddings.
Milvus searches for the most similar vector embeddings to the specified ones.
limit The total number of entities to return.
You can use this parameter in combination with offset in param to enable pagination.
The sum of this value and offset in param should be less than 16,384.

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 0,
            "distance": 1.4093276262283325,
            "entity": {}
        },
        {
            "id": 4,
            "distance": 0.9902134537696838,
            "entity": {}
        },
        {
            "id": 1,
            "distance": 0.8519943356513977,
            "entity": {}
        },
        {
            "id": 5,
            "distance": 0.7972343564033508,
            "entity": {}
        },
        {
            "id": 2,
            "distance": 0.5928734540939331,
            "entity": {}
        }
    ]
]

{"searchResults": [[
    {
        "score": 1.263043,
        "fields": {
            "vector": [
                0.9533119,
                0.02538395,
                0.76714665,
                0.35481733,
                0.9845762
            ],
            "id": 740
        }
    },
    {
        "score": 1.2377806,
        "fields": {
            "vector": [
                0.7411156,
                0.08687937,
                0.8254139,
                0.08370924,
                0.99095553
            ],
            "id": 640
        }
    },
    {
        "score": 1.1869997,
        "fields": {
            "vector": [
                0.87928146,
                0.05324632,
                0.6312755,
                0.28005534,
                0.9542448
            ],
            "id": 455
        }
    }
]]}

[
  { score: 1.7463608980178833, id: '854' },
  { score: 1.744946002960205, id: '425' },
  { score: 1.7258622646331787, id: '718' }
]

The output showcases the top 5 neighbors nearest to your query vector, including their unique IDs and the calculated distances.

A bulk-vector search extends the single-vector search concept by allowing multiple query vectors to be searched in a single request. This type of search is ideal for scenarios where you need to find similar vectors for a set of query vectors, significantly reducing the time and computational resources required.

In a bulk-vector search, you can include several query vectors in the data field. The system processes these vectors in parallel, returning a separate result set for each query vector, each set containing the closest matches found within the collection.

Here is an example of searching for two distinct sets of the most similar entities from two query vectors:

Python Java Node.js 
# Bulk-vector search
res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    data=[
        [0.19886812562848388, 0.06023560599112088, 0.6976963061752597, 0.2614474506242501, 0.838729485096104],
        [0.3172005263489739, 0.9719044792798428, -0.36981146090600725, -0.4860894583077995, 0.95791889146345]
    ], # Replace with your query vectors
    limit=2, # Max. number of search results to return
    search_params={"metric_type": "IP", "params": {}} # Search parameters
)

result = json.dumps(res, indent=4)
print(result)

// 5. Batch vector search
query_vectors = Arrays.asList(
    Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f),
    Arrays.asList(0.19886812562848388f, 0.06023560599112088f, 0.6976963061752597f, 0.2614474506242501f, 0.838729485096104f)
);

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .topK(2)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 5. Batch vector search
var query_vectors = [
    [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592],
    [0.19886812562848388, 0.06023560599112088, 0.6976963061752597, 0.2614474506242501, 0.838729485096104]
]

res = await client.search({
    collection_name: "quick_setup",
    data: query_vectors,
    limit: 2,
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 1,
            "distance": 1.3017789125442505,
            "entity": {}
        },
        {
            "id": 7,
            "distance": 1.2419954538345337,
            "entity": {}
        }
    ], # Result set 1
    [
        {
            "id": 3,
            "distance": 2.3358664512634277,
            "entity": {}
        },
        {
            "id": 8,
            "distance": 0.5642921924591064,
            "entity": {}
        }
    ] # Result set 2
]

// Two sets of vectors are returned as expected

{"searchResults": [
    [
        {
            "score": 1.263043,
            "fields": {
                "vector": [
                    0.9533119,
                    0.02538395,
                    0.76714665,
                    0.35481733,
                    0.9845762
                ],
                "id": 740
            }
        },
        {
            "score": 1.2377806,
            "fields": {
                "vector": [
                    0.7411156,
                    0.08687937,
                    0.8254139,
                    0.08370924,
                    0.99095553
                ],
                "id": 640
            }
        }
    ],
    [
        {
            "score": 1.8654699,
            "fields": {
                "vector": [
                    0.4671427,
                    0.8378432,
                    0.98844475,
                    0.82763994,
                    0.9729997
                ],
                "id": 638
            }
        },
        {
            "score": 1.8581753,
            "fields": {
                "vector": [
                    0.735541,
                    0.60140246,
                    0.86730254,
                    0.93152493,
                    0.98603314
                ],
                "id": 855
            }
        }
    ]
]}

[
  [
    { score: 2.3590476512908936, id: '854' },
    { score: 2.2896690368652344, id: '59' }
  [
    { score: 2.664059638977051, id: '59' },
    { score: 2.59483003616333, id: '854' }
  ]
]

The results include two sets of nearest neighbors, one for each query vector, showcasing the efficiency of bulk-vector searches in handling multiple query vectors at once.

Partition search narrows the scope of your search to a specific subset or partition of your collection. This is particularly useful for organized datasets where data is segmented into logical or categorical divisions, allowing for faster search operations by reducing the volume of data to scan.

To conduct a partition search, simply include the name of the target partition in partition_names of your search request. This specifies that the search operation only considers vectors within the specified partition.

Here is an example of searching for entities in red:

Python Java Node.js 
# 6.2 Search within a partition
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = client.search(
    collection_name="quick_setup",
    data=[query_vector],
    limit=5,
    search_params={"metric_type": "IP", "params": {"level": 1}},
    partition_names=["red"]
)

print(res)

// 6.3 Search within partitions
query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .partitionNames(Arrays.asList("red"))
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 6.2 Search within partitions
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    partition_names: ["red"],
    limit: 5,
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 16,
            "distance": 0.9200337529182434,
            "entity": {}
        },
        {
            "id": 14,
            "distance": 0.4505271911621094,
            "entity": {}
        },
        {
            "id": 15,
            "distance": 0.19924677908420563,
            "entity": {}
        },
        {
            "id": 17,
            "distance": 0.0075093843042850494,
            "entity": {}
        },
        {
            "id": 13,
            "distance": -0.14609718322753906,
            "entity": {}
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.1677284,
            "fields": {
                "vector": [
                    0.9986977,
                    0.17964739,
                    0.49086612,
                    0.23155272,
                    0.98438674
                ],
                "id": 1435
            }
        },
        {
            "score": 1.1476475,
            "fields": {
                "vector": [
                    0.6952647,
                    0.13417172,
                    0.91045254,
                    0.119336545,
                    0.9338931
                ],
                "id": 1291
            }
        },
        {
            "score": 1.0969629,
            "fields": {
                "vector": [
                    0.3363194,
                    0.028906643,
                    0.6675426,
                    0.030419827,
                    0.9735209
                ],
                "id": 1168
            }
        },
        {
            "score": 1.0741848,
            "fields": {
                "vector": [
                    0.9980543,
                    0.36063594,
                    0.66427994,
                    0.17359233,
                    0.94954175
                ],
                "id": 1164
            }
        },
        {
            "score": 1.0584627,
            "fields": {
                "vector": [
                    0.7187005,
                    0.12674773,
                    0.987718,
                    0.3110777,
                    0.86093885
                ],
                "id": 1085
            }
        }
    ],
    [
        {
            "score": 1.8030131,
            "fields": {
                "vector": [
                    0.59726167,
                    0.7054632,
                    0.9573117,
                    0.94529945,
                    0.8664103
                ],
                "id": 1203
            }
        },
        {
            "score": 1.7728865,
            "fields": {
                "vector": [
                    0.6672442,
                    0.60448086,
                    0.9325822,
                    0.80272985,
                    0.8861626
                ],
                "id": 1448
            }
        },
        {
            "score": 1.7536311,
            "fields": {
                "vector": [
                    0.59663296,
                    0.77831805,
                    0.8578314,
                    0.88818026,
                    0.9030075
                ],
                "id": 1010
            }
        },
        {
            "score": 1.7520742,
            "fields": {
                "vector": [
                    0.854198,
                    0.72294194,
                    0.9245805,
                    0.86126596,
                    0.7969224
                ],
                "id": 1219
            }
        },
        {
            "score": 1.7452049,
            "fields": {
                "vector": [
                    0.96419,
                    0.943535,
                    0.87611496,
                    0.8268136,
                    0.79786557
                ],
                "id": 1149
            }
        }
    ]
]}

[
  { score: 3.0258803367614746, id: '1201' },
  { score: 3.004319190979004, id: '1458' },
  { score: 2.880324363708496, id: '1187' },
  { score: 2.8246407508850098, id: '1347' },
  { score: 2.797295093536377, id: '1406' }
]

Then, search for entities in blue:

Python Java Node.js 
res = client.search(
    collection_name="quick_setup",
    data=[query_vector],
    limit=5,
    search_params={"metric_type": "IP", "params": {"level": 1}},
    partition_names=["blue"]
)

print(res)

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .partitionNames(Arrays.asList("blue"))
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    partition_names: ["blue"],
    limit: 5,
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 20,
            "distance": 2.363696813583374,
            "entity": {}
        },
        {
            "id": 26,
            "distance": 1.0665391683578491,
            "entity": {}
        },
        {
            "id": 23,
            "distance": 1.066049575805664,
            "entity": {}
        },
        {
            "id": 29,
            "distance": 0.8353596925735474,
            "entity": {}
        },
        {
            "id": 28,
            "distance": 0.7484277486801147,
            "entity": {}
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.1628494,
            "fields": {
                "vector": [
                    0.7442872,
                    0.046407282,
                    0.71031404,
                    0.3544345,
                    0.9819991
                ],
                "id": 1992
            }
        },
        {
            "score": 1.1470042,
            "fields": {
                "vector": [
                    0.5505825,
                    0.04367262,
                    0.9985836,
                    0.18922359,
                    0.93255126
                ],
                "id": 1977
            }
        },
        {
            "score": 1.1450152,
            "fields": {
                "vector": [
                    0.89994013,
                    0.052991092,
                    0.8645576,
                    0.6406729,
                    0.95679337
                ],
                "id": 1573
            }
        },
        {
            "score": 1.1439825,
            "fields": {
                "vector": [
                    0.9253267,
                    0.15890503,
                    0.7999555,
                    0.19126713,
                    0.898583
                ],
                "id": 1552
            }
        },
        {
            "score": 1.1029172,
            "fields": {
                "vector": [
                    0.95661926,
                    0.18777144,
                    0.38115507,
                    0.14323527,
                    0.93137646
                ],
                "id": 1823
            }
        }
    ],
    [
        {
            "score": 1.8005109,
            "fields": {
                "vector": [
                    0.5953582,
                    0.7794224,
                    0.9388869,
                    0.79825854,
                    0.9197286
                ],
                "id": 1888
            }
        },
        {
            "score": 1.7714822,
            "fields": {
                "vector": [
                    0.56805456,
                    0.89422905,
                    0.88187534,
                    0.914824,
                    0.8944365
                ],
                "id": 1648
            }
        },
        {
            "score": 1.7561421,
            "fields": {
                "vector": [
                    0.83421993,
                    0.39865613,
                    0.92319834,
                    0.42695504,
                    0.96633124
                ],
                "id": 1688
            }
        },
        {
            "score": 1.7553532,
            "fields": {
                "vector": [
                    0.89994013,
                    0.052991092,
                    0.8645576,
                    0.6406729,
                    0.95679337
                ],
                "id": 1573
            }
        },
        {
            "score": 1.7543385,
            "fields": {
                "vector": [
                    0.16542226,
                    0.38248396,
                    0.9888778,
                    0.80913955,
                    0.9501492
                ],
                "id": 1544
            }
        }
    ]
]}

[
  { score: 2.8421106338500977, id: '1745' },
  { score: 2.838560104370117, id: '1782' },
  { score: 2.8134000301361084, id: '1511' },
  { score: 2.718268871307373, id: '1679' },
  { score: 2.7014894485473633, id: '1597' }
]

The data in red differs from that in blue. Therefore, the search results will be constrained to the specified partition, reflecting the unique characteristics and data distribution of that subset.

Search with output fields

Search with output fields allows you to specify which attributes or fields of the matched vectors should be included in the search results.

You can specify output_fields in a request to return results with specific fields.

Here is an example of returning results with color attribute values:

Python Java Node.js 
# Search with output fields
res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    data=[[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]],
    limit=5, # Max. number of search results to return
    search_params={"metric_type": "IP", "params": {}}, # Search parameters
    output_fields=["color"] # Output fields to return
)

result = json.dumps(res, indent=4)
print(result)

// 7. Search with output fields
query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .outputFields(Arrays.asList("color"))
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 7. Search with output fields
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    limit: 5,
    output_fields: ["color"],
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 0,
            "distance": 1.4093276262283325,
            "entity": {
                "color": "pink_8682"
            }
        },
        {
            "id": 16,
            "distance": 1.0159327983856201,
            "entity": {
                "color": "yellow_1496"
            }
        },
        {
            "id": 4,
            "distance": 0.9902134537696838,
            "entity": {
                "color": "red_4794"
            }
        },
        {
            "id": 14,
            "distance": 0.9803846478462219,
            "entity": {
                "color": "green_2899"
            }
        },
        {
            "id": 1,
            "distance": 0.8519943356513977,
            "entity": {
                "color": "red_7025"
            }
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.263043,
            "fields": {}
        },
        {
            "score": 1.2377806,
            "fields": {}
        },
        {
            "score": 1.1869997,
            "fields": {}
        },
        {
            "score": 1.1748955,
            "fields": {}
        },
        {
            "score": 1.1720343,
            "fields": {}
        }
    ]
]}


[
  { score: 3.036271572113037, id: '59', color: 'orange' },
  { score: 3.0267879962921143, id: '1745', color: 'blue' },
  { score: 3.0069446563720703, id: '854', color: 'black' },
  { score: 2.984386682510376, id: '718', color: 'black' },
  { score: 2.916019916534424, id: '425', color: 'purple' }
]

Alongside the nearest neighbors, the search results will include the specified field color, providing a richer set of information for each matching vector.

Filtered search applies scalar filters to vector searches, allowing you to refine the search results based on specific criteria. You can find more about filter expressions in Boolean Expression Rules and examples in Get & Scalar Query.

Use the like operator

The like operator enhances string searches by evaluating patterns including prefixes, infixes, and suffixes:

  • Prefix matching: To find values starting with a specific prefix, use the syntax 'like "prefix%"'.
  • Infix matching: To find values containing a specific sequence of characters anywhere within the string, use the syntax 'like "%infix%"'.
  • Suffix matching: To find values ending with a specific suffix, use the syntax 'like "%suffix"'.

For single-character matching, underscore (_) acts as a wildcard for one character, e.g., 'like "y_llow"'.

Special characters in search strings

If you want to search for a string that contains special characters like underscores (_) or percent signs (%), which are normally used as wildcards in search patterns (_ for any single character and % for any sequence of characters), you must escape these characters to treat them as literal characters. Use a backslash (\) to escape special characters, and remember to escape the backslash itself. For instance:

  • To search for a literal underscore, use \\_.
  • To search for a literal percent sign, use \\%.

So, if you need to search for the text "_version_", your query should be formatted as 'like "\\_version\\_"' to ensure the underscores are treated as part of the search term and not as wildcards.

Filter results whose color is prefixed with red:

Python Java Node.js 
# Search with filter
res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    data=[[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]],
    limit=5, # Max. number of search results to return
    search_params={"metric_type": "IP", "params": {}}, # Search parameters
    output_fields=["color"], # Output fields to return
    filter='color like "red%"'
)

result = json.dumps(res, indent=4)
print(result)

// 8. Filtered search
query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .outputFields(Arrays.asList("color_tag"))
    .filter("color_tag like \"red%\"")
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 8. Filtered search
// 8.1 Filter with "like" operator and prefix wildcard
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    limit: 5,
    filters: "color_tag like \"red%\"",
    output_fields: ["color_tag"]
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 4,
            "distance": 0.9902134537696838,
            "entity": {
                "color": "red_4794"
            }
        },
        {
            "id": 1,
            "distance": 0.8519943356513977,
            "entity": {
                "color": "red_7025"
            }
        },
        {
            "id": 6,
            "distance": -0.4113418459892273,
            "entity": {
                "color": "red_9392"
            }
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.1869997,
            "fields": {"color_tag": "red_3026"}
        },
        {
            "score": 1.1677284,
            "fields": {"color_tag": "red_9030"}
        },
        {
            "score": 1.1476475,
            "fields": {"color_tag": "red_3744"}
        },
        {
            "score": 1.0969629,
            "fields": {"color_tag": "red_4168"}
        },
        {
            "score": 1.0741848,
            "fields": {"color_tag": "red_9678"}
        }
    ]
]}

[
  { score: 2.5080761909484863, id: '1201', color_tag: 'red_8904' },
  { score: 2.491129159927368, id: '425', color_tag: 'purple_8212' },
  { score: 2.4889798164367676, id: '1458', color_tag: 'red_6891' },
  { score: 2.42964243888855, id: '724', color_tag: 'black_9885' },
  { score: 2.4004223346710205, id: '854', color_tag: 'black_5990' }
]

Filter results whose color contains the letters ll anywhere within the string:

Python Java Node.js 
# Infix match on color field
res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    data=[[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]],
    limit=5, # Max. number of search results to return
    search_params={"metric_type": "IP", "params": {}}, # Search parameters
    output_fields=["color"], # Output fields to return
    filter='color like "%ll%"' # Filter on color field, infix match on "ll"
)

result = json.dumps(res, indent=4)
print(result)

// 8. Filtered search
query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .outputFields(Arrays.asList("color_tag"))
    .filter("color like \"%ll%\"")
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 8. Filtered search
// 8.1 Filter with "like" operator and prefix wildcard
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    limit: 5,
    filters: "color_tag like \"%ll%\"",
    output_fields: ["color_tag"]
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 5,
            "distance": 0.7972343564033508,
            "entity": {
                "color": "yellow_4222"
            }
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.1869997,
            "fields": {"color_tag": "yellow_4222"}
        }
    ]
]}

[
  { score: 2.5080761909484863, id: '1201', color_tag: 'yellow_4222' }
]

Range search allows you to find vectors that lie within a specified distance range from your query vector.

By setting radius and optionally range_filter, you can adjust the breadth of your search to include vectors that are somewhat similar to the query vector, providing a more comprehensive view of potential matches.

  • radius: Defines the outer boundary of your search space. Only vectors that are within this distance from the query vector are considered potential matches.

  • range_filter: While radius sets the outer limit of the search, range_filter can be optionally used to define an inner boundary, creating a distance range within which vectors must fall to be considered matches.

Python Java Node.js 
# Conduct a range search
search_params = {
    "metric_type": "IP",
    "params": {
        "radius": 0.8, # Radius of the search circle
        "range_filter": 1.0 # Range filter to filter out vectors that are not within the search circle
    }
}

res = client.search(
    collection_name="quick_setup", # Replace with the actual name of your collection
    data=[[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]],
    limit=3, # Max. number of search results to return
    search_params=search_params, # Search parameters
    output_fields=["color"], # Output fields to return
)

result = json.dumps(res, indent=4)
print(result)

// 9. Range search
query_vectors = Arrays.asList(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f));

searchReq = SearchReq.builder()
    .collectionName("quick_setup")
    .data(query_vectors)
    .outputFields(Arrays.asList("color_tag"))
    .searchParams(Map.of("radius", 0.1, "range", 1.0))
    .topK(5)
    .build();

searchResp = client.search(searchReq);

System.out.println(JSONObject.toJSON(searchResp));

// 9. Range search
query_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]

res = await client.search({
    collection_name: "quick_setup",
    data: [query_vector],
    limit: 5,
    params: {
        radius: 0.1,
        range: 1.0
    },
    output_fields: ["color_tag"]
})

console.log(res.results)

The output is similar to the following:

Python Java Node.js 
[
    [
        {
            "id": 4,
            "distance": 0.9902134537696838,
            "entity": {
                "color": "red_4794"
            }
        },
        {
            "id": 14,
            "distance": 0.9803846478462219,
            "entity": {
                "color": "green_2899"
            }
        },
        {
            "id": 1,
            "distance": 0.8519943356513977,
            "entity": {
                "color": "red_7025"
            }
        }
    ]
]

{"searchResults": [
    [
        {
            "score": 1.263043,
            "fields": {"color_tag": "green_2052"}
        },
        {
            "score": 1.2377806,
            "fields": {"color_tag": "purple_3709"}
        },
        {
            "score": 1.1869997,
            "fields": {"color_tag": "red_3026"}
        },
        {
            "score": 1.1748955,
            "fields": {"color_tag": "black_1646"}
        },
        {
            "score": 1.1720343,
            "fields": {"color_tag": "green_4853"}
        }
    ]
]}

[
  { score: 2.3387961387634277, id: '718', color_tag: 'black_7154' },
  { score: 2.3352415561676025, id: '1745', color_tag: 'blue_8741' },
  { score: 2.290485382080078, id: '1408', color_tag: 'red_2324' },
  { score: 2.285870313644409, id: '854', color_tag: 'black_5990' },
  { score: 2.2593345642089844, id: '1309', color_tag: 'red_8458' }
]

You will observe that all the entities returned have a distance that falls within the range of 0.8 to 1.0 from the query vector.

The parameter settings for radius and range_filter vary with the metric type in use.

Metric TypeCharactericsticsRange Search Settings
L2Smaller L2 distances indicate higher similarity.To exclude the closest vectors from results, ensure that:
range_filter <= distance < radius
IPLarger IP distances indicate higher similarity.To exclude the closest vectors from results, ensure that:
radius < distance <= range_filter
COSINELarger cosine value indicates higher similarity.To exclude the closest vectors from results, ensure that:
radius < distance <= range_filter
JACCARDSmaller Jaccard distances indicate higher similarity.To exclude the closest vectors from results, ensure that:
range_filter <= distance < radius
HAMMINGSmaller Hamming distances indicate higher similarity.To exclude the closest vectors from results, ensure that:
range_filter <= distance < radius

To learn more about distance metric types, refer to Similarity Metrics.

In Milvus, grouping search is designed to improve comprehensiveness and accuracy for search results.

Consider a scenario in RAG, where loads of documents are split into various passages, and each passage is represented by one vector embedding. Users want to find the most relevant passages to prompt the LLMs accurately. The ordinary Milvus search function can meet this requirement, but it may result in highly skewed and biased results: most of the passages come from only a few documents, and the comprehensiveness of the search results is very poor. This can seriously impair the accuracy or even correctness of the results given by the LLM and influence the LLM users’ experience negatively.

Grouping search can effectively solve this problem. By passing a group_by_field and group_size, Milvus users can bucket the search results into several groups and ensure that the number of entities from each group does not exceed a specific group_size. This feature can significantly enhance the comprehensiveness and fairness of search results, noticeably improving the quality of LLM output.

Here is the example code to group search results by field:

# Connect to Milvus
client = MilvusClient(uri='http://localhost:19530') # Milvus server address

# Load data into collection
client.load_collection("group_search") # Collection name

# Group search results
res = client.search(
    collection_name="group_search", # Collection name
    data=[[0.14529211512077012, 0.9147257273453546, 0.7965055218724449, 0.7009258593102812, 0.5605206522382088]], # Query vector
    search_params={
    "metric_type": "L2",
    "params": {"nprobe": 10},
    }, # Search parameters
    limit=5, # Max. number of groups to return
    group_by_field="doc_id", # Group results by document ID
    group_size=2, # returned at most 2 passages per document, the default value is 1
    group_strict_size=True, # ensure every group contains exactly 3 passages
    output_fields=["doc_id", "passage_id"]
)

# Retrieve the values in the `doc_id` column
doc_ids = [result['entity']['doc_id'] for result in res[0]]
passage_ids = [result['entity']['passage_id'] for result in res[0]]

print(doc_ids)
print(passage_ids)

The output is similar to the following:

["doc_11", "doc_11", "doc_7", "doc_7", "doc_3", "doc_3", "doc_2", "doc_2", "doc_8", "doc_8"]
[5, 10, 11, 10, 9, 6, 5, 4, 9, 2]

In the given output, it can be observed that for each document, exactly two passages are retrieved and a total of 5 documents collectively make up the results.

For comparison, let’s comment out the group-related parameters and conduct a regular search:

# Connect to Milvus
client = MilvusClient(uri='http://localhost:19530') # Milvus server address

# Load data into collection
client.load_collection("group_search") # Collection name

# Search without `group_by_field`
res = client.search(
    collection_name="group_search", # Collection name
    data=query_passage_vector, # Replace with your query vector
    search_params={
    "metric_type": "L2",
    "params": {"nprobe": 10},
    }, # Search parameters
    limit=5, # Max. number of search results to return
    # group_by_field="doc_id", # Group results by document ID
    # group_size=2, 
    # group_strict_size=True,
    output_fields=["doc_id", "passage_id"]
)

# Retrieve the values in the `doc_id` column
doc_ids = [result['entity']['doc_id'] for result in res[0]]
passage_ids = [result['entity']['passage_id'] for result in res[0]]

print(doc_ids)
print(passage_ids)

The output is similar to the following:

["doc_11", "doc_11", "doc_11", "doc_11", "doc_11"]
[1, 10, 3, 12, 9]

In the given output, it can be observed that “doc_11” completely dominated the search results, overshadowing the high-quality paragraphs from other documents, which can be a poor prompt to LLM.

One more point to note: by default, grouping_search will return results instantly when it has enough groups, which may lead to the number of results in each group not being sufficient to meet the group_size. If you care about the number of results for each group, set group_strict_size=True as shown in the code above. This will make Milvus strive to obtain enough results for each group, at a slight cost to performance.

Limitations

  • Indexing: This grouping feature works only for collections that are indexed with the HNSW, IVF_FLAT, or FLAT type. For more information, refer to In-memory Index.

  • Vector: Currently, grouping search does not support a vector field of the BINARY_VECTOR type. For more information on data types, refer to Supported data types.

  • Field: Currently, grouping search allows only for a single column. You cannot specify multiple field names in the group_by_field config. Additionally, grouping search is incompatible with data types of JSON, FLOAT, DOUBLE, ARRAY, or vector fields.

  • Performance Impact: Be mindful that performance degrades with increasing query vector counts. Using a cluster with 2 CPU cores and 8 GB of memory as an example, the execution time for grouping search increases proportionally with the number of input query vectors.

  • Functionality: Currently, grouping search is not supported by range search, search iterators

Search parameters

In the above searches except the range search, the default search parameters apply. In normal cases, you do not need to manually set search parameters.

# In normal cases, you do not need to set search parameters manually
# Except for range searches.
search_parameters = {
    'metric_type': 'L2',
    'params': {
        'nprobe': 10,
        'level': 1，
        'radius': 1.0
        'range_filter': 0.8
    }
}

The following table lists all possible settings in the search parameters.

Parameter NameParameter Description
metric_typeHow to measure similarity between vector embeddings.
Possible values are IP, L2, COSINE, JACCARD, and HAMMING, and defaults to that of the loaded index file.
params.nprobeNumber of units to query during the search.
The value falls in the range [1, nlist[1]].
params.levelSearch precision level.
Possible values are 1, 2, and 3, and defaults to 1. Higher values yield more accurate results but slower performance.
params.radiusDefines the outer boundary of your search space. Only vectors that are within this distance from the query vector are considered potential matches.
The value range is determined by the metric_type parameter. For instance, if metric_type is set to L2, the valid value range is [0, ∞]. If metric_type is set to COSINE, the valid value range is [-1, 1]. For more information, refer to Similarity Metrics.
params.range_filterWhile radius sets the outer limit of the search, range_filter can be optionally used to define an inner boundary, creating a distance range within which vectors must fall to be considered matches.
The value range is determined by the metric_type parameter. For instance, if metric_type is set to L2, the valid value range is [0, ∞]. If metric_type is set to COSINE, the valid value range is [-1, 1]. For more information, refer to Similarity Metrics.

notes

[1] Number of cluster units after indexing. When indexing a collection, Milvus sub-divides the vector data into multiple cluster units, the number of which varies with the actual index settings.

[2] Number of entities to return in a search.

Übersetzt vonDeepL

Tabelle der Inhalte

Try Managed Milvus for Free

Zilliz Cloud is hassle-free, powered by Milvus and 10x faster.

Get Started
Feedback

War diese Seite hilfreich?