elasticsearch/docs/reference/elasticsearch/mapping-reference/semantic-text.md

navigation_title	mapped_pages
Semantic text	https://www.elastic.co/guide/en/elasticsearch/reference/current/semantic-text.html

Semantic text field type [semantic-text]

The semantic_text field type automatically generates embeddings for text content using an inference endpoint. Long passages are automatically chunked to smaller sections to enable the processing of larger corpuses of text.

The semantic_text field type specifies an inference endpoint identifier that will be used to generate embeddings. You can create the inference endpoint by using the Create {{infer}} API. This field type and the semantic query type make it simpler to perform semantic search on your data. The semantic_text field type may also be queried with match, sparse_vector or knn queries.

If you don’t specify an inference endpoint, the inference_id field defaults to .elser-2-elasticsearch, a preconfigured endpoint for the elasticsearch service.

Using semantic_text, you won’t need to specify how to generate embeddings for your data, or how to index it. The {{infer}} endpoint automatically determines the embedding generation, indexing, and query to use. Newly created indices with semantic_text fields using dense embeddings will be quantized to bbq_hnsw automatically.

If you use the preconfigured .elser-2-elasticsearch endpoint, you can set up semantic_text with the following API request:

PUT my-index-000001
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text"
      }
    }
  }
}

To use a custom {{infer}} endpoint instead of the default .elser-2-elasticsearch, you must Create {{infer}} API and specify its inference_id when setting up the semantic_text field type.

PUT my-index-000002
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "my-openai-endpoint" <1>
      }
    }
  }
}

1. The inference_id of the {{infer}} endpoint to use to generate embeddings.

The recommended way to use semantic_text is by having dedicated {{infer}} endpoints for ingestion and search. This ensures that search speed remains unaffected by ingestion workloads, and vice versa. After creating dedicated {{infer}} endpoints for both, you can reference them using the inference_id and search_inference_id parameters when setting up the index mapping for an index that uses the semantic_text field.

PUT my-index-000003
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "my-elser-endpoint-for-ingest",
        "search_inference_id": "my-elser-endpoint-for-search"
      }
    }
  }
}

Parameters for semantic_text fields [semantic-text-params]

inference_id

(Optional, string) {{infer-cap}} endpoint that will be used to generate embeddings for the field. By default, .elser-2-elasticsearch is used. This parameter cannot be updated. Use the Create {{infer}} API to create the endpoint. If search_inference_id is specified, the {{infer}} endpoint will only be used at index time.

search_inference_id

(Optional, string) {{infer-cap}} endpoint that will be used to generate embeddings at query time. You can update this parameter by using the Update mapping API. Use the Create {{infer}} API to create the endpoint. If not specified, the {{infer}} endpoint defined by inference_id will be used at both index and query time.

index_options

(Optional, object) Specifies the index options to override default values for the field. Currently, dense_vector index options are supported. For text embeddings, index_options may match any allowed dense_vector index options.

chunking_settings

(Optional, object) Settings for chunking text into smaller passages. If specified, these will override the chunking settings set in the {{infer-cap}} endpoint associated with inference_id. If chunking settings are updated, they will not be applied to existing documents until they are reindexed. To completely disable chunking, use the none chunking strategy.

Valid values for chunking_settings:

type

Indicates the type of chunking strategy to use. Valid values are none, word or sentence. Required.

max_chunk_size

The maximum number of words in a chunk. Required for word and sentence strategies.

overlap

The number of overlapping words allowed in chunks. This cannot be defined as more than half of the max_chunk_size. Required for word type chunking settings.

sentence_overlap

The number of overlapping sentences allowed in chunks. Valid values are 0 or 1. Required for sentence type chunking settings

::::{warning} When using the none chunking strategy, if the input exceeds the maximum token limit of the underlying model, some services (such as OpenAI) may return an error. In contrast, the elastic and elasticsearch services will automatically truncate the input to fit within the model's limit. ::::

{{infer-cap}} endpoint validation [infer-endpoint-validation]

The inference_id will not be validated when the mapping is created, but when documents are ingested into the index. When the first document is indexed, the inference_id will be used to generate underlying indexing structures for the field.

::::{warning} Removing an {{infer}} endpoint will cause ingestion of documents and semantic queries to fail on indices that define semantic_text fields with that {{infer}} endpoint as their inference_id. Trying to delete an {{infer}} endpoint that is used on a semantic_text field will result in an error. ::::

Text chunking [auto-text-chunking]

{{infer-cap}} endpoints have a limit on the amount of text they can process. To allow for large amounts of text to be used in semantic search, semantic_text automatically generates smaller passages if needed, called chunks.

Each chunk refers to a passage of the text and the corresponding embedding generated from it. When querying, the individual passages will be automatically searched for each document, and the most relevant passage will be used to compute a score.

For more details on chunking and how to configure chunking settings, see Configuring chunking in the Inference API documentation.

You can pre-chunk the input by sending it to Elasticsearch as an array of strings. Example:

PUT test-index
{
  "mappings": {
    "properties": {
      "my_semantic_field": {
        "type": "semantic_text",
        "chunking_settings": {
          "strategy": "none"    <1>
        }
      }
    }
  }
}

1. Disable chunking on my_semantic_field.

PUT test-index/_doc/1
{
    "my_semantic_field": ["my first chunk", "my second chunk", ...]    <1>
    ...
}

1. The text is pre-chunked and provided as an array of strings. Each element in the array represents a single chunk that will be sent directly to the inference service without further chunking.

Important considerations:

• When providing pre-chunked input, ensure that you set the chunking strategy to none to avoid additional processing.
• Each chunk should be sized carefully, staying within the token limit of the inference service and the underlying model.
• If a chunk exceeds the model's token limit, the behavior depends on the service:
  • Some services (such as OpenAI) will return an error.
  • Others (such as elastic and elasticsearch) will automatically truncate the input.

Refer to this tutorial to learn more about semantic search using semantic_text.

Extracting Relevant Fragments from Semantic Text [semantic-text-highlighting]

You can extract the most relevant fragments from a semantic text field by using the highlight parameter in the Search API.

POST test-index/_search
{
    "query": {
        "match": {
            "my_semantic_field": "Which country is Paris in?"
        }
    },
    "highlight": {
        "fields": {
            "my_semantic_field": {
                "number_of_fragments": 2,  <1>
                "order": "score"           <2>
            }
        }
    }
}

1. Specifies the maximum number of fragments to return.
2. Sorts highlighted fragments by score when set to score. By default, fragments will be output in the order they appear in the field (order: none).

Highlighting is supported on fields other than semantic_text. However, if you want to restrict highlighting to the semantic highlighter and return no fragments when the field is not of type semantic_text, you can explicitly enforce the semantic highlighter in the query:

PUT test-index
{
    "query": {
        "match": {
            "my_field": "Which country is Paris in?"
        }
    },
    "highlight": {
        "fields": {
            "my_field": {
                "type": "semantic",         <1>
                "number_of_fragments": 2,
                "order": "score"
            }
        }
    }
}

1. Ensures that highlighting is applied exclusively to semantic_text fields.

Customizing semantic_text indexing [custom-indexing]

semantic_text uses defaults for indexing data based on the {{infer}} endpoint specified. It enables you to quickstart your semantic search by providing automatic {{infer}} and a dedicated query so you don’t need to provide further details.

If you want to override those defaults and customize the embeddings that semantic_text indexes, you can do so by modifying <<semantic-text-params, parameters>>:

• Use index_options to specify alternate index options such as specific dense_vector quantization methods
• Use chunking_settings to override the chunking strategy associated with the {{infer}} endpoint, or completely disable chunking using the none type

Here is an example of how to set these parameters for a text embedding endpoint:

PUT my-index-000004
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "my-text-embedding-endpoint",
        "index_options": {
          "dense_vector": {
            "type": "int4_flat"
          }
        },
        "chunking_settings": {
          "type": "none"
        }
      }
    }
  }
}

Updates to semantic_text fields [update-script]

For indices containing semantic_text fields, updates that use scripts have the following behavior:

• Are supported through the Update API.
• Are not supported through the Bulk API and will fail. Even if the script targets non-semantic_text fields, the update will fail when the index contains a semantic_text field.

copy_to and multi-fields support [copy-to-support]

The semantic_text field type can serve as the target of copy_to fields, be part of a multi-field structure, or contain multi-fields internally. This means you can use a single field to collect the values of other fields for semantic search.

For example, the following mapping:

PUT test-index
{
    "mappings": {
        "properties": {
            "source_field": {
                "type": "text",
                "copy_to": "infer_field"
            },
            "infer_field": {
                "type": "semantic_text",
                "inference_id": ".elser-2-elasticsearch"
            }
        }
    }
}

can also be declared as multi-fields:

PUT test-index
{
    "mappings": {
        "properties": {
            "source_field": {
                "type": "text",
                "fields": {
                    "infer_field": {
                        "type": "semantic_text",
                        "inference_id": ".elser-2-elasticsearch"
                    }
                }
            }
        }
    }
}

Limitations [limitations]

semantic_text field types have the following limitations:

• semantic_text fields are not currently supported as elements of nested fields.
• semantic_text fields can’t currently be set as part of Dynamic templates.
• semantic_text fields are not supported with Cross-Cluster Search (CCS) or Cross-Cluster Replication (CCR).