Agent Search release notes

Agent Search: Gemini 3.5 Flash answer generation

You can generate answers with the Gemini 3.5 Flash model.

For more information, see Answer generation model versions and lifecycle and Gemini 3.5 Flash.

Agent Search: Decrease thresholds for configurable pricing

You can decrease the storage size and queries per minute (QPM) subscription thresholds for configurable pricing. Previously, you could only increase these thresholds.

Decreased thresholds take effect at the start of the next billing cycle.

For more information, see Modify subscription thresholds. This feature is generally available (GA).

Agent Search: Prefix and partial matching for filtering search queries (Preview)

You can configure schema fields to support prefix matching and partial matching in filter expressions:

• Prefix matching lets you filter search results based on whether a field value starts with a specific string.

• Partial matching lets you filter results based on whether the query contains some of the words in the field value. Partial matching doesn't require a perfect match like the ANY operator does.

This feature is in Public Preview. For more information, see Configure field settings.

Agent Search: EXISTS filter for filtering search queries (Preview)

You can use the EXISTS filter to filter search results for documents. Specifying EXISTS for a field means that a document can only be returned in a search request if the field has a value and that value is not the default. This filter is available for custom search and for media search. Use EXISTS with other filters such as ANY and IN to create expressions to scope the documents that can be returned in a search query.

This feature is in Public Preview. For more information, see Filter custom search for structured or unstructured data, Filter website search, and Filter media search.

Agent Search: Table and image annotation in layout parser

The table annotation and image annotation features of the layout parser are generally available (GA).

You can ask the layout parser to annotate images or tables with a descriptive block of text describing the information in the image or table. The annotation can then be used as a source in a generated answer. For more information, see Layout parser.

Agent Search: Agent Search for healthcare is deprecated

Agent Search for healthcare is deprecated.

For a comprehensive, managed solution, consider building custom search apps on Agent Search. Or, if you require fine-grained control over the underlying retrieval mechanisms and are prepared for a more customer-managed integration, use Agent Retrieval (formerly known as Vector Search 2.0).

Agent Search: Weight searchable fields (Preview)

You can specify a weight for searchable fields in your schema to indicate their relative importance in search results.

This feature is in Public Preview. For more information, see Weight searchable fields.

Agent Search: Stream answers using agentic retrieval (GA with allowlist)

You can use agentic retrieval with the streaming answer method.

Agentic retrieval can return better results as compared to the standard streaming answer method. This is because agentic retrieval can do multi-pass searches across multiple data stores. The agent plans and executes searches sequentially, choosing the best tools, such as Google Search and Google Maps, for each step. Agentic retrieval also enables multi-turn search queries (follow-up questions) on blended search apps.

For more information, see Stream answers using agentic retrieval.

This feature is GA with an allowlist, available for select customers.

Agent Search: MCP server (GA)

Agent Search has a Model Context Protocol (MCP) server hosted at the following endpoint: https://discoveryengine.googleapis.com/mcp

This feature is generally available (GA). For more information, see MCP Reference: discoveryengine.googleapis.com.

Agent Search: Dense reciprocal rank for custom ranking

You can use the dense reciprocal rank transformation function, drr, to customize search result ranking. It's an improvement on the reciprocal rank function, rr. Using the dense reciprocal rank function leads to higher quality ranking when there are duplicate signal values.

Duplicate signal values are more common when the ranking formula contains the following types of signal:

• The boosting_factor signal
• The geo_distance() function signal
• Categorical and integer custom signals

This feature is generally available (GA). For more information, see Customize search results ranking.

Agent Search: Geodistance function for custom ranking (GA)

The geo_distance function can be used in custom ranking formulas to calculate the distance between a source location and a destination location. The function supports query locations extracted from natural language, explicitly provided coordinates, and addresses.

This feature is generally available (GA). For more information, see Custom ranking: Geodistance—a derived signal.

Agent Search: Filter searches by document-level relevance (GA)

When searching in your Agent Search app, you can specify document-level relevance filters so that only the documents that meet the filter threshold are returned as results.

You can specify either the relevance threshold or semantic-relevance threshold to filter documents by relevance based on keyword and semantic search similarity.

This feature is Generally Available (GA). For more information, see Filter searches by document-level relevance.

Agent Search: Renamed from Vertex AI Search

The Vertex AI Search product has been renamed as Agent Search in the following contexts:

• The documentation set. See What is Agent Search?

What has not changed:

• The user interface in the Google Cloud console is still referred to as Vertex AI Search and AI Applications. See Vertex AI Search.

• The APIs still use the Discovery Engine API endpoints. See APIs and reference.

Despite the rebrand, the product functionality remains the same.

Vertex AI Search: Gemini 3.1 Pro and Gemini 3 Flash for answer generation (Preview)

You can generate answers with the Gemini 3.1 Pro (Preview) and Gemini 3 Flash (Preview) models.

For more information, see Answer generation model versions and lifecycle, Gemini 3.1 Pro, and Gemini 3 Flash.

Vertex AI Search: Gemini 3 Pro (Preview) for answer generation discontinued

The Gemini 3 Pro (Preview) model has been discontinued and is no longer available for answer generation. If you have been using that model, upgrade to the Gemini 3.1 Pro (Preview) model.

For information about available models, see Answer generation model versions and lifecycle.

Vertex AI Search: Change the pricing model for a project

There are two pricing models for apps and data stores: the general model (pay-as-you-go consumption-based) and the configurable subscription model (a monthly subscription).

You can switch from configurable pricing to general pricing for a project if all its data stores and apps use the general model. For more information, see Disable configurable pricing for project.

Vertex AI Search: Gemini 3 Pro for answer generation (Preview)

You can generate answers with the Gemini 3 Pro model.

For more information, see Answer generation model versions and lifecycle and Gemini 3 Pro.

Vertex AI Search: MCP server (Preview)

Vertex AI Search has a Model Context Protocol (MCP) server hosted at the following endpoint: https://discoveryengine.googleapis.com/mcp

This feature is in Public Preview. For more information, see MCP Reference: discoveryengine.googleapis.com.

Vertex AI Search: Change the pricing model for apps and data stores

There are two pricing models for apps and data stores: the general model (pay-as-you-go consumption-based) and the configurable subscription model (a monthly subscription).

If you have enabled configurable pricing for your project, you can change the pricing model for apps and data stores from configurable to general, as well as from general to configurable. See Disable configurable pricing.

Vertex AI Search: Filter searches by document-level relevance (Preview)

When searching in your Vertex AI Search app, you can specify document-level relevance filters so that only the documents that meet the filter threshold are returned as results.

You can specify either the relevance threshold or semantic-relevance threshold to filter documents by relevance based on keyword and semantic search similarity.

This feature is in Public Preview. For more information, see Filter searches by document-level relevance.

Vertex AI Search: Configurable pricing for custom search (GA)

Configurable pricing is available for custom search apps and data stores.

Configurable pricing offers a flexible alternative to the default pay-as-you-go model. Configurable pricing lets you to choose a subscription that fits your needs. There are two subscriptions, one for storage and one for search queries, plus add-ons that let you add more features according to your needs.

For storage, the minimum subscription is 50 GiB/month and the available add-on is semantic embedding. For search queries, the minimum subscription is 1000 queries per minute per project, and the available add-ons are semantic query, KPI & personalization, and AI overview.

After you set up configurable pricing for a project, apply configurable pricing to apps and data stores as needed. In a project, you can have some apps and data store using the general pricing and others using configurable pricing.

Configurable pricing is generally available (GA). For more information, see Enable configurable pricing for custom search and Vertex AI Search pricing.

Vertex AI Search: Natural language query filters (GA)

For queries on structured data stores, the natural language queries can be reformulated as filters and a residual query. For example, "Find a coffee shop serving banana bread" becomes "query": "banana bread", "filter": "type": ANY("cafe").

The natural-language query understanding feature only applies to custom search apps attached to a single, structured, data store.

By default, a hard restriction filter is applied, but a softer, boost-like filter can be used instead.

This feature is generally available (GA). For more information, see Filter with natural language understanding.

Vertex AI Search: Allowlist fields for natural language query understanding

You can specify an allowlist of fields to be used for filter extraction. If you have some fields, perhaps for internal use, that you don't want to be used in filters, then specify an allowlist to restrict the fields that can be used.

For more information, see Specify fields for natural-language queries.

Vertex AI Search: Updated file size restrictions for unstructured data

The file size restrictions for unstructured data import have been unified. For unstructured data, you can import files up to 200 MB regardless of the parser type.

For more information, see Unstructured data.

Vertex AI Search: Gemini layout parser (Preview)

For data stores with unstructured documents, you can use Gemini to get layout analysis and content extraction on PDF files. Layout parsing with Gemini provides high quality table recognition, improved reading order and more accurate text recognition.

This feature is in Public Preview. For more information, see Parse and chunk documents.

Vertex AI Search: Layout parser support for DOCX, PPTX, and XLSX (GA)

With the layout parser, support for parsing DOCX, PPTX, and XLSX file formats is Generally Available (GA). Both the layout and digital parsers can parse PDF, HTML, DOCX, PPTX, and XLSX files. For more information about the parsers, see Parse and chunk documents.

Vertex AI Search: Renamed from AI Applications

The AI Applications product has been renamed as Vertex AI Search in the following contexts:

• The documentation set. See What is Vertex AI Search?
• The marketing collateral. See Vertex AI Search.

What has not changed:

• The user interface in the Google Cloud console is still referred to as AI Applications. See AI Applications.
• The APIs still use the Discovery Engine API endpoints. See APIs and reference.

Despite the rebrand, the product functionality remains the same.

Vertex AI Search: Custom ranking of search results (GA)

You can modify the ranking behavior of your search app using custom ranking. Custom ranking lets you provide a mathematical expression that relies on a set of model-computed signals (such as semantic relevance score and keyword similarity score) and document-based signals (such as a custom field like distance or document age). The resulting ranking of the search results is more considered and likely matches a user's needs better than a purely embedding-based ranking.

For more information, see Customize search results ranking.

Vertex AI Search: Advanced autocomplete (GA)

Use advanced autocomplete to enable autocomplete on blended search apps.

For more information, see Configure advanced autocomplete. This feature is Generally Available (GA).

Vertex AI Search: gemini-2.5-flash/answer_gen/v1 model is the default model

Model version gemini-2.5-flash/answer_gen/v1 is the default model for generating answers in Vertex AI Search.

For more information, see Answer generation model versions and lifecycle.

AI Applications: Quotas

AI Applications offers the following allocation quotas in the global region (global), the US multi-region (us), and the EU multi-region (eu) under the Discovery Engine API:

• Number of data stores per project
• Number of documents per project
• Number of engines per project
• Number of user events

The number of data stores, documents, user events, and engines across all locations can't exceed the total per-project quota for that resource.

For more information, see Quotas.

Vertex AI Search: Search for an exact match (GA)

To search for an exact match, you can enclose your search query in double quotes ("). For example, when you search for "Mary had a little lamb", Vertex AI Search looks for the phrase exactly as it is. It doesn't return search results that contain Mary had lamb, which has missing words; or a little lamb had Mary, which has the words in a different order.

This feature is Generally Available when you use the engines.servingConfigs.search method to get search results for custom data, media data, and healthcare data.

Vertex AI Search: gemini-2.5-flash/answer_gen/v1 model

You can generate answers with the Gemini 2.5 Flash (gemini-2.5-flash) model. This model is tuned to address context-based question and answering tasks.

For more information, see Answer generation model versions and lifecycle.

AI Applications: Custom search and recommendations

The vertical-agnostic apps, formerly known as generic search and recommendations, are renamed to custom search and recommendations. You'll see this new name in the product console and the documentation set. The functionality and the endpoints remain the same.

Vertex AI Search: Skip layout parsing for types of HTML content (GA)

The layout parser can skip parsing specific types of HTML content. By excluding less relevant content such as boilerplate, you can improve data quality. The layout parser can exclude based on HTML tags and IDs and on CSS classes.

This feature is Generally Available (GA) and accessible only through the API. For more information, see Exclude HTML content.

Vertex AI Search: Adjust autocomplete settings to reduce risk of PII leaks (Public Preview)

If you use either the search history or user events model for autocomplete suggestions and you have concerns about your users entering their personally identifiable information (PII) as search queries, then see Reduce the risk of returning suggestions that contain PII.

This feature is in Public Preview.

Vertex AI Search: Media search automatically chooses the results mode

Media search can respond to queries as full queries or as partial (search-as-you-type) queries. You can specify which query mode to use or you can specify the auto mode.

With auto mode, Media search makes the choice for you by assessing and weighting various signals such as query length and content. Use the auto mode if you have one search box that some users expect to respond with search-as-you-type suggestions and others expect to respond as a full query search.

Auto mode is the default mode. For more information, see Get search results for media.

Vertex AI Search: Obtain claim-level grounding scores (GA)

Claim-level scores from the check grounding API is Generally Available (GA). In addition to the answer-level support score, you can obtain a support score for each claim in an answer candidate.

For more information, see Obtain claim-level scores for an answer candidate.

Vertex AI Search: Stream Google Cloud Storage buckets to data stores

In addition to one time and periodic imports from Cloud Storage, you can stream unstructured data from Cloud Storage into a data store. This lets you serve results from the bucket to your users in near real time.

Streaming must be set up at the bucket-level (not at the folder- or file-level), and the bucket may only contain unstructured data.

For general information about creating data stores, see Create a search data store.

Vertex AI Search: Grounded generation with the generateGroundedContent API

The generateGroundedContent API to that grounds your answers with your inline text, Vertex AI Search data store, and Google Search is no longer available.

Instead, to generate grounded answers, Google recommends that you use the Generally Available groundContent API. You can either ground your answers with Google Search or with your own data. For more information, see Overview.

AI Applications: Renamed from Vertex AI Agent Builder

The Vertex AI Agent Builder product has been renamed AI Applications. You'll see this new name in the product console, the documentation set, and the marketing collateral. The product functionality and endpoints remain the same.

Vertex AI Search: Generate and return charts in answers and with follow-ups (Public Preview)

The answer method can include a chart in an answer, as well as text. The chart is generated from the data in the data store. A chart is generated if there is sufficient data, and the query either asks for a chart or the answer is sufficiently complex that the method itself determines that a chart is helpful.

This feature is in Public Preview and is only available through the API. For more information, see Generate charts for answers.

Vertex AI Search: Return corpus images in answers and with follow-ups (Public Preview)

The answer method can return images in answers, along with text.

If appropriate, one image from the data store can be returned with the answer. Citations can also include images from the data store.

This feature is restricted to queries made to unstructured data stores where the layout parser is in effect and is only available through the API. For more information, see Retrieve existing images from the data store.

Vertex AI Search: Ranking visibility and custom ranking (Private Preview)

You can access the signals that contribute to your ranking and then tune these signals to customize the ranking. To tune the ranking signals, you can modify predefined ranking expressions or specify custom ranking expressions.

Ranking visibility and custom ranking is a Private Preview feature.

Vertex AI Search: Document-relevance scores for search results (GA)

You can ask to have a relevance score returned for each search result associated with a query. The returned score can be used to do post-search ranking or filtering of the results. This feature is available for search apps associated with structured and unstructured data stores.

This feature is Generally Available (GA). For more information, see Get document-relevance score with search results.

Vertex AI Search: Personalize responses from the answer method (GA)

When making a query call to the answer method, you can provide information about the user to personalize the generated answer.

This feature is Generally Available (GA). For more information, see Personalize answers.

Vertex AI Search: Stream answers (GA)

Vertex AI Search now supports answer streaming. This feature returns generated answers in sequential parts, reducing the perception of latency. As the end users read the first part of the answer, the subsequent parts of the answer are being generated.

This feature also includes many of the features of the original answer method.

This feature is Generally Available to all customers. For more information, see Stream answers.

Vertex AI Search: gemini-2.0-flash-001 model for the answer method

The Gemini 2.0 Flash (gemini-2.0-flash-001) model is available for answer generation in Vertex AI Search. This model is tuned to address context-based question and answering tasks.

For more information, see Answer generation model versions and lifecycle.

Vertex AI Search: Identity mapping store (GA)

You can map your identity provider (IDP) to external identities from third-party applications that aren't managed by your identity provider. This information allows Google to enforce access control correctly by syncing users from your IDP and third-party applications. For more information, see the IdentityMappingStore resource in the API docs.

Vertex AI Search: gemini-1.5-flash-001/answer_gen/v2 for healthcare

The gemini-1.5-flash-001/answer_gen/v2 model is available for answer generation in healthcare search apps.

For more information, see Available models.

Vertex AI Search: View widget metrics on the Analytics page (GA with allowlist)

You can view metrics from the widget on the Analytics page.

This feature is available to select Google Cloud customers (GA with allowlist). For more information, see View search analytics.

Vertex AI Agent Builder: Grounding is available in more languages (GA with allowlist)

The grounded generation API supports more than 35 languages.

This feature is available to select Google Cloud customers (GA with allowlist). For general information about grounding, see Generate grounded answers with RAG. For available languages, see Languages.

Vertex AI Agent Builder: Additional inputs for generating grounded answers (GA with allowlist)

You can specify a language code and a latitude-longitude value when making calls to the grounded generation API.

If the language can't be determined from the query, then the language code is used to set the language for the answer. If the language code is not present, then the latitude-longitude value is used to set the language.

The latitude-longitude value is also used to answer location-related queries, such as "restaurants near me".

This feature is available to select Google Cloud customers (GA with allowlist). For more information, see Generate grounded answers with RAG.

Vertex AI Search: gemini-1.5-flash-002-high-fidelity model (Public Preview)

The gemini-1.5-flash-002-high-fidelity model is available for grounded answer generation with RAG. This model is based on the gemini-1.5-flash-002 model and has been further tuned to address context-based question and answering tasks. This model is suitable for specialized industries, such as financial services, healthcare, and insurance.

This model is available in Public Preview.

For more information, see High fidelity models.

Vertex AI Search: Boost controls for media recommendations (Public Preview)

Boost controls are used to affect the order in which recommendations are listed. Boost controls use filters on string and boolean values in the schema data to determine what media content to boost or bury. The boost value (-1 to 1) determines whether the content should be placed lower (buried) or higher (boosted) in the list of recommendations returned.

Boost controls are attached to serving configs and applied to recommend method calls.

The boost feature is in Public Preview and is available through the API. For more information about the feature, see Boost and bury media recommendations.

Vertex AI Search: Index and refresh web pages using sitemap (Public Preview)

If advanced website indexing is enabled in your data store, you can submit and use sitemaps and sitemap indexes to index and refresh the web pages in your data store. This feature supports only XML sitemaps and sitemap indexes.

This feature is in Public Preview and is available through the API. For more information about the feature, see Index and refresh web pages using sitemaps.

Vertex AI Search: Check ingested data quality for media recommendations (GA)

You can check the quality of your ingested data for media recommendations through the Google Cloud console. These checks are not blocking but can suggest ways that your data can be improved. This feature is Generally Available (GA).

Previously, this check was only available through API method calls.

For more information, see Check data quality for media recommendations.

Vertex AI Search: Stream answers (GA with allowlist)

The answer streaming method can return generated answers in sequential parts. This reduces the perception of latency. As the end users read the first part of the answer, the subsequent parts of the answer are being generated.

The answer streaming method also includes many of the features of the original answer method.

This feature is Generally Available to select Google customers (GA with allowlist). For more information, see Stream answers.

Vertex AI Search: Return only well-grounded answers with summaries and follow-ups (GA)

With the answer method, you can choose to filter out poorly-grounded answers. There are two filter levels: choose to return only answers with high grounding scores (at the risk of losing some helpful answers) or choose a lower filter to get more answers.

This feature is Generally Available (GA). For more information, see Show only well-grounded answers.

Vertex AI Search: Get grounding scores for answers with summaries and follow-ups (GA)

The answer method can return aggregated grounding scores for answers and individual grounding scores for claims.

This feature is Generally Available (GA). For more information, see