With Iterators

Milvus provides search and query iterators for iterating through a large volume of entities. Since Milvus limits TopK to 16384, users can use iterators to return large numbers or even whole entities in a collection in batch mode.

Overview

Iterators are efficient tool for scanning a whole collection or iterating through a large volume of entities by specifying primary key values or a filter expression. Compared to a search or query call with offset and limit parameters, using iterators is more efficient and scalable.

Benefits of using iterators

• Simplicity: Eliminates the complex offset and limit settings.

• Efficiency: Provides scalable data retrieval by fetching only the data in need.

• Consistency: Ensures a consistent dataset size with boolean filters.

Preparations

The following preparation step connects to Milvus and inserts randomly generated entities into a collection.

Step 1: Create a collection

from pymilvus import MilvusClient

# 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,
)

import com.google.gson.Gson;
import com.google.gson.JsonObject;
import io.milvus.orm.iterator.QueryIterator;
import io.milvus.orm.iterator.SearchIterator;
import io.milvus.response.QueryResultsWrapper;
import io.milvus.v2.client.MilvusClientV2;
import io.milvus.v2.client.ConnectConfig;
import io.milvus.v2.common.ConsistencyLevel;
import io.milvus.v2.common.IndexParam;
import io.milvus.v2.service.collection.request.CreateCollectionReq;
import io.milvus.v2.service.collection.request.DropCollectionReq;
import io.milvus.v2.service.vector.request.*;
import io.milvus.v2.service.vector.request.data.FloatVec;
import io.milvus.v2.service.vector.response.InsertResp;
import io.milvus.v2.service.vector.response.QueryResp;

import java.util.*;

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

// 1. Connect to Milvus server
ConnectParam connectParam = ConnectParam.newBuilder()
        .withUri(CLUSTER_ENDPOINT)
        .build();

MilvusServiceClient client  = new MilvusServiceClient(connectParam);

// 2. Create a collection
CreateCollectionReq quickSetupReq = CreateCollectionReq.builder()
        .collectionName("quick_setup")
        .dimension(5)
        .build();
client.createCollection(quickSetupReq);

Step 2: Insert randomly generated entities

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

for i in range(10000):
    current_color = random.choice(colors)
    current_tag = random.randint(1000, 9999)
    data.append({
        "id": i,
        "vector": [ random.uniform(-1, 1) for _ in range(5) ],
        "color": current_color,
        "tag": current_tag,
        "color_tag": f"{current_color}_{str(current_tag)}"
    })

print(data[0])

# Output
#
# {
#     "id": 0,
#     "vector": [
#         -0.5705990742218152,
#         0.39844925120642083,
#         -0.8791287928610869,
#         0.024163154953680932,
#         0.6837669917169638
#     ],
#     "color": "purple",
#     "tag": 7774,
#     "color_tag": "purple_7774"
# }

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

print(res)

# Output
#
# {
#     "insert_count": 10000,
#     "ids": [
#         0,
#         1,
#         2,
#         3,
#         4,
#         5,
#         6,
#         7,
#         8,
#         9,
#         "(9990 more items hidden)"
#     ]
# }

// 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<10000; 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 + "_" + (rand.nextInt(8999) + 1000));
    data.add(row);
}

InsertResp insertR = client.insert(InsertReq.builder()
        .collectionName("quick_setup")
        .data(data)
        .build());
System.out.println(insertR.getInsertCnt());

// Output
// 10000

Search with iterator

Iterators make similarity searches more scalable.

1. Initialize the search iterator to define the search parameters and output fields.

2. Use the next() method within a loop to paginate through the search results.

   • If the method returns an empty array, the loop ends, and no more pages are available.

   • All results carry the specified output fields.

3. Manually call the close() method to close the iterator once all data has been retrieved.

from pymilvus import Collection,connections

# 4. Search with iterator
connections.connect(host="127.0.0.1", port=19530)
collection = Collection("quick_setup")

query_vectors = [[0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]]
search_params = {
    "metric_type": "IP",
    "params": {"nprobe": 10}
}

iterator = collection.search_iterator(
    data=query_vectors,
    anns_field="vector",
    batch_size=10,
    param=search_params,
    output_fields=["color_tag"],
    limit=300
)
# search 300 entities totally with 10 entities per page

results = []

while True:
    result = iterator.next()
    if not result:
        iterator.close()
        break
        
    results.extend(result)
    
    for hit in result:
        results.append(hit.to_dict())

print(results)

# Output
#
# [
#     {
#         "id": 1756,
#         "distance": 2.0642056465148926,
#         "entity": {
#             "color_tag": "black_9109"
#         }
#     },
#     {
#         "id": 6488,
#         "distance": 1.9437453746795654,
#         "entity": {
#             "color_tag": "purple_8164"
#         }
#     },
#     {
#         "id": 3338,
#         "distance": 1.9107104539871216,
#         "entity": {
#             "color_tag": "brown_8121"
#         }
#     }
# ]

// 4. Search with iterators
SearchIteratorReq iteratorReq = SearchIteratorReq.builder()
        .collectionName("quick_setup")
        .vectorFieldName("vector")
        .batchSize(10L)
        .vectors(Collections.singletonList(new FloatVec(Arrays.asList(0.3580376395471989f, -0.6023495712049978f, 0.18414012509913835f, -0.26286205330961354f, 0.9029438446296592f))))
        .params("{\"level\": 1}")
        .metricType(IndexParam.MetricType.COSINE)
        .outputFields(Collections.singletonList("color_tag"))
        .topK(300)
        .build();

SearchIterator searchIterator = client.searchIterator(iteratorReq);

List<QueryResultsWrapper.RowRecord> results = new ArrayList<>();
while (true) {
    List<QueryResultsWrapper.RowRecord> batchResults = searchIterator.next();
    if (batchResults.isEmpty()) {
        searchIterator.close();
        break;
    }

    results.addAll(batchResults);
}
System.out.println(results.size());

// Output
// 300

Parameter	Description
data	A list of vector embeddings. Milvus searches for the most similar vector embeddings to the specified ones.
anns_field	The name of the vector field in the current collection.
batch_size	The number of entities to return each time you call next() on the current iterator. The value defaults to 1000. Set it to a proper value to control the number of entities to return per iteration.
param	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_iterator().
output_fields	A list of field names to include in each entity in return. The value defaults to None. If left unspecified, only the primary field is included.
limit	The total number of entities to return. The value defaults to -1, indicating all matching entities will be in return.

Parameter	Description
withCollectionName	Set the collection name. Collection name cannot be empty or null.
withVectorFieldName	Set target vector field by name. Field name cannot be empty or null.
withVectors	Set the target vectors. Up to 16384 vectors allowed.
withBatchSize	The number of entities to return each time you call next() on the current iterator. The value defaults to 1000. Set it to a proper value to control the number of entities to return per iteration.
withParams	Specifies the parameters of search in JSON format. For more information, refer to searchIterator().

Query with an iterator

# 6. Query with iterator
iterator = collection.query_iterator(
    batch_size=10, # Controls the size of the return each time you call next()
    expr="color_tag like \"brown_8\"",
    output_fields=["color_tag"]
)

results = []

while True:
    result = iterator.next()
    if not result:
        iterator.close()
        break
        
    results.extend(result)
    
# 8. Check the search results
print(len(results))

print(results[:3])

# Output
#
# [
#     {
#         "color_tag": "brown_8785",
#         "id": 94
#     },
#     {
#         "color_tag": "brown_8568",
#         "id": 176
#     },
#     {
#         "color_tag": "brown_8721",
#         "id": 289
#     }
# ]

// 5. Query with iterators
QueryIterator queryIterator = client.queryIterator(QueryIteratorReq.builder()
        .collectionName("quick_setup")
        .expr("color_tag like \"brown_8%\"")
        .batchSize(50L)
        .outputFields(Arrays.asList("vector", "color_tag"))
        .build());

results.clear();
while (true) {
    List<QueryResultsWrapper.RowRecord> batchResults = queryIterator.next();
    if (batchResults.isEmpty()) {
        queryIterator.close();
        break;
    }

    results.addAll(batchResults);
}

System.out.println(results.subList(0, 3));

// Output
// [
//  [color_tag:brown_8975, vector:[0.93425006, 0.42161798, 0.1603949, 0.86406225, 0.30063087], id:104],
//  [color_tag:brown_8292, vector:[0.075261295, 0.51725155, 0.13842249, 0.13178307, 0.90713704], id:793],
//  [color_tag:brown_8763, vector:[0.80366623, 0.6534371, 0.6446101, 0.094082, 0.1318503], id:1157]
// ]

Parameter	Description
batch_size	The number of entities to return each time you call next() on the current iterator. The value defaults to 1000. Set it to a proper value to control the number of entities to return per iteration.
expr	A scalar filtering condition to filter matching entities. The value defaults to None, indicating that scalar filtering is ignored. To build a scalar filtering condition, refer to Boolean Expression Rules.
output_fields	A list of field names to include in each entity in return. The value defaults to None. If left unspecified, only the primary field is included.
limit	The total number of entities to return. The value defaults to -1, indicating all matching entities will be in return.

Parameter	Description
withCollectionName	Set the collection name. Collection name cannot be empty or null.
withExpr	Set the expression to query entities. To build a scalar filtering condition, refer to Boolean Expression Rules.
withBatchSize	The number of entities to return each time you call next() on the current iterator. The value defaults to 1000. Set it to a proper value to control the number of entities to return per iteration.
addOutField	Specifies an output scalar field (Optional).