Skip to content

Update doc for batch semantic highlighting #11137

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Oct 2, 2025
Merged

Update doc for batch semantic highlighting #11137

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
b091423
Update doc for batch semantic highlighting
junqiu-lei Sep 30, 2025
9e1d576
Doc review
kolchfa-aws Oct 1, 2025
be014fb
Update _search-plugins/searching-data/highlight.md
kolchfa-aws Oct 1, 2025
ef86d02
Update _search-plugins/searching-data/highlight.md
kolchfa-aws Oct 1, 2025
7464a71
Update stats api for batch semantic highlighting
junqiu-lei Oct 1, 2025
0f1a4d1
Formatting fixes
kolchfa-aws Oct 1, 2025
09ebfaf
clear message for remote model id
junqiu-lei Oct 1, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
207 changes: 133 additions & 74 deletions _search-plugins/searching-data/highlight.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,16 +172,142 @@
The `plain` highlighter is based on the standard Lucene highlighter. It requires the highlighted fields to be stored either individually or in the `_source` field. The `plain` highlighter mirrors the query matching logic, in particular word importance and positions in phrase queries. It works for most use cases but may be slow for large fields because it has to reanalyze the text to be highlighted.

### The `semantic` highlighter
**Introduced 3.0**
{: .label .label-purple }

The `semantic` highlighter uses machine learning (ML) models to identify and highlight the most semantically relevant sentences or passages within a text field, based on the query's meaning. This goes beyond traditional lexical matching offered by other highlighters. It does not rely on offsets from postings or term vectors but instead uses a deployed ML model (specified by the `model_id`) to perform inference on the field content. This approach allows you to highlight contextually relevant text even when exact terms don't match the query. Highlighting is performed at the sentence level.

Before using the `semantic` highlighter, you must configure and deploy a sentence highlighting model. For more information about using ML models in OpenSearch, see [Integrating ML models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/integrating-ml-models/). For information about OpenSearch-provided sentence highlighting models, see [Semantic sentence highlighting models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/pretrained-models/#semantic-sentence-highlighting-models).
The `semantic` highlighter supports two processing modes:

- **Single inference mode (default)**: Processes each document individually using one ML inference call per document. Supports both local and externally hosted models.
- [**Batch inference mode**](#batch-inference-mode): Processes all documents in a single ML inference call, significantly improving performance for multi-document results.

For production environments, we recommend using externally hosted models with batch inference enabled for optimal performance and scalability.
{: .tip}

Before using the `semantic` highlighter, you must configure and deploy a sentence highlighting model. For more information about using ML models in OpenSearch, see [Integrating ML models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/integrating-ml-models/). For information about OpenSearch-provided sentence highlighting models, see [Semantic sentence highlighting models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/pretrained-models/#semantic-sentence-highlighting-models).
{: .note}

To use the `semantic` highlighter, you must specify a `model_id` in the `highlight.options` object. The model determines which parts of the text are semantically similar to the query.
### Basic usage (single inference mode)

To use the `semantic` highlighter, set the `type` to `semantic` in the `fields` object and provide the `model_id` of the deployed sentence transformer or question-answering model within the global `highlight.options` object. The following example uses a `neural` query to find documents related to "treatments for neurodegenerative diseases" and then applies semantic highlighting using the specified `sentence_model_id`:

Check failure on line 193 in _search-plugins/searching-data/highlight.md

View workflow job for this annotation

GitHub Actions / style-job 

[vale] reported by reviewdog 🐶
[OpenSearch.Spelling] Error: neurodegenerative. If you are referencing a setting, variable, format, function, or repository, surround it with tic marks.

Raw Output:
{"message": "[OpenSearch.Spelling] Error: neurodegenerative. If you are referencing a setting, variable, format, function, or repository, surround it with tic marks.", "location": {"path": "_search-plugins/searching-data/highlight.md", "range": {"start": {"line": 193, "column": 315}}}, "severity": "ERROR"}

```json
POST /neural-search-index/_search
{
"_source": {
"excludes": ["text_embedding"]
},
"query": {
"neural": {
"text_embedding": {
"query_text": "treatments for neurodegenerative diseases",
"model_id": "your-text-embedding-model-id",
"k": 5
}
}
},
"highlight": {
"fields": {
"text": {
"type": "semantic"
}
},
"options": {
"model_id": "your-sentence-model-id"
}
}
}
```
{% include copy-curl.html %}

The response includes a `highlight` object for each hit, indicating the most semantically relevant sentence by emphasizing it with <em> tags. Note that model IDs are placeholders:

```json
{
"took": 628,
"timed_out": false,
"_shards": { ... },
"hits": {
"total": { "value": 5, "relation": "eq" },
"max_score": 0.4841726,
"hits": [
{
"_index": "neural-search-index",
"_id": "srL7G5YBmDiZSe-G2pDc",
"_score": 0.4841726,
"_source": {
"text": "Alzheimer's disease is a progressive neurodegenerative disorder characterized by accumulation of amyloid-beta plaques and neurofibrillary tangles in the brain. Early symptoms include short-term memory impairment, followed by language difficulties, disorientation, and behavioral changes. While traditional treatments such as cholinesterase inhibitors and memantine provide modest symptomatic relief, they do not alter disease progression. Recent clinical trials investigating monoclonal antibodies targeting amyloid-beta, including aducanumab, lecanemab, and donanemab, have shown promise in reducing plaque burden and slowing cognitive decline. Early diagnosis using biomarkers such as cerebrospinal fluid analysis and PET imaging may facilitate timely intervention and improved outcomes."
},
"highlight": {
"text": [
"Alzheimer's disease is a progressive neurodegenerative disorder ... <em>Recent clinical trials investigating monoclonal antibodies targeting amyloid-beta, including aducanumab, lecanemab, and donanemab, have shown promise in reducing plaque burden and slowing cognitive decline.</em> Early diagnosis using biomarkers ..."
]
}
},
// ... other hits with highlighted sentences ...
]
}
}
```

For a step-by-step guide, see the [semantic highlighting tutorial]({{site.url}}{{site.baseurl}}/tutorials/vector-search/semantic-highlighting-tutorial/).

### Batch inference mode
**Introduced 3.3**
{: .label .label-purple }

To improve performance when highlighting multiple documents, enable batch inference mode. Batch inference mode processes all matching documents in a single ML model inference call, reducing latency and improving throughput compared to single inference mode (which makes one call per document).

Batch inference mode requires an externally hosted model with batch processing capabilities. Local models do not support batch inference. For information about externally hosted models, see [Connecting to externally hosted models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/remote-models/).
{: .note}

First, configure the cluster setting (one-time setup):

```json
PUT _cluster/settings
{
"persistent": {
"search.pipeline.enabled_system_generated_factories": ["semantic-highlighter"]
}
}
```
{% include copy-curl.html %}

Then set `batch_inference` to `true` in `highlight.options`:

```json
POST /neural-search-index/_search
{
"_source": {
"excludes": ["text_embedding"]
},
"query": {
"neural": {
"text_embedding": {
"query_text": "treatments for neurodegenerative diseases",
"model_id": "your-text-embedding-model-id",
"k": 5
}
}
},
"highlight": {
"fields": {
"text": {
"type": "semantic"
}
},
"options": {
"model_id": "your-remote-sentence-model-id",
"batch_inference": true
}
}
}
```
{% include copy-curl.html %}

Batch inference provides responses in the same format as single inference.

## Highlighting options

The following table describes the highlighting options you can specify on a global or field level. Field-level settings override global settings.
Expand Down Expand Up @@ -210,8 +336,10 @@
`no_match_size` | Specifies the number of characters, starting from the beginning of the field, to return if there are no matching fragments to highlight. Default is 0.
`phrase_limit` | The number of matching phrases in a document that are considered. Limits the number of phrases to be analyzed by the `fvh` highlighter in order to avoid consuming a lot of memory. If `matched_fields` are used, `phrase_limit` specifies the number of phrases for each matched field. A higher `phrase_limit` leads to increased query time and more memory consumption. Valid only for the `fvh` highlighter. Default is 256.
`max_analyzer_offset` | Specifies the maximum number of characters to be analyzed by a highlight request. The remaining text will not be processed. If the text to be highlighted exceeds this offset, then an empty highlight is returned. The maximum number of characters that will be analyzed for a highlight request is defined by `index.highlight.max_analyzed_offset`. When this limit is reached, an error is returned. Set the `max_analyzer_offset` to a lower value than `index.highlight.max_analyzed_offset` to avoid the error.
`options` | A global object containing highlighter-specific options.
`options.model_id` | The ID of the deployed ML model to use for highlighting. Required. Valid only for the `semantic` highlighter.
`options` | A global object containing highlighter-specific options.
`options.batch_inference` | When set to `true`, enables batch inference mode for semantic highlighting, processing all documents in a single ML inference call instead of one call per document. Requires an externally hosted model with batch processing capabilities (local models are not supported) and the system processor factory to be enabled via cluster setting: `search.pipeline.enabled_system_generated_factories: ["semantic-highlighter"]`. Default is `false`. Valid only for the `semantic` highlighter.

Check warning on line 340 in _search-plugins/searching-data/highlight.md

View workflow job for this annotation

GitHub Actions / style-job 

[vale] reported by reviewdog 🐶
[OpenSearch.LatinismsSubstitution] Use 'using, through, by accessing, or by choosing' instead of 'via'.

Raw Output:
{"message": "[OpenSearch.LatinismsSubstitution] Use 'using, through, by accessing, or by choosing' instead of 'via'.", "location": {"path": "_search-plugins/searching-data/highlight.md", "range": {"start": {"line": 340, "column": 345}}}, "severity": "WARNING"}
`options.max_inference_batch_size` | Specifies the maximum number of documents to include in each inference request to the model server when using batch inference mode. If the number of documents to process exceeds this value, the documents will be processed iteratively in batches of this size. Default is `100`. Valid only for the `semantic` highlighter with `batch_inference` enabled.
`options.model_id` | The ID of the deployed ML model to use for highlighting. Required for the `semantic` highlighter. When `options.batch_inference` is set to `true`, the model must be an externally hosted model with batch processing capabilities.

The unified highlighter's sentence scanner splits sentences larger than `fragment_size` at the first word boundary after `fragment_size` is reached. To return whole sentences without splitting them, set `fragment_size` to 0.
{: .note}
Expand Down Expand Up @@ -973,79 +1101,10 @@
}
```

## Using the `semantic` highlighter

The `semantic` highlighter uses the specified ML model to find passages in text that are semantically relevant to the search query, even if there are no exact keyword matches. Highlighting occurs at the sentence level.

To use the `semantic` highlighter, set the `type` to `semantic` in the `fields` object and provide the `model_id` of the deployed sentence transformer or question-answering model within the global `highlight.options` object.

The following example uses a `neural` query to find documents related to "treatments for neurodegenerative diseases" and then applies semantic highlighting using the specified `sentence_model_id`:

```json
POST neural-search-index/_search
{
"_source": {
"excludes": ["text_embedding"]
},
"query": {
"neural": {
"text_embedding": {
"query_text": "treatments for neurodegenerative diseases",
"model_id": "your-text-embedding-model-id",
"k": 5
}
}
},
"highlight": {
"fields": {
"text": {
"type": "semantic"
}
},
"options": {
"model_id": "your-sentence-model-id"
}
}
}
```
{% include copy-curl.html %}

The response includes a `highlight` object for each hit, indicating the most semantically relevant sentence by emphasizing it with <em> tags. Note that model IDs are placeholders:

```json
{
"took": 628,
"timed_out": false,
"_shards": { ... },
"hits": {
"total": { "value": 5, "relation": "eq" },
"max_score": 0.4841726,
"hits": [
{
"_index": "neural-search-index",
"_id": "srL7G5YBmDiZSe-G2pDc",
"_score": 0.4841726,
"_source": {
"text": "Alzheimer's disease is a progressive neurodegenerative disorder characterized by accumulation of amyloid-beta plaques and neurofibrillary tangles in the brain. Early symptoms include short-term memory impairment, followed by language difficulties, disorientation, and behavioral changes. While traditional treatments such as cholinesterase inhibitors and memantine provide modest symptomatic relief, they do not alter disease progression. Recent clinical trials investigating monoclonal antibodies targeting amyloid-beta, including aducanumab, lecanemab, and donanemab, have shown promise in reducing plaque burden and slowing cognitive decline. Early diagnosis using biomarkers such as cerebrospinal fluid analysis and PET imaging may facilitate timely intervention and improved outcomes."
},
"highlight": {
"text": [
"Alzheimer's disease is a progressive neurodegenerative disorder ... <em>Recent clinical trials investigating monoclonal antibodies targeting amyloid-beta, including aducanumab, lecanemab, and donanemab, have shown promise in reducing plaque burden and slowing cognitive decline.</em> Early diagnosis using biomarkers ..."
]
}
},
// ... other hits with highlighted sentences ...
]
}
}
```

The highlighted fragments in the example response have been truncated for brevity. The `semantic` highlighter returns the full sentence containing the most relevant passage.

## Query limitations

Note the following limitations:

- When extracting terms to highlight, highlighters don't reflect the Boolean logic of a query. Therefore, for some complex Boolean queries, such as nested Boolean queries and queries using `minimum_should_match`, OpenSearch may highlight terms that don't correspond to query matches.
- The `fvh` highlighter does not support span queries.
- The `semantic` highlighter requires a deployed ML model specified by `model_id` in the `highlight.options`. It does not use traditional offset methods (postings, term vectors) and relies solely on model inference.
- The `semantic` highlighter requires a deployed ML model specified by `model_id` in the `highlight.options`. It does not use traditional offset methods (postings, term vectors) and relies solely on model inference. For batch inference mode (`batch_inference: true`), you must use an externally hosted model with batch processing capabilities.
13 changes: 11 additions & 2 deletions _tutorials/vector-search/semantic-highlighting-tutorial.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,9 @@ POST /_plugins/_ml/models/_register?deploy=true

Monitor the deployment status using the Tasks API. Note the semantic highlighting model ID; you'll use it in the following steps.

For production environments, consider using an externally hosted model instead of a locally deployed model. Externally hosted models offer better scalability, resource isolation, and support for advanced features like batch inference. For information about deploying externally hosted models, see [Connecting to externally hosted models]({{site.url}}{{site.baseurl}}/ml-commons-plugin/remote-models/).
{: .tip}

## Step 3 (Optional): Configure an ingest pipeline

To automatically generate embeddings during indexing, create an [ingest pipeline]({{site.url}}{{site.baseurl}}/ingest-pipelines/):
Expand Down Expand Up @@ -213,8 +216,6 @@ POST /neural-search-index/_search
```
{% include copy-curl.html %}

## Step 6: Interpret the results

The search results include a `highlight` object within each hit. The specified `text` field in the `highlight` object contains the original text, with the most semantically relevant sentences wrapped in `<em>` tags by default:

```json
Expand Down Expand Up @@ -267,3 +268,11 @@ The search results include a `highlight` object within each hit. The specified `
```

The `semantic` highlighter identifies the sentence determined by the model to be semantically relevant to the query ("treatments for neurodegenerative diseases") within the context of each retrieved document. You can customize the highlight tags using the `pre_tags` and `post_tags` parameters if needed. For more information, see [Changing the highlighting tags]({{site.url}}{{site.baseurl}}/search-plugins/searching-data/highlight/#changing-the-highlighting-tags).

### Using batch inference mode for highlighting

For improved performance when highlighting multiple documents in production environments, consider enabling batch inference mode. This processes all documents in a single ML inference call instead of one call per document. For more information, see [Batch inference mode]({{site.url}}{{site.baseurl}}/search-plugins/searching-data/highlight#batch-inference-mode).

## Next steps

For more information about semantic highlighting options and configuration, see [Using the semantic highlighter]({{site.url}}{{site.baseurl}}/search-plugins/searching-data/highlight#the-semantic-highlighter).
Loading