Introduction What is semantic search Model details and performance benchmark Languages Supported Set up an automatic semantic enrichment index for serverless collections Data ingestion and search Configuring permissions for automatic semantic enrichment Query Rewrites Limitations of automatic semantic enrichment Pricing

Automatic semantic enrichment for Serverless

Introduction

The automatic semantic enrichment feature can help improve search relevance by up to 20% over lexical search. Automatic semantic enrichment eliminates the undifferentiated heavy lifting of managing your own ML (machine learning) model infrastructure and integration with the search engine. The feature is available for all three serverless collection types: Search, Time Series, and Vector.

What is semantic search

Traditional search engines rely on word-to-word matching (referred to as lexical search) to find results for queries. Although this works well for specific queries such as television model numbers, it struggles with more abstract searches. For example, when searching for "shoes for the beach," a lexical search merely matches individual words "shoes," "beach," "for," and "the" in catalog items, potentially missing relevant products like "water-resistant sandals" or "surf footwear" that don't contain the exact search terms.

Semantic search returns query results that incorporate not just keyword matching, but the intent and contextual meaning of the user's search. For example, if a user searches for "how to treat a headache," a semantic search system might return the following results:

Migraine remedies
Pain management techniques
Over-the-counter pain relievers

Model details and performance benchmark

While this feature handles the technical complexities behind the scenes without exposing the underlying model, we provide transparency through a brief model description and benchmark results to help you make informed decisions about feature adoption in your critical workloads.

Automatic semantic enrichment uses a service-managed, pre-trained sparse model that works effectively without requiring custom fine-tuning. The model analyzes the fields you specify, expanding them into sparse vectors based on learned associations from diverse training data. The expanded terms and their significance weights are stored in native Lucene index format for efficient retrieval. We’ve optimized this process using document-only mode, where encoding happens only during data ingestion. Search queries are merely tokenized rather than processed through the sparse model, making the solution both cost-effective and performant.

Our performance validation during feature development used the MS MARCO passage retrieval dataset, featuring passages averaging 334 characters. For relevance scoring, we measured average Normalized Discounted Cumulative Gain (NDCG) for the first 10 search results (ndcg@10) on the BEIR benchmark for English content and average ndcg@10 on MIRACL for multilingual content. We assessed latency through client-side, 90th-percentile (p90) measurements and search response p90 took values. These benchmarks provide baseline performance indicators for both search relevance and response times. Here are the key benchmark numbers -

English language - Relevance improvement of 20% over lexical search. It also lowered P90 search latency by 7.7% over lexical search (BM25 is 26 ms, and automatic semantic enrichment is 24 ms).
Multi-lingual - Relevance improvement of 105% over lexical search, whereas P90 search latency increased by 38.4% over lexical search (BM25 is 26 ms, and automatic semantic enrichment is 36 ms).

Given the unique nature of each workload, we encourage you to evaluate this feature in your development environment using your own benchmarking criteria before making implementation decisions.

Languages Supported

The feature supports English. In addition, the model also supports Arabic, Bengali, Chinese, Finnish, French, Hindi, Indonesian, Japanese, Korean, Persian, Russian, Spanish, Swahili, and Telugu.

Set up an automatic semantic enrichment index for serverless collections

Setting up an index with automatic semantic enrichment enabled for your text fields is easy, and you can manage it through the console, APIs, and Amazon CloudFormation templates during new index creation. To enable it for an existing index, you need to recreate the index with automatic semantic enrichment enabled for text fields.

Console experience - The Amazon console allows you to easily create an index with automatic semantic enrichment fields. Once you select a collection, you will find the create index button at the top of the console. Once you click the create index button, you will find options to define automatic semantic enrichment fields. In one index, you can have combinations of automatic semantic enrichment for English and multilingual, as well as lexical fields.

API experience - To create an automatic semantic enrichment index using the Amazon Command Line Interface (Amazon CLI), use the create-index command:


aws opensearchserverless create-index \
--id [collection_id] \
--index-name [index_name] \
--index-schema [index_body] \

In the following example index-schema, the title_semantic field has a field type set to text and has parameter semantic_enrichment set to status ENABLED. Setting the semantic_enrichment parameter enables automatic semantic enrichment on the title_semantic field. You can use the language_options field to specify either english or multi-lingual.



    aws opensearchserverless create-index \
    --id XXXXXXXXX \
    --index-name 'product-catalog' \
    --index-schema '{
    "mappings": {
        "properties": {
            "product_id": {
                "type": "keyword"
            },
            "title_semantic": {
                "type": "text",
                "semantic_enrichment": {
                    "status": "ENABLED",
                    "language_options": "english"
                }
            },
            "title_non_semantic": {
                "type": "text"
            }
        }
    }
}'

To describe the created index, use the following command:


aws opensearchserverless get-index \
--id [collection_id] \
--index-name [index_name] \

You can also use Amazon CloudFormation templates (Type:AWS::OpenSearchServerless::CollectionIndex) to create semantic search during collection provisioning as well as after the collection is created.

Data ingestion and search

Once you've created an index with automatic semantic enrichment enabled, the feature works automatically during data ingestion process, no additional configuration required.

Data ingestion: When you add documents to your index, the system automatically:

Analyzes the text fields you designated for semantic enrichment
Generates semantic encodings using OpenSearch Service managed sparse model
Stores these enriched representations alongside your original data

This process uses OpenSearch's built-in ML connectors and ingest pipelines, which are created and managed automatically behind the scenes.

Search: The semantic enrichment data is already indexed, so queries run efficiently without invoking the ML model again. This means you get improved search relevance with no additional search latency overhead.

Configuring permissions for automatic semantic enrichment

Before creating an automated semantic enrichment index, you need to configure the required permissions. This section explains the permissions needed and how to set them up.

IAM policy permissions

Use the following Amazon Identity and Access Management (IAM) policy to grant the necessary permissions for working with automatic semantic enrichment:

Key permissions

The aoss:*Index permissions enable index management
The aoss:APIAccessAll permission allows OpenSearch API operations
To restrict permissions to a specific collection, replace "Resource": "*" with the collection's ARN

Configure data access permissions

To set up an index for automatic semantic enrichment, you must have appropriate data access policies that grant permission to access index, pipeline, and model collection resources. For more information about data access policies, see Data access control for Amazon OpenSearch Serverless. For the procedure to configure a data access policy, see Creating data access policies (console).

Data access permissions


[
    {
        "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"
                ]
            }
        ],
        "Principal": [
            "arn:aws:iam::account_id:role/role_name"
        ]
    },
    {
        "Description": "Create model permission",
        "Rules": [
            {
                "ResourceType": "model",
                "Resource": ["model/collection_name/*"],
                "Permission": ["aoss:CreateMLResource"]
            }
        ],
        "Principal": [
            "arn:aws:iam::account_id:role/role_name"
        ]
    },
]

Network access permissions

To allow service APIs to access private collections, you must configure network policies that permit the required access between the service API and the collection. For more information about network policies, see Network access for Amazon OpenSearch Serverless .


[
   {
      "Description":"Enable automatic semantic enrichment in a private collection",
      "Rules":[
         {
            "ResourceType":"collection",
            "Resource":[
               "collection/collection_name"
            ]
         }
      ],
      "AllowFromPublic":false,
      "SourceServices":[
         "aoss.amazonaws.com"
      ],
   }
]

To configure network access permissions for a private collection

Sign in to the OpenSearch Service console at https://console.aws.amazon.com/aos/home.
In the left navigation, choose Network policies. Then do one of the following:
- Choose an existing policy name and choose Edit
- Choose Create network policy and configure the policy details
In the Access type area, choose Private (recommended), and then select Amazon service private access.
In the search field, choose Service, and then choose aoss.amazonaws.com.
In the Resource type area, select the Enable access to OpenSearch endpoint box.
For Search collection(s), or input specific prefix term(s), in the search field, select Collection Name. Then enter or select the name of the collections to associate with the network policy.
Choose Create for a new network policy or Update for an existing network policy.

Query Rewrites

Automatic semantic enrichment automatically converts your existing “match” queries to semantic search queries without requiring query modifications. If a match query is part of a compound query, the system traverses your query structure, finds match queries, and replaces them with neural sparse queries. Currently, the feature only supports replacing “match” queries, whether it’s a standalone query or part of a compound query. “multi_match” is not supported. In addition, the feature supports all compound queries to replace their nested match queries. Compound queries include: bool, boosting, constant_score, dis_max, function_score, and hybrid.

Limitations of automatic semantic enrichment

Automatic semantic search is most effective when applied to small-to-medium sized fields containing natural language content, such as movie titles, product descriptions, reviews, and summaries. Although semantic search enhances relevance for most use cases, it might not be optimal for certain scenarios. Consider following limitations when deciding whether to implement automatic semantic enrichment for your specific use case.

Very long documents – The current sparse model processes only the first 8,192 tokens of each document for English. For multilingual documents, it’s 512 tokens. For lengthy articles, consider implementing document chunking to ensure complete content processing.
Log analysis workloads – Semantic enrichment significantly increases index size, which might be unnecessary for log analysis where exact matching typically suffices. The additional semantic context rarely improves log search effectiveness enough to justify the increased storage requirements.

Pricing

OpenSearch Serverless bills automatic semantic enrichment based on OpenSearch Compute Units (OCUs) consumed during sparse vector generation at indexing time. You’re charged only for actual usage during indexing. You can monitor this consumption using the Amazon CloudWatch metric SemanticSearchOCU. For specific details about model token limits, volume throughput per OCU, and example of sample calculation, visit OpenSearch Service Pricing.

Javascript is disabled or is unavailable in your browser.

To use the Amazon Web Services Documentation, Javascript must be enabled. Please refer to your browser's Help pages for instructions.

Configuring permissions for collections

Creating collections