Configure Neural Search and Hybrid Search on OpenSearch Serverless - Amazon OpenSearch Service
Neural SearchHybrid SearchNeural and hybrid queriesConfigure permissions

Configure Neural Search and Hybrid Search on OpenSearch Serverless

Neural Search

Amazon OpenSearch Serverless supports Neural Search functionality for semantic search operations on your data. Neural Search uses machine learning models to understand the semantic meaning and context of your queries, providing more relevant search results than traditional keyword-based searches. This section explains how to configure Neural Search in OpenSearch Serverless, including the required permissions, supported processors, and key differences from the standard OpenSearch implementation.

With Neural Search you can perform semantic search on your data, which considers semantic meaning to understand the intent of your search queries. This capability is powered by the following components:

  • Text embedding ingest pipeline processor

  • Neural query

  • Neural sparse query

With hybrid search, you can improve search relevance by combining keyword and semantic search capabilities. To use hybrid search, create a search pipeline that processes your search results and combines document scores. For more information, see Search pipelines on the OpenSearch Documentation website. Use the following components to implement hybrid search:

  • Normalization search pipeline processor

    Supported normalization techniques

    • min_max

    • l2

    Supported combination techniques

    • arithmetic_mean

    • geometric_mean

    • harmonic_mean

    For more information about normalization and combination techniques, see Request body fields on the OpenSearch Documentation website.

  • Hybrid query

Neural and hybrid queries

By default, OpenSearch calculates document scores using the keyword-based Okapi BM25 algorithm, which works well for search queries that contain keywords. Neural Search provides new query types for natural language queries and the ability to combine both semantic and keyword search.

Example : neural

"neural": {
  "vector_field": {
    "query_text": "query_text",
    "query_image": "image_binary",
    "model_id": "model_id",
    "k": 100
  }
}

For more information, see Neural query on the OpenSearch Documentation website.

Example : hybrid

"hybrid": {
      "queries": [
        array of lexical, neural, or combined queries
      ]
    }

For more information, see Hybrid query on the OpenSearch Documentation website.

To configure semantic search components in Amazon OpenSearch Serverless, follow the steps in the Neural Search tutorial on the OpenSearch Documentation website. Keep in mind these important differences:

  • OpenSearch Serverless supports only remote models. You must configure connectors to remotely hosted models. You don't need to deploy or remove remote models. For more information, see Getting started with semantic and hybrid search on the OpenSearch Documentation website.

  • Expect up to 15 seconds of latency when you search against your vector index or search for recently created search and ingest pipelines.

Configure permissions

Neural Search in OpenSearch Serverless requires the following permissions. For more information, see Supported policy permissions.

Example : Neural search policy
JSON
{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "NeuralSearch",
            "Effect": "Allow",
            "Action": [
              "aoss:CreateIndex",
              "aoss:CreateCollectionItems",
              "aoss:CreateMLResource",
              "aoss:DescribeCollectionItems",
              "aoss:UpdateCollectionItems",
              "aoss:DeleteIndex",
              "aoss:DeleteCollectionItems",
              "aoss:DeleteMLResource",
              "aoss:APIAccessAll",
            ],
            "Resource": "*"
        }
    ]
}

  • aoss:*Index – Creates a vector index where text embeddings are stored.

  • aoss:*CollectionItems – Creates ingest and search pipelines.

  • aoss:*MLResource – Creates and registers text embedding models.

  • aoss:APIAccessAll – Provides access to OpenSearch APIs for search and ingest operations.

The following describes the collection data access policies required for neural search. Replace the placeholder values with your specific information.

Example : Data access policy
[
    {
        "Description": "Create index permission",
        "Rules": [
            {
                "ResourceType": "index",
                "Resource": ["index/collection_name/*"],
                "Permission": [
                  "aoss:CreateIndex", 
                  "aoss:DescribeIndex",
                  "aoss:UpdateIndex",
                  "aoss:DeleteIndex"
                ]
            }
        ],
        "Principal": [
            "arn:aws:iam::account_id:role/role_name"
        ]
    },
    {
        "Description": "Create pipeline permission",
        "Rules": [
            {
                "ResourceType": "collection",
                "Resource": ["collection/collection_name"],
                "Permission": [
                  "aoss:CreateCollectionItems",
                  "aoss:DescribeCollectionItems",
                  "aoss:UpdateCollectionItems",
                  "aoss:DeleteCollectionItems"
                ]
            }
        ],
        "Principal": [
            "arn:aws:iam::account_id:role/role_name"
        ]
    },
    {
        "Description": "Create model permission",
        "Rules": [
            {
                "ResourceType": "model",
                "Resource": ["model/collection_name/*"],
                "Permission": ["aoss:CreateMLResources"]
            }
        ],
        "Principal": [
            "arn:aws:iam::account_id:role/role_name"
        ]
    }
]