Query public index to get nearest neighbors

After you've created and deployed the index, you can run queries to get the nearest neighbors.

Here are some examples for a match query to find the top nearest neighbors using the k-nearest neighbors algorithm (k-NN).

Example queries for public endpoint

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation. 

def vector_search_find_neighbors(
    project: str,
    location: str,
    index_endpoint_name: str,
    deployed_index_id: str,
    queries: List[List[float]],
    num_neighbors: int,
) -> List[
    List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]
]:
    """Query the vector search index.

    Args:
        project (str): Required. Project ID
        location (str): Required. The region name
        index_endpoint_name (str): Required. Index endpoint to run the query
        against.
        deployed_index_id (str): Required. The ID of the DeployedIndex to run
        the queries against.
        queries (List[List[float]]): Required. A list of queries. Each query is
        a list of floats, representing a single embedding.
        num_neighbors (int): Required. The number of neighbors to return.

    Returns:
        List[List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]] - A list of nearest neighbors for each query.
    """
    # Initialize the Vertex AI client
    aiplatform.init(project=project, location=location)

    # Create the index endpoint instance from an existing endpoint.
    my_index_endpoint = aiplatform.MatchingEngineIndexEndpoint(
        index_endpoint_name=index_endpoint_name
    )

    # Query the index endpoint for the nearest neighbors.
    return my_index_endpoint.find_neighbors(
        deployed_index_id=deployed_index_id,
        queries=queries,
        num_neighbors=num_neighbors,
    )

Command-line

The publicEndpointDomainName listed below can be found at Deploy and is formatted as <number>.<region>-<number>.vdb.vertexai.goog. 


  $ curl -X POST -H "Content-Type: application/json" -H "Authorization: Bearer `gcloud auth print-access-token`" https://1957880287.us-central1-181224308459.vdb.vertexai.goog/v1/projects/181224308459/locations/us-central1/indexEndpoints/3370566089086861312:findNeighbors -d '{deployed_index_id: "test_index_public1", queries: [{datapoint: {datapoint_id: "0", feature_vector: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}, neighbor_count: 5}]}'

This curl example demonstrates how to call from http(s) clients, although public endpoint supports dual protocol for restful and grpc_cli. 


  $ curl -X POST -H "Content-Type: application/json" -H "Authorization: Bearer `gcloud auth print-access-token`" https://1957880287.us-central1-181224308459.vdb.vertexai.goog/v1/projects/${PROJECT_ID}/locations/us-central1/indexEndpoints/${INDEX_ENDPOINT_ID}:readIndexDatapoints -d '{deployed_index_id:"test_index_public1", ids: ["606431", "896688"]}'

This curl example demonstrates how to query with token and numeric restricts.


  $ curl -X POST -H "Content-Type: application/json" -H "Authorization: Bearer `gcloud auth print-access-token`"  https://${PUBLIC_ENDPOINT_DOMAIN}/v1/projects/${PROJECT_ID}/locations/${LOCATION}/indexEndpoints/${INDEX_ENDPOINT_ID}:findNeighbors -d '{deployed_index_id:"${DEPLOYED_INDEX_ID}", queries: [{datapoint: {datapoint_id:"x", feature_vector: [1, 1], "sparse_embedding": {"values": [111.0,111.1,111.2], "dimensions": [10,20,30]}, numeric_restricts: [{namespace: "int-ns", value_int: -2, op: "GREATER"}, {namespace: "int-ns", value_int: 4, op: "LESS_EQUAL"}, {namespace: "int-ns", value_int: 0, op: "NOT_EQUAL"}], restricts: [{namespace: "color", allow_list: ["red"]}]}}]}'

Console

Use these instructions to query an index deployed to a public endpoint from the console.

  1. In the Vertex AI section of the Google Cloud console, go to the Deploy and Use section. Select Vector Search.

    Go to Vector Search

  2. Select the index you want to query. The Index info page opens.
  3. Scroll down to the Deployed indexes section and select the deployed index you want to query. The Deployed index info page opens.
  4. From the Query index section, select whether to query by a dense embedding value, a sparse embedding value, a hybrid embedding value (dense and sparse embeddings), or a specific data point.
  5. Enter the query parameters for the type of query you selected. For example, if you're querying by a dense embedding, enter the embedding vector to query by.
  6. Execute the query using the provided curl command, or by running with Cloud Shell.
  7. If using Cloud Shell, select Run in Cloud Shell.
  8. Run in Cloud Shell.
  9. The results return nearest neighbors.

Hybrid queries

Hybrid search uses both dense and sparse embeddings for searches based on combination of keyword search and semantic search.

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation. 

def vector_search_find_neighbors_hybrid_queries(
    project: str,
    location: str,
    index_endpoint_name: str,
    deployed_index_id: str,
    num_neighbors: int,
) -> List[
    List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]
]:
    """Query the vector search index using example hybrid queries.

    Args:
        project (str): Required. Project ID
        location (str): Required. The region name
        index_endpoint_name (str): Required. Index endpoint to run the query
        against.
        deployed_index_id (str): Required. The ID of the DeployedIndex to run
        the queries against.
        num_neighbors (int): Required. The number of neighbors to return.

    Returns:
        List[List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]] - A list of nearest neighbors for each query.
    """
    # Initialize the Vertex AI client
    aiplatform.init(project=project, location=location)

    # Create the index endpoint instance from an existing endpoint.
    my_index_endpoint = aiplatform.MatchingEngineIndexEndpoint(
        index_endpoint_name=index_endpoint_name
    )

    # Query hybrid datapoints, sparse-only datapoints, and dense-only datapoints.
    hybrid_queries = [
        aiplatform.matching_engine.matching_engine_index_endpoint.HybridQuery(
            dense_embedding=[1, 2, 3],
            sparse_embedding_dimensions=[10, 20, 30],
            sparse_embedding_values=[1.0, 1.0, 1.0],
            rrf_ranking_alpha=0.5,
        ),
        aiplatform.matching_engine.matching_engine_index_endpoint.HybridQuery(
            dense_embedding=[1, 2, 3],
            sparse_embedding_dimensions=[10, 20, 30],
            sparse_embedding_values=[0.1, 0.2, 0.3],
        ),
        aiplatform.matching_engine.matching_engine_index_endpoint.HybridQuery(
            sparse_embedding_dimensions=[10, 20, 30],
            sparse_embedding_values=[0.1, 0.2, 0.3],
        ),
        aiplatform.matching_engine.matching_engine_index_endpoint.HybridQuery(
            dense_embedding=[1, 2, 3]
        ),
    ]

    return my_index_endpoint.find_neighbors(
        deployed_index_id=deployed_index_id,
        queries=hybrid_queries,
        num_neighbors=num_neighbors,
    )

Queries with filtering and crowding

Filtering vector matches lets you restrict your nearest neighbor results to specific categories. Filters can also designate categories to exclude from your results.

Per-crowding neighbor limits can increase result diversity by limiting the number of results returned from any single crowding tag in your index data.

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation. 

def vector_search_find_neighbors_filtering_crowding(
    project: str,
    location: str,
    index_endpoint_name: str,
    deployed_index_id: str,
    queries: List[List[float]],
    num_neighbors: int,
    filter: List[aiplatform.matching_engine.matching_engine_index_endpoint.Namespace],
    numeric_filter: List[
        aiplatform.matching_engine.matching_engine_index_endpoint.NumericNamespace
    ],
    per_crowding_attribute_neighbor_count: int,
) -> List[
    List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]
]:
    """Query the vector search index with filtering and crowding.

    Args:
        project (str): Required. Project ID
        location (str): Required. The region name
        index_endpoint_name (str): Required. Index endpoint to run the query
        against.
        deployed_index_id (str): Required. The ID of the DeployedIndex to run
        the queries against.
        queries (List[List[float]]): Required. A list of queries. Each query is
        a list of floats, representing a single embedding.
        num_neighbors (int): Required. The number of neighbors to return.
        filter (List[Namespace]): Required. A list of Namespaces for filtering
        the matching results. For example,
        [Namespace("color", ["red"], []), Namespace("shape", [], ["square"])]
        will match datapoints that satisfy "red color" but not include
        datapoints with "square shape".
        numeric_filter (List[NumericNamespace]): Required. A list of
        NumericNamespaces for filtering the matching results. For example,
        [NumericNamespace(name="cost", value_int=5, op="GREATER")] will limit
        the matching results to datapoints with cost greater than 5.
        per_crowding_attribute_neighbor_count (int): Required. The maximum
        number of returned matches with the same crowding tag.

    Returns:
        List[List[aiplatform.matching_engine.matching_engine_index_endpoint.MatchNeighbor]] - A list of nearest neighbors for each query.
    """
    # Initialize the Vertex AI client
    aiplatform.init(project=project, location=location)

    # Create the index endpoint instance from an existing endpoint.
    my_index_endpoint = aiplatform.MatchingEngineIndexEndpoint(
        index_endpoint_name=index_endpoint_name
    )

    # Query the index endpoint for the nearest neighbors.
    return my_index_endpoint.find_neighbors(
        deployed_index_id=deployed_index_id,
        queries=queries,
        num_neighbors=num_neighbors,
        filter=filter,
        numeric_filter=numeric_filter,
        per_crowding_attribute_neighbor_count=per_crowding_attribute_neighbor_count,
    )

Query-time settings that impact performance

The following query-time parameters can affect latency, availability, and cost when using Vector Search. This guidance applies to most cases. However, always experiment with your configurations to make sure that they work for your use case.

For parameter definitions, see Index configuration parameters.

Parameter About Performance impact
approximateNeighborsCount

Tells the algorithm the number of approximate results to retrieve from each shard.

The value of approximateNeighborsCount should always be greater than the value of setNeighborsCount. If the value of setNeighborsCount is small, 10 times that value is recommended for approximateNeighborsCount. For larger setNeighborsCount values, a smaller multiplier can be used.

Increasing the value of approximateNeighborsCount can affect performance in the following ways:

  • Recall: Increased
  • Latency: Potentially increased
  • Availability: No impact
  • Cost: Can increase because more data is processed during a search

Decreasing the value of approximateNeighborsCount can affect performance in the following ways:

  • Recall: Decreased
  • Latency: Potentially decreases
  • Availability: No impact
  • Cost: Can decrease cost because less data is processed during a search
setNeighborCount Specifies the number of results that you want the query to return.

Values less than or equal to 300 remain performant in most use cases. For larger values, test for your specific use case.

fractionLeafNodesToSearch Controls the percentage of leaf nodes to visit when searching for nearest neighbors. This is related to the leafNodeEmbeddingCount in that the more embeddings per leaf node, the more data examined per leaf.

Increasing the value of fractionLeafNodesToSearch can affect performance in the following ways:

  • Recall: Increased
  • Latency: Increased
  • Availability: No impact
  • Cost: Can increase because higher latency occupies more machine resources

Decreasing the value of fractionLeafNodesToSearch can affect performance in the following ways:

  • Recall: Decreased
  • Latency: Decreased
  • Availability: No impact
  • Cost: Can decrease because lower latency occupies fewer machine resources

What's next

To see an end-to-end example of how to create an index, how to deploy it to a public endpoint, and how to query, see the official notebook: Using Vector Search and Vertex AI Embeddings for Text for StackOverflow Questions.