Vertex AI Search and Conversation has been renamed to Vertex AI Agent Builder. We are in the process of updating content to reflect the new branding.

Get search summaries

This page shows how to use the API to get search summaries with your search results. It also explains the options that are available with search summaries. For unstructured and website data only.

For information about getting search results, see Get search results. For information about getting search summaries using the answer method (Preview), see Get answers and follow-ups.

For information about getting generative AI answers for your healthcare data queries, see Search using natural language query with generative AI answer.

Before you begin

Depending on the type of app you have, complete the following requirements:

For an unstructured search app: Turn on the Advanced LLM features.
For a website search app: Turn on the following features:
- Enterprise edition features.
- Advanced LLM features.
- Advanced website indexing. Requires domain verification.

Get a search summary

A search summary is a short summarization of the top one or more search results returned in a search response. The summary itself is taken from the extractive answers returned in the response. Therefore, to get a summary, you must also get extractive answers with your search results. For more information, see Get extractive answers (Preview).

The following image shows the summary when PDFs in a data store are queried with the summaryResultCount set to 5. The summary content can vary depending on the app configurations.

Query is quote Define operating
expenses endquote. The search summary section shows a summary extracted from the
top results. — **Figure 1.** Sample widget with a search summary.

Search summaries can include Markdown-formatted text. For this reason, consider using a Markdown parser in your application to render Markdown text.

To get a search summary, follow these steps:

Submit a search request that includes contentSearchSpec.summarySpec and specifies values for summaryResultCount and maxExtractiveAnswerCount. For more information about submitting a search request, see Get search results.

In the following example, summarySpec indicates that you want a search summary and that the summary should be generated from the top three search results.
```
"contentSearchSpec":
 {
   "summarySpec":
   {
     "summaryResultCount": 3
   },
   "extractiveContentSpec": { "maxExtractiveAnswerCount" : 1}
 }
```
- summaryResultCount: The number of top results to generate the search summary from. If the number of results returned is less than summaryResultCount, the summary is generated from all of the results.
- maxExtractiveAnswerCount: The number of extractive answers to return for each search result. The default value is 0 and the maximum is 1.

Get the summary from the search response. One summary property is returned in each response.

Here is an example of a summary returned at the end of a search response:

"summary":
{
  "summaryText": "BigQuery is Google Cloud's fully managed and completely
  serverless enterprise data warehouse. BigQuery supports all data types,
  works across clouds, and has built-in machine learning and business
  intelligence, all within a unified platform."
}

Generate summaries from semantic chunks

You can turn on use_semantic_chunks to generate summaries from the most relevant document chunks. Using semantic chunks for summary generation increases recall and retrieval compared the default behavior of using extractive answers.

When semantic chunking is turned on for summaries, the response returns the summary and the content of each chunk that the summary used.

To use semantic chunks for summary generation, follow these steps:

Submit a search request that includes contentSearchSpec.summarySpec and specifies "use_semantic_chunks": true. For more information about submitting a search request, see Get search results.

The following example of summarySpec indicates that you want a search summary that uses semantic chunks, how many results to include, and whether to include citations.
```
"contentSearchSpec":
 {
   "summarySpec":
   {
     "useSemanticChunks": SEMANTIC_CHUNK_BOOLEAN,
     "summaryResultCount": SUMMARY_RESULT_COUNT,
     "includeCitations": CITATIONS_BOOLEAN,
   }
 }
```
- SEMANTIC_CHUNK_BOOLEAN: A boolean that specifies whether to use semantic chunks to generate the search summary. If set to true, semantic chunks are used.
- SUMMARY_RESULT_COUNT: The number of top results to generate the search summary from. The maximum value is 10.
- CITATIONS_BOOLEAN: A boolean that specifies whether citations are returned. If you turned on chunk mode when you created your data store, then citations refer to chunks. Otherwise, citations refer to source documents. For more about chunk mode, see Parse and chunk documents.

Get the summary from the search response.

Here is an example of a search response that includes a summary that is generated from chunks and includes citations. The references part of the response contains the content of the chunks that the summary is generated from.

Response

{
  "results": [
    {
      "id": "123xyz",
      "document": {
        "name": "projects/exampleproject/locations/global/collections/default_collection/dataStores/exampledatastore/branches/0/documents/123xyz",
        "id": "123xyz",
        "derivedStructData": {
          "link": "gs://examplebucket/alphabet-investor-pdfs/2004_google_annual_report.pdf"
        }
      }
    }
  ],
  "totalSize": 8375,
  "attributionToken": "abcdefg",
  "nextPageToken": "hijklmnop",
  "guidedSearchResult": {},
  "summary": {
    "summaryText": "Google's search technology uses a combination of techniques to determine the importance of a web page independent of a particular search query and to determine the relevance of that page to a particular search query. [1]",
    "summaryWithMetadata": {
      "summary": "Google's search technology uses a combination of techniques to determine the importance of a web page independent of a particular search query and to determine the relevance of that page to a particular search query.",
      "citationMetadata": {
        "citations": [
          {
            "endIndex": "216",
            "sources": [
              {}
            ]
          }
        ]
      },
      "references": [
        {
          "document": "projects/exampleproject/locations/global/collections/default_collection/dataStores/exampledatastore/branches/0/documents/123xyz",
          "chunkContents": [
            {
              "content": "Groups contains more than 1 billion messages from Usenet Internet discussion groups dating back to 1981.The\ndiscussions in these groups cover a broad range of discourse and provide a comprehensive look at evolving\nviewpoints, debate and advice on many subjects.The new Google Groups adds in the ability to create your own\ngroups for you and your friends and an improved user interface.Google Mobile.Google Mobile offers people the ability to search and view both the "mobile web,"\nconsisting of pages created specifically for wireless devices, and the entire Google index of more than 8 billion\nweb pages.Google Mobile works on devices that support WAP, WAP 2.0, i-mode or j-sky mobile Internet\nprotocols.In addition, users can access a variety of information using Google SMS by typing a query to the\nGoogle shortcode.Google Mobile is available through many wireless and mobile phone services worldwide.",
              "pageIdentifier": "17"
            },
            {
              "content": "Google Labs is our playground for our engineers and for adventurous Google users.On Google\nLabs, we post product prototypes and solicit feedback on how the technology could be used or improved.Current Google Labs examples include:Google Personalized Search—provides customized search results based on an individual user's interests.Froogle Wireless—gives people the ability to search for product information from their mobile phones\nand other wireless devices.Google Maps—enables users to see maps, get directions, and find local businesses and services quickly\nand easily.Google Maps has several unique features, including draggable maps, integrated local search\nfrom Google Local, and keyboard shortcuts.Google Scholar—enables users to search specifically for scholarly literature, including peer-reviewed\npapers, theses, books, preprints, abstracts and technical reports from all broad areas of research.Google\nScholar can be used to find articles from a wide variety of academic publishers, professional societies,\npreprint repositories and universities, as well as scholarly articles available across the web.Google Suggest—guesses what you're typing and offers suggestions in real time.This is similar to\nGoogle's "Did you mean?"feature that offers alternative spellings for your query after you search, except\nthat it works in real time.",
              "pageIdentifier": "17"
            },
            {
              "content": "Groups contains more than 1 billion messages from Usenet Internet discussion groups dating back to 1981.The\ndiscussions in these groups cover a broad range of discourse and provide a comprehensive look at evolving\nviewpoints, debate and advice on many subjects.The new Google Groups adds in the ability to create your own\ngroups for you and your friends and an improved user interface.Google Mobile.Google Mobile offers people the ability to search and view both the "mobile web,"\nconsisting of pages created specifically for wireless devices, and the entire Google index of more than 8 billion\nweb pages.Google Mobile works on devices that support WAP, WAP 2.0, i-mode or j-sky mobile Internet\nprotocols.In addition, users can access a variety of information using Google SMS by typing a query to the\nGoogle shortcode.Google Mobile is available through many wireless and mobile phone services worldwide.\n\nGoogle Local.Google Local enables users to find relevant local businesses near a city, postal code, or specific\naddress.This service combines Yellow Page listings with information found on web pages, and plots their\nlocations on interactive maps.Google Print.Google Print brings information online that had previously not been available to web\nsearchers.Under this program, we enable a number of publishers to host their content and show their\npublications at the top of our search results.",
              "pageIdentifier": "17"
            },
            {
              "content": "Votes cast by important web pages with high PageRank weigh more heavily and are\nmore influential in deciding the PageRank of pages on the web.Text-Matching Techniques.Our technology employs text-matching techniques that compare search queries\nwith the content of web pages to help determine relevance.Our text-based scoring techniques do far more than\ncount the number of times a search term appears on a web page.For example, our technology determines the\nproximity of individual search terms to each other on a given web page, and prioritizes results that have the\nsearch terms near each other.Many other aspects of a page's content are factored into the equation, as is the\ncontent of pages that link to the page in question.By combining query independent measures such as PageRank\nwith our text-matching techniques, we are able to deliver search results that are relevant to what people are\ntrying to find.\n\nAdvertising Technology\nOur advertising program serves millions of relevant, targeted ads each day based on search terms people\n\nenter or content they view on the web.The key elements of our advertising technology include:\n\nGoogle AdWords Auction System.We use the Google AdWords auction system to enable advertisers to\nautomatically deliver relevant, targeted advertising.",
              "pageIdentifier": "21"
            },
            {
              "content": "Votes cast by important web pages with high PageRank weigh more heavily and are\nmore influential in deciding the PageRank of pages on the web.Text-Matching Techniques.Our technology employs text-matching techniques that compare search queries\nwith the content of web pages to help determine relevance.Our text-based scoring techniques do far more than\ncount the number of times a search term appears on a web page.For example, our technology determines the\nproximity of individual search terms to each other on a given web page, and prioritizes results that have the\nsearch terms near each other.Many other aspects of a page's content are factored into the equation, as is the\ncontent of pages that link to the page in question.By combining query independent measures such as PageRank\nwith our text-matching techniques, we are able to deliver search results that are relevant to what people are\ntrying to find.\n\nAdvertising Technology\nOur advertising program serves millions of relevant, targeted ads each day based on search terms people\n\nenter or content they view on the web.The key elements of our advertising technology include:",
              "pageIdentifier": "21"
            },
            {
              "content": "Google Maps—enables users to see maps, get directions, and find local businesses and services quickly\nand easily.Google Maps has several unique features, including draggable maps, integrated local search\nfrom Google Local, and keyboard shortcuts.Google Scholar—enables users to search specifically for scholarly literature, including peer-reviewed\npapers, theses, books, preprints, abstracts and technical reports from all broad areas of research.Google\nScholar can be used to find articles from a wide variety of academic publishers, professional societies,\npreprint repositories and universities, as well as scholarly articles available across the web.Google Suggest—guesses what you're typing and offers suggestions in real time.This is similar to\nGoogle's "Did you mean?"feature that offers alternative spellings for your query after you search, except\nthat it works in real time.Google Video—includes thousands of programs that play on our TVs every day.Google Video enables\nyou to search a growing archive of televised content—everything from sports to dinosaur\ndocumentaries to news shows.\n\n6",
              "pageIdentifier": "17"
            },
            {
              "content": "Every search query we process involves the automated\nexecution of an auction, resulting in our advertising system often processing hundreds of millions of auctions per\nday.To determine whether an ad is relevant to a particular query, this system weighs an advertiser's willingness\nto pay for prominence in the ad listings (the CPC) and interest from users in the ad as measured by the click\nthrough rate and other factors.If an ad does not attract user clicks, it moves to a less prominent position on the\npage, even if the advertiser offers to pay a high amount.This prevents advertisers with irrelevant ads from\n"squatting" in top positions to gain exposure.Conversely, more relevant, well-targeted ads that are clicked on\nfrequently move up in ranking, with no need for advertisers to increase their bids.Because we are paid only\nwhen users click on ads, the AdWords ranking system aligns our interests equally with those of our advertisers\nand our users.The more relevant and useful the ad, the better for our users, for our advertisers and for us.\n\nThe AdWords auction system also incorporates our AdWords discounter, which automatically lowers the\namount advertisers actually pay to the minimum needed to maintain their ad position.",
              "pageIdentifier": "21"
            },
            {
              "content": "Web Search Technology\nOur web search technology uses a combination of techniques to determine the importance of a web page\nindependent of a particular search query and to determine the relevance of that page to a particular search\nquery.We do not explain how we do ranking in great detail because some people try to manipulate our search\nresults for their own gain, rather than in an attempt to provide high-quality information to users.\n\nRanking Technology.One element of our technology for ranking web pages is called PageRank.While we\ndeveloped much of our ranking technology after Google was formed, PageRank was developed at Stanford\nUniversity with the involvement of our founders, and was therefore published as research.Most of our current\nranking technology is protected as trade-secret.PageRank is a query-independent technique for determining the\nimportance of web pages by looking at the link structure of the web.PageRank treats a link from web page A to\nweb page B as a "vote" by page A in favor of page B.The PageRank of a page is the sum of the PageRank of the\npages that link to it.The PageRank of a web page also depends on the importance (or PageRank) of the other\nweb pages casting the votes.",
              "pageIdentifier": "21"
            },
            {
              "content": "The Company recognizes as revenue the fees charged advertisers each time a user clicks on one of the text\nbased ads that are displayed next to the search results on Google web sites.Effective January 1, 2004, the\nCompany offered a single pricing structure to all of its advertisers based on the AdWords cost per click model.\n\nGoogle AdSense is the program through which the Company distributes its advertisers' text-based ads for\ndisplay on the web sites of the Google Network members.In accordance with Emerging Issues Task Force\n("EITF") Issue No. 99 19, Reporting Revenue Gross as a Principal Versus Net as an Agent, the Company recognizes\nas revenues the fees it receives from its advertisers.This revenue is reported gross primarily because the\nCompany is the primary obligor to its advertisers.\n\nThe Company generates fees from search services through a variety of contractual arrangements, which\ninclude per-query search fees and search service hosting fees.Revenues from set up and support fees and search\nservice hosting fees are recognized on a straight-line basis over the term of the contract, which is the expected\nperiod during which these services will be provided.The Company's policy is to recognize revenues from per\nquery search fees in the period queries are made and results are delivered.\n\nThe Company provides search services pursuant to certain AdSense agreements.",
              "pageIdentifier": "85"
            },
            {
              "content": "On Google Print pages, we provide links to book sellers that may\noffer the full versions of these publications for sale, and we show content-targeted ads that are served through\nthe Google AdSense program.Google Desktop Search.Google Desktop Search enables our users to perform a full text search on the\ncontents of their own computer, including email, files, instant messenger chats and web browser history.Users\ncan use this service to view web pages they have visited even when they are not online.Google Alerts.Google Alerts are email updates of the latest relevant Google results (web, news, etc.) based\non the user's choice of query or topic.Typical uses include monitoring a developing news story, keeping current\non a competitor or industry, getting the latest on a celebrity or event, or keeping tabs on a favorite sports team.Google Labs.Google Labs is our playground for our engineers and for adventurous Google users.On Google\nLabs, we post product prototypes and solicit feedback on how the technology could be used or improved.Current Google Labs examples include:Google Personalized Search—provides customized search results based on an individual user's interests.Froogle Wireless—gives people the ability to search for product information from their mobile phones\nand other wireless devices.",
              "pageIdentifier": "17"
            }
          ]
        }
      ]
    }
  }
}

Get citations

Citations, when specified, are numbers that are placed inline in a search summary. These numbers indicate from which search results specific sentences in the summary are taken.

To get citations, follow these steps:

Submit a search request that includes contentSearchSpec.summarySpec and specifies "includeCitations": true. For more information about submitting a search request, see Get search results.

In the following example, summarySpec indicates that you want a search summary, that the summary should be generated from the top three search results, and that citations should be included in the summary.
```
"contentSearchSpec":
 {
   "summarySpec":
   {
     "summaryResultCount": 3,
     "includeCitations": true
   },
   "extractiveContentSpec": { "maxExtractiveAnswerCount" : 1}
 }
```
- summaryResultCount: The number of top results to generate the search summary from. If the number of results returned is less than summaryResultCount, the summary is generated from all of the results. The maximum value is 5.
- includeCitations: A boolean that specifies whether citations are returned.
- maxExtractiveAnswerCount: The number of extractive answers to return for each search result. The default value is 0 and the maximum is 1.

Get the summary, with citations, from the search response. One summary property is returned in each response.

Here is an example of a summary, with citations and citation metadata, returned at the end of a search response:

"summary": {
 "summaryText": "BigQuery is Google Cloud's fully managed and completely
  serverless enterprise data warehouse [1]. BigQuery supports all data types,
  works across clouds, and has built-in machine learning and business
  intelligence, all within a unified platform [2, 3].",
 "summaryWithMetadata": {
   "summary": "BigQuery is Google Cloud's fully managed and completely
   serverless enterprise data warehouse. BigQuery supports all data types,
   works across clouds, and has built-in machine learning and business
   intelligence, all within a unified platform.",
   "citationMetadata": {
     "citations": [
       {
         "startIndex": "0",
         "endIndex": "101",
         "sources": [
           {
             "uri": "gs://example-dataset/html/6344007140738632642.html",
             "title": "About BigQuery",
             "id": "b6344007140738632642",
             "referenceIndex": "0"
           },
           {
             "uri": "gs://example-dataset/html/1365490014946172719.html",
             "title": "Google Cloud article",
             "id": "b1365490014946172719",
             "referenceIndex": "1"
           },
           {
             "uri": "gs://example-dataset/html/2687910668117268120.html",
             "title": "BigQuery document",
             "id": "a2687910668117268120",
             "referenceIndex": "2"
           }
         ]
       },
       {
         "startIndex": "103",
         "endIndex": "230",
         "sources": [
           {
             "referenceIndex": "0"
            },
           {
             "referenceIndex": "1"
           },
           {
             "referenceIndex": "2",
           }
         ]
       }
     ]
   },
   "references": [
   {
     "title": "Sports in the United States",
     "docName": "projects/123/locations/global/collections/default_collection/dataStores/ds-123/branches/0/documents/b6344007140738632642",
     "uri": "https://example.com/bigqueryA"
   },
   {
     "title": "Sports in the United States",
     "docName": "projects/123/locations/global/collections/default_collection/dataStores/ds-123/branches/0/documents/b1365490014946172719",
     "uri": "https://example.com/bigqueryB"
   },
   {
     "title": "Sports in the United States",
     "docName": "projects/123/locations/global/collections/default_collection/dataStores/ds-123/branches/0/documents/a268791066811726812",
     "uri": "https://example.com/bigqueryC"
   }
 ]
}
}

summaryText: The search summary, with citation numbers. The citation numbers refer to the returned search results and are 1-indexed. For example, [1] means that the sentence is attributed to the first search result. [2, 3] means that the sentence is attributed to both the second and third search results.
citations: For each sentence in the summary that has a citation, lists the metadata for that citation.
startIndex: Indicates the start of the sentence, measured in unicode bytes.
endIndex: Indicates the end of the sentence, measured in unicode bytes.
sources: Lists the referenceIndex for each source that was included in the sentence's citation. referenceIndex is the index number assigned to a source. The first source's referenceIndex isn't always explicitly returned in the response. Because referenceIndex is 0-indexed, the first source always has a referenceIndex of 0.
references: Lists metadata for each reference that was cited in the summary. Metadata includes title, docName, and uri.

Ignore adversarial queries

Adversarial queries include negative comments or are designed to generate unsafe, policy-violating output. You can specify that no search summaries should be returned for adversarial queries. When an adversarial query is ignored, the summaryText property contains boilerplate text indicating that no search summary is returned. Search documents are returned for adversarial queries even though search summaries are not.

To specify that no search summaries should be returned for adversarial queries, follow these steps:

Submit a search request that includes contentSearchSpec.summarySpec and specifies "ignoreAdversarialQuery": true. For more information about submitting a search request, see Get search results.

In the following example, summarySpec indicates that you want a search summary, that the summary should be generated from the top three search results, but that no summary should be returned for adversarial queries.
```
"contentSearchSpec":
 {
   "summarySpec":
   {
     "summaryResultCount": 3,
     "ignoreAdversarialQuery": true
   },
   "extractiveContentSpec": { "maxExtractiveAnswerCount" : 1}
 }
```
- summaryResultCount: The number of top results to generate the search summary from. If the number of results returned is less than summaryResultCount, the summary is generated from all of the results. The maximum value is 5.
- ignoreAdversarialQuery: A boolean that specifies that no search summaries should be returned for adversarial queries.
- maxExtractiveAnswerCount: The number of extractive answers to return for each search result. The default value is 0 and the maximum is 1.
See the summary property that is returned for an adversarial search request.

Here is an example:
```
"summary":
{
  "summaryText": "We do not have a summary for your query. Here are some
  search results.",
  "summarySkippedReasons": [
   "ADVERSARIAL_QUERY_IGNORED"
 ]
}
```
- summaryText: Boilerplate text indicating that no search summary is returned.
- summarySkippedReasons: An enumeration with values for summary-skipped reasons.

Ignore non-summary seeking queries

Non-summary seeking queries return results that are not suitable for summarization. For example, "why is the sky blue" and "Who is the best soccer player in the world?" are summary-seeking queries, but "SFO airport" and "world cup 2026" are not. They are most likely navigational queries. You can specify that no search summaries should be returned for non-summary seeking queries. Search documents are returned for non-summary seeking queries even though search summaries are not.

To specify that no search summaries should be returned for non-summary seeking queries, follow these steps:

Submit a search request that includes contentSearchSpec.summarySpec and specifies "ignoreNonSummarySeekingQuery": true. For more information about submitting a search request, see Get search results.

In the following example, summarySpec indicates that you want a search summary, the summary should be generated from the top three search results, but that no summary should be returned for non-summary seeking queries.
```
"contentSearchSpec":
 {
   "summarySpec":
   {
     "summaryResultCount": 3,
     "ignoreNonSummarySeekingQuery": true
   },
   "extractiveContentSpec": { "maxExtractiveAnswerCount" : 1}
 }
```
- summaryResultCount: The number of top results to generate the search summary from. If the number of results returned is less than summaryResultCount, the summary is generated from all of the results. The maximum value is 5.
- ignoreNonSummarySeekingQuery: A boolean that specifies that no search summaries should be returned for non-summary seeking queries.
- maxExtractiveAnswerCount: The number of extractive answers to return for each search result. The default value is 0 and the maximum is 1.
See the summary property that is returned for a non-summary seeking search request.

Here is an example:
```
"summary":
{
  "summaryText": "We do not have a summary for your query. Here are some
  search results.",
  "summarySkippedReasons": [
    "NON_SUMMARY_SEEKING_QUERY_IGNORED"
 ]
}
```
- summaryText: Boilerplate text indicating that no search summary is returned.
- summarySkippedReasons: An enumeration with values for summary-skipped reasons.

Get customized summaries

You can get customized summaries by providing natural-language instructions. This feature is available for the search widget by using the console, and for the API by entering instructions in the modelPromptSpec.preamble field.

You can request customizations such as length, level of detail, style of output (such as "simple"), language of output, focus of answer, and format (such as tables, bullets, and XML).

For example, for the query What are the benefits of bigquery?, you could enter customization instructions like in the following examples:

Example instruction: please show the answer format in an ordered list

Example answer:

The benefits of BigQuery are:

1. Increased business productivity
2. IT staff productivity
3. Reduced infrastructure costs
4. Loading and exporting data are free

Example instruction: give a very detailed answer

Example answer:

BigQuery is a serverless, highly scalable, and cost-effective cloud data warehouse that enables businesses to analyze all their data very quickly. It is a fully managed service that takes care of provisioning, scaling, and managing the underlying infrastructure, so you can focus on analyzing your data. BigQuery is also integrated with other Google Cloud services, such as Google Kubernetes Engine, Cloud Data Fusion, and Cloud Dataproc, making it easy to build and deploy data pipelines.

Here are some of the benefits of using BigQuery:

* **Fast and scalable:** BigQuery can process petabytes of data very quickly, and it can scale to handle even the most demanding workloads. * **Cost-effective:** BigQuery is a very cost-effective way to store and analyze data. You only pay for the data that you use, and there are no upfront costs or commitments. * **Secure:** BigQuery is a secure platform that meets the needs of even the most security-conscious organizations. * **Easy to use:** BigQuery is easy to use, even for non-technical users. It has a simple and intuitive user interface, and it supports a variety of data sources. * **Integrated with other Google Cloud services:** BigQuery is integrated with other Google Cloud services, making it easy to build and deploy data pipelines.

If you are looking for a fast, scalable, and cost-effective way to analyze your data, then BigQuery is a great option.

Best practices for customized summaries

If you plan to use this feature, do the following:

Request only one customization at a time. Don't combine customizations—for example, requesting an HTML table in French.
Google recommends that you impose limits on what customizations your end users can request—for example, by offering a selector with a set of predefined customizations.

Customize summaries

You can get customized summaries for only the search widget by using the console or, for any search request, by using the API.

To get a customized summary, follow these steps:

Console

In the Google Cloud console, go to the Agent Builder page.

Agent Builder
Click the name of the app that you want to edit.
Go to Configurations > UI.
Make sure your search widget's Search type is set to Search with an answer or Search with follow-ups. This feature isn't available if Search is selected.
Turn on Enable summary customization.
To enter summary instructions, do one of the following:
- Enter free-form instructions: Enter your own natural language instructions in the Preamble field.
- Use template instructions: Click Replace with a template and select one of the predefined template instructions. The predefined template appears in the Preamble field after you select it.
Test the customized summary generation for your app by searching in the Preview pane.
To reset to the last saved set of instructions, click Reset preamble.
To save your settings to the widget, click Save and publish.

REST

Submit a search request that includes contentSearchSpec.summarySpec and specifies the customization instruction in modelPromptSpec.preamble. For more information about submitting a search request, see Get search results.

In the following example, summarySpec indicates that you want a search summary, the summary should be generated from the top three search results, and the summary should be customized as though it is being explained to a 10-year-old.
```
"contentSearchSpec":
  {
    "summarySpec":
    {
      "summaryResultCount": 3,
      "modelPromptSpec":
      {
        "preamble": "explain like you would to a ten year old"
      }
    }
  }
```
- summaryResultCount: The number of top results to generate the search summary from. If the number of results returned is less than summaryResultCount, the summary is generated from all of the results. The maximum value is 5.
- preamble: The instruction for customization.

Get the customized summary from the search response.

Here is an example of a customized summary that is returned:

"summary":
{
  "summaryText": "BigQuery is a serverless data warehouse that helps you
  analyze all your data very quickly. It's very easy to use and you don't
  need to worry about managing servers or infrastructure. BigQuery is also
  very scalable, so you can analyze large datasets without any problems."
}

summaryText: The customized search summary.

Specify the summarization model

You can specify the model that you want to use to generate summaries.

You can specify a stable, GA model that is based on either the text-bison model or gemini-pro model, or you can specify a preview model that is based on the gemini-pro model. For available model versions, see Answer generation model versions and lifecycle.

To change the model version:

Submit a search request that includes ContentSearchSpec.SummarySpec.ModelSpec to specify the model version.
```
"contentSearchSpec": {
  "summarySpec": {
    "modelSpec": {
      "version": "MODEL_VERSION"
     }
   }
 }
```
- MODEL_VERSION: Specifies which model to use to generate summaries. Supported values are:
  - stable: string. Default specification when no value is specified. stable points to a GA model version that's fine-tuned for answer generation. Which model stable points to changes as new GA model versions are released and previous model versions are discontinued. For the up-to-date version that stable points to, see Answer generation model versions and lifecycle.
  - preview: string. preview points to the latest Gemini model for question and answering. For more information about Gemini, see Overview of multimodal models.
  - To specify a certain model version instead, enter the version name, such as text-bison@001/answer_gen/v1. For supported versions, see Answer generation model versions and lifecycle.

For example, the following search request specifies preview as the model version:

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
"https://discoveryengine.googleapis.com/v1/projects/exampleproject/locations/global/collections/default_collection/dataStores/exampledatastore/servingConfigs/default_search:search" \
-d '{
  "query": "what is bigquery",
  "contentSearchSpec": {
    "summarySpec": {
      "modelSpec": {
        "version": "preview"
      }
    }
  }
}'

Limitations of search summaries

You might encounter the following limitations when using search summaries:

Because LLMs are used to generate search summaries and citations, the limitations of LLMs also apply to Vertex AI Search summaries.

For general information about these LLM limitations, see PaLM API limitations in the Vertex AI documentation.
Search queries that require complex logical or analytical reasoning or understanding of the world can lead to search summaries that contain incorrect information (hallucinations) or information that is not present in the unstructured or website data.
Some statements in the search summary might not contain a citation and, in rare cases, citations might be incorrectly attributed to a statement.
Complex documents might be incorrectly parsed by the LLM. In this case, the summary might be incomplete or incorrect.
Because customization instructions are in natural language, adherence to instructions can't be guaranteed for all requests.