@ai-sdk/google-vertex 5.0.0-beta.5 → 5.0.0-beta.52

package/src/anthropic/google-vertex-anthropic-provider.ts

@@ -1,7 +1,7 @@
 import {
-  LanguageModelV3,
+  LanguageModelV4,
   NoSuchModelError,
-  ProviderV3,
+  ProviderV4,
 } from '@ai-sdk/provider';
 import {
   FetchFunction,
@@ -78,24 +78,47 @@ export const vertexAnthropicTools = {
    * Creates a web search tool that gives Claude direct access to real-time web content.
    */
   webSearch_20250305: anthropicTools.webSearch_20250305,
+
+  /**
+   * Creates a tool search tool that uses regex patterns to find tools.
+   *
+   * The tool search tool enables Claude to work with hundreds or thousands of tools
+   * by dynamically discovering and loading them on-demand.
+   *
+   * Use `providerOptions: { anthropic: { deferLoading: true } }` on other tools
+   * to mark them for deferred loading.
+   */
+  toolSearchRegex_20251119: anthropicTools.toolSearchRegex_20251119,
+
+  /**
+   * Creates a tool search tool that uses BM25 (natural language) to find tools.
+   *
+   * The tool search tool enables Claude to work with hundreds or thousands of tools
+   * by dynamically discovering and loading them on-demand.
+   *
+   * Use `providerOptions: { anthropic: { deferLoading: true } }` on other tools
+   * to mark them for deferred loading.
+   */
+  toolSearchBm25_20251119: anthropicTools.toolSearchBm25_20251119,
 };
-export interface GoogleVertexAnthropicProvider extends ProviderV3 {
+export interface GoogleVertexAnthropicProvider extends ProviderV4 {
   /**
    * Creates a model for text generation.
    */
-  (modelId: GoogleVertexAnthropicMessagesModelId): LanguageModelV3;
+  (modelId: GoogleVertexAnthropicMessagesModelId): LanguageModelV4;
 
   /**
    * Creates a model for text generation.
    */
-  languageModel(modelId: GoogleVertexAnthropicMessagesModelId): LanguageModelV3;
+  languageModel(modelId: GoogleVertexAnthropicMessagesModelId): LanguageModelV4;
 
   /**
    * Anthropic tools supported by Google Vertex.
    * Note: Only a subset of Anthropic tools are available on Vertex.
    * Supported tools: bash_20241022, bash_20250124, textEditor_20241022,
    * textEditor_20250124, textEditor_20250429, textEditor_20250728,
-   * computer_20241022, webSearch_20250305
+   * computer_20241022, webSearch_20250305, toolSearchRegex_20251119,
+   * toolSearchBm25_20251119
    */
   tools: typeof vertexAnthropicTools;
 
@@ -169,7 +192,7 @@ export function createVertexAnthropic(
         }`,
       transformRequestBody: args => {
         // Remove model from args and add anthropic version
-        const { model, ...rest } = args;
+        const { model: _model, ...rest } = args;
         return {
           ...rest,
           anthropic_version: 'vertex-2023-10-16',
@@ -179,6 +202,8 @@ export function createVertexAnthropic(
       supportedUrls: () => ({}),
       // force the use of JSON tool fallback for structured outputs since beta header isn't supported
       supportsNativeStructuredOutput: false,
+      // Vertex Anthropic doesn't support strict mode on tool definitions.
+      supportsStrictTools: false,
     });
 
   const provider = function (modelId: GoogleVertexAnthropicMessagesModelId) {
@@ -191,7 +216,7 @@ export function createVertexAnthropic(
     return createChatModel(modelId);
   };
 
-  provider.specificationVersion = 'v3' as const;
+  provider.specificationVersion = 'v4' as const;
   provider.languageModel = createChatModel;
   provider.chat = createChatModel;
   provider.messages = createChatModel;

package/src/edge/google-vertex-provider-edge.ts

@@ -11,8 +11,7 @@ import {
 
 export type { GoogleVertexProvider };
 
-export interface GoogleVertexProviderSettings
-  extends GoogleVertexProviderSettingsOriginal {
+export interface GoogleVertexProviderSettings extends GoogleVertexProviderSettingsOriginal {
   /**
    * Optional. The Google credentials for the Google Cloud service account. If
    * not provided, the Google Vertex provider will use environment variables to

package/src/google-vertex-config.ts

@@ -3,6 +3,6 @@ import { FetchFunction, Resolvable } from '@ai-sdk/provider-utils';
 export interface GoogleVertexConfig {
   provider: string;
   baseURL: string;
-  headers: Resolvable<Record<string, string | undefined>>;
+  headers?: Resolvable<Record<string, string | undefined>>;
   fetch?: FetchFunction;
 }

package/src/google-vertex-embedding-model.ts

@@ -1,5 +1,5 @@
 import {
-  EmbeddingModelV3,
+  EmbeddingModelV4,
   TooManyEmbeddingValuesForCallError,
 } from '@ai-sdk/provider';
 import {
@@ -8,6 +8,9 @@ import {
   postJsonToApi,
   resolve,
   parseProviderOptions,
+  serializeModelOptions,
+  WORKFLOW_SERIALIZE,
+  WORKFLOW_DESERIALIZE,
 } from '@ai-sdk/provider-utils';
 import { z } from 'zod/v4';
 import { googleVertexFailedResponseHandler } from './google-vertex-error';
@@ -17,14 +20,28 @@ import {
 } from './google-vertex-embedding-options';
 import { GoogleVertexConfig } from './google-vertex-config';
 
-export class GoogleVertexEmbeddingModel implements EmbeddingModelV3 {
-  readonly specificationVersion = 'v3';
+export class GoogleVertexEmbeddingModel implements EmbeddingModelV4 {
+  readonly specificationVersion = 'v4';
   readonly modelId: GoogleVertexEmbeddingModelId;
   readonly maxEmbeddingsPerCall = 2048;
   readonly supportsParallelCalls = true;
 
   private readonly config: GoogleVertexConfig;
 
+  static [WORKFLOW_SERIALIZE](model: GoogleVertexEmbeddingModel) {
+    return serializeModelOptions({
+      modelId: model.modelId,
+      config: model.config,
+    });
+  }
+
+  static [WORKFLOW_DESERIALIZE](options: {
+    modelId: string;
+    config: GoogleVertexConfig;
+  }) {
+    return new GoogleVertexEmbeddingModel(options.modelId, options.config);
+  }
+
   get provider(): string {
     return this.config.provider;
   }
@@ -42,8 +59,8 @@ export class GoogleVertexEmbeddingModel implements EmbeddingModelV3 {
     headers,
     abortSignal,
     providerOptions,
-  }: Parameters<EmbeddingModelV3['doEmbed']>[0]): Promise<
-    Awaited<ReturnType<EmbeddingModelV3['doEmbed']>>
+  }: Parameters<EmbeddingModelV4['doEmbed']>[0]): Promise<
+    Awaited<ReturnType<EmbeddingModelV4['doEmbed']>>
   > {
     let googleOptions = await parseProviderOptions({
       provider: 'vertex',
@@ -71,7 +88,7 @@ export class GoogleVertexEmbeddingModel implements EmbeddingModelV3 {
     }
 
     const mergedHeaders = combineHeaders(
-      await resolve(this.config.headers),
+      this.config.headers ? await resolve(this.config.headers) : undefined,
       headers,
     );
 

package/src/google-vertex-embedding-options.ts

@@ -10,6 +10,8 @@ export type GoogleVertexEmbeddingModelId =
   | 'text-multilingual-embedding-002'
   | 'text-embedding-004'
   | 'text-embedding-005'
+  | 'gemini-embedding-001'
+  | 'gemini-embedding-2-preview'
   | (string & {});
 
 export const googleVertexEmbeddingModelOptions = z.object({

package/src/google-vertex-image-model.ts

@@ -1,10 +1,10 @@
 import type { GoogleLanguageModelOptions } from '@ai-sdk/google';
 import { GoogleGenerativeAILanguageModel } from '@ai-sdk/google/internal';
 import {
-  ImageModelV3,
-  ImageModelV3File,
-  LanguageModelV3Prompt,
-  SharedV3Warning,
+  ImageModelV4,
+  ImageModelV4File,
+  LanguageModelV4Prompt,
+  SharedV4Warning,
 } from '@ai-sdk/provider';
 import {
   Resolvable,
@@ -16,6 +16,9 @@ import {
   parseProviderOptions,
   postJsonToApi,
   resolve,
+  serializeModelOptions,
+  WORKFLOW_SERIALIZE,
+  WORKFLOW_DESERIALIZE,
 } from '@ai-sdk/provider-utils';
 import { z } from 'zod/v4';
 import { googleVertexFailedResponseHandler } from './google-vertex-error';
@@ -33,8 +36,22 @@ interface GoogleVertexImageModelConfig {
 }
 
 // https://cloud.google.com/vertex-ai/generative-ai/docs/image/generate-images
-export class GoogleVertexImageModel implements ImageModelV3 {
-  readonly specificationVersion = 'v3';
+export class GoogleVertexImageModel implements ImageModelV4 {
+  readonly specificationVersion = 'v4';
+
+  static [WORKFLOW_SERIALIZE](model: GoogleVertexImageModel) {
+    return serializeModelOptions({
+      modelId: model.modelId,
+      config: model.config,
+    });
+  }
+
+  static [WORKFLOW_DESERIALIZE](options: {
+    modelId: string;
+    config: GoogleVertexImageModelConfig;
+  }) {
+    return new GoogleVertexImageModel(options.modelId, options.config);
+  }
 
   get maxImagesPerCall(): number {
     if (isGeminiModel(this.modelId)) {
@@ -53,8 +70,8 @@ export class GoogleVertexImageModel implements ImageModelV3 {
   ) {}
 
   async doGenerate(
-    options: Parameters<ImageModelV3['doGenerate']>[0],
-  ): Promise<Awaited<ReturnType<ImageModelV3['doGenerate']>>> {
+    options: Parameters<ImageModelV4['doGenerate']>[0],
+  ): Promise<Awaited<ReturnType<ImageModelV4['doGenerate']>>> {
     if (isGeminiModel(this.modelId)) {
       return this.doGenerateGemini(options);
     }
@@ -72,10 +89,10 @@ export class GoogleVertexImageModel implements ImageModelV3 {
     abortSignal,
     files,
     mask,
-  }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<
-    Awaited<ReturnType<ImageModelV3['doGenerate']>>
+  }: Parameters<ImageModelV4['doGenerate']>[0]): Promise<
+    Awaited<ReturnType<ImageModelV4['doGenerate']>>
   > {
-    const warnings: Array<SharedV3Warning> = [];
+    const warnings: Array<SharedV4Warning> = [];
 
     if (size != null) {
       warnings.push({
@@ -164,7 +181,10 @@ export class GoogleVertexImageModel implements ImageModelV3 {
     const currentDate = this.config._internal?.currentDate?.() ?? new Date();
     const { value: response, responseHeaders } = await postJsonToApi({
       url: `${this.config.baseURL}/models/${this.modelId}:predict`,
-      headers: combineHeaders(await resolve(this.config.headers), headers),
+      headers: combineHeaders(
+        this.config.headers ? await resolve(this.config.headers) : undefined,
+        headers,
+      ),
       body,
       failedResponseHandler: googleVertexFailedResponseHandler,
       successfulResponseHandler: createJsonResponseHandler(
@@ -212,10 +232,10 @@ export class GoogleVertexImageModel implements ImageModelV3 {
     abortSignal,
     files,
     mask,
-  }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<
-    Awaited<ReturnType<ImageModelV3['doGenerate']>>
+  }: Parameters<ImageModelV4['doGenerate']>[0]): Promise<
+    Awaited<ReturnType<ImageModelV4['doGenerate']>>
   > {
-    const warnings: Array<SharedV3Warning> = [];
+    const warnings: Array<SharedV4Warning> = [];
 
     if (mask != null) {
       throw new Error(
@@ -268,7 +288,7 @@ export class GoogleVertexImageModel implements ImageModelV3 {
       }
     }
 
-    const languageModelPrompt: LanguageModelV3Prompt = [
+    const languageModelPrompt: LanguageModelV4Prompt = [
       { role: 'user', content: userContent },
     ];
 
@@ -433,9 +453,9 @@ export type GoogleVertexImageModelOptions = z.infer<
 >;
 
 /**
- * Helper to convert ImageModelV3File data to base64 string
+ * Helper to convert ImageModelV4File data to base64 string
  */
-function getBase64Data(file: ImageModelV3File): string {
+function getBase64Data(file: ImageModelV4File): string {
   if (file.type === 'url') {
     throw new Error(
       'URL-based images are not supported for Google Vertex image editing. Please provide the image data directly.',

package/src/google-vertex-options.ts

@@ -21,7 +21,6 @@ export type GoogleVertexModelId =
   | 'gemini-1.0-pro-002'
   // Preview models
   | 'gemini-2.0-flash-lite-preview-02-05'
-  | 'gemini-2.5-flash-lite-preview-09-2025'
   | 'gemini-2.5-flash-preview-09-2025'
   | 'gemini-3-pro-preview'
   | 'gemini-3-pro-image-preview'

package/src/google-vertex-provider-node.ts

@@ -7,8 +7,7 @@ import {
   GoogleVertexProviderSettings as GoogleVertexProviderSettingsOriginal,
 } from './google-vertex-provider';
 
-export interface GoogleVertexProviderSettings
-  extends GoogleVertexProviderSettingsOriginal {
+export interface GoogleVertexProviderSettings extends GoogleVertexProviderSettingsOriginal {
   /**
    * Optional. The Authentication options provided by google-auth-library.
    * Complete list of authentication options is documented in the

package/src/google-vertex-provider.ts

@@ -1,9 +1,9 @@
 import { GoogleGenerativeAILanguageModel } from '@ai-sdk/google/internal';
 import {
-  Experimental_VideoModelV3,
-  ImageModelV3,
-  LanguageModelV3,
-  ProviderV3,
+  Experimental_VideoModelV4,
+  ImageModelV4,
+  LanguageModelV4,
+  ProviderV4,
 } from '@ai-sdk/provider';
 import {
   FetchFunction,
@@ -47,23 +47,23 @@ function createExpressModeFetch(
   };
 }
 
-export interface GoogleVertexProvider extends ProviderV3 {
+export interface GoogleVertexProvider extends ProviderV4 {
   /**
    * Creates a model for text generation.
    */
-  (modelId: GoogleVertexModelId): LanguageModelV3;
+  (modelId: GoogleVertexModelId): LanguageModelV4;
 
-  languageModel: (modelId: GoogleVertexModelId) => LanguageModelV3;
+  languageModel: (modelId: GoogleVertexModelId) => LanguageModelV4;
 
   /**
    * Creates a model for image generation.
    */
-  image(modelId: GoogleVertexImageModelId): ImageModelV3;
+  image(modelId: GoogleVertexImageModelId): ImageModelV4;
 
   /**
    * Creates a model for image generation.
    */
-  imageModel(modelId: GoogleVertexImageModelId): ImageModelV3;
+  imageModel(modelId: GoogleVertexImageModelId): ImageModelV4;
 
   tools: typeof googleVertexTools;
 
@@ -77,12 +77,12 @@ export interface GoogleVertexProvider extends ProviderV3 {
   /**
    * Creates a model for video generation.
    */
-  video(modelId: GoogleVertexVideoModelId): Experimental_VideoModelV3;
+  video(modelId: GoogleVertexVideoModelId): Experimental_VideoModelV4;
 
   /**
    * Creates a model for video generation.
    */
-  videoModel(modelId: GoogleVertexVideoModelId): Experimental_VideoModelV3;
+  videoModel(modelId: GoogleVertexVideoModelId): Experimental_VideoModelV4;
 }
 
 export interface GoogleVertexProviderSettings {
@@ -231,7 +231,7 @@ export function createVertex(
     return createChatModel(modelId);
   };
 
-  provider.specificationVersion = 'v3' as const;
+  provider.specificationVersion = 'v4' as const;
   provider.languageModel = createChatModel;
   provider.embeddingModel = createEmbeddingModel;
   provider.textEmbeddingModel = createEmbeddingModel;

package/src/google-vertex-video-model.ts

@@ -1,7 +1,7 @@
 import {
   AISDKError,
-  type Experimental_VideoModelV3,
-  type SharedV3Warning,
+  type Experimental_VideoModelV4,
+  type SharedV4Warning,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
@@ -53,8 +53,8 @@ interface GoogleVertexVideoModelConfig {
   };
 }
 
-export class GoogleVertexVideoModel implements Experimental_VideoModelV3 {
-  readonly specificationVersion = 'v3';
+export class GoogleVertexVideoModel implements Experimental_VideoModelV4 {
+  readonly specificationVersion = 'v4';
 
   get provider(): string {
     return this.config.provider;
@@ -71,10 +71,10 @@ export class GoogleVertexVideoModel implements Experimental_VideoModelV3 {
   ) {}
 
   async doGenerate(
-    options: Parameters<Experimental_VideoModelV3['doGenerate']>[0],
-  ): Promise<Awaited<ReturnType<Experimental_VideoModelV3['doGenerate']>>> {
+    options: Parameters<Experimental_VideoModelV4['doGenerate']>[0],
+  ): Promise<Awaited<ReturnType<Experimental_VideoModelV4['doGenerate']>>> {
     const currentDate = this.config._internal?.currentDate?.() ?? new Date();
-    const warnings: SharedV3Warning[] = [];
+    const warnings: SharedV4Warning[] = [];
 
     const vertexOptions = (await parseProviderOptions({
       provider: 'vertex',

package/src/maas/edge/google-vertex-maas-provider-edge.ts

@@ -0,0 +1,65 @@
+import { FetchFunction, resolve } from '@ai-sdk/provider-utils';
+import {
+  generateAuthToken,
+  GoogleCredentials,
+} from '../../edge/google-vertex-auth-edge';
+import {
+  createVertexMaas as createVertexMaasOriginal,
+  GoogleVertexMaasProvider,
+  GoogleVertexMaasProviderSettings as GoogleVertexMaasProviderSettingsOriginal,
+} from '../google-vertex-maas-provider';
+
+export type { GoogleVertexMaasProvider };
+
+export interface GoogleVertexMaasProviderSettings extends GoogleVertexMaasProviderSettingsOriginal {
+  /**
+   * Optional. The Google credentials for the Google Cloud service account. If
+   * not provided, the Google Vertex provider will use environment variables to
+   * load the credentials.
+   */
+  googleCredentials?: GoogleCredentials;
+}
+
+/**
+ * Create a Google Vertex AI MaaS (Model as a Service) provider instance for Edge runtimes.
+ * Uses the OpenAI-compatible Chat Completions API for partner and open models.
+ * Automatically handles Google Cloud authentication.
+ *
+ * @see https://cloud.google.com/vertex-ai/generative-ai/docs/maas/use-open-models
+ */
+export function createVertexMaas(
+  options: GoogleVertexMaasProviderSettings = {},
+): GoogleVertexMaasProvider {
+  // Create a custom fetch wrapper that adds auth headers
+  const customFetch: FetchFunction = async (url, init) => {
+    const token = await generateAuthToken(options.googleCredentials);
+    const resolvedHeaders = await resolve(options.headers);
+    const authHeaders = {
+      ...resolvedHeaders,
+      Authorization: `Bearer ${token}`,
+    };
+
+    // Merge auth headers with existing headers from init
+    const fetchInit = {
+      ...init,
+      headers: {
+        ...init?.headers,
+        ...authHeaders,
+      },
+    };
+
+    // Call the original fetch or user's custom fetch
+    return (options.fetch ?? fetch)(url, fetchInit);
+  };
+
+  return createVertexMaasOriginal({
+    ...options,
+    fetch: customFetch,
+    headers: undefined, // Don't pass headers, we handle them in fetch
+  });
+}
+
+/**
+ * Default Google Vertex AI MaaS provider instance for Edge runtimes.
+ */
+export const vertexMaas = createVertexMaas();

package/src/maas/edge/index.ts

@@ -0,0 +1,9 @@
+export {
+  createVertexMaas,
+  vertexMaas,
+} from './google-vertex-maas-provider-edge';
+export type {
+  GoogleVertexMaasProvider,
+  GoogleVertexMaasProviderSettings,
+} from './google-vertex-maas-provider-edge';
+export type { GoogleVertexMaasModelId } from '../google-vertex-maas-options';

package/src/maas/google-vertex-maas-options.ts

@@ -0,0 +1,15 @@
+// https://cloud.google.com/vertex-ai/generative-ai/docs/maas/use-open-models
+export type GoogleVertexMaasModelId =
+  | 'deepseek-ai/deepseek-r1-0528-maas'
+  | 'deepseek-ai/deepseek-v3.1-maas'
+  | 'deepseek-ai/deepseek-v3.2-maas'
+  | 'openai/gpt-oss-120b-maas'
+  | 'openai/gpt-oss-20b-maas'
+  | 'meta/llama-4-maverick-17b-128e-instruct-maas'
+  | 'meta/llama-4-scout-17b-16e-instruct-maas'
+  | 'minimax/minimax-m2-maas'
+  | 'qwen/qwen3-coder-480b-a35b-instruct-maas'
+  | 'qwen/qwen3-next-80b-a3b-instruct-maas'
+  | 'qwen/qwen3-next-80b-a3b-thinking-maas'
+  | 'moonshotai/kimi-k2-thinking-maas'
+  | (string & {});

package/src/maas/google-vertex-maas-provider-node.ts

@@ -0,0 +1,64 @@
+import { FetchFunction, resolve } from '@ai-sdk/provider-utils';
+import { GoogleAuthOptions } from 'google-auth-library';
+import { generateAuthToken } from '../google-vertex-auth-google-auth-library';
+import {
+  createVertexMaas as createVertexMaasOriginal,
+  GoogleVertexMaasProvider,
+  GoogleVertexMaasProviderSettings as GoogleVertexMaasProviderSettingsOriginal,
+} from './google-vertex-maas-provider';
+
+export type { GoogleVertexMaasProvider };
+
+export interface GoogleVertexMaasProviderSettings extends GoogleVertexMaasProviderSettingsOriginal {
+  /**
+   * Optional. The Authentication options provided by google-auth-library.
+   * Complete list of authentication options is documented in the
+   * GoogleAuthOptions interface:
+   * https://github.com/googleapis/google-auth-library-nodejs/blob/main/src/auth/googleauth.ts.
+   */
+  googleAuthOptions?: GoogleAuthOptions;
+}
+
+/**
+ * Create a Google Vertex AI MaaS (Model as a Service) provider instance for Node.js.
+ * Uses the OpenAI-compatible Chat Completions API for partner and open models.
+ * Automatically handles Google Cloud authentication.
+ *
+ * @see https://cloud.google.com/vertex-ai/generative-ai/docs/maas/use-open-models
+ */
+export function createVertexMaas(
+  options: GoogleVertexMaasProviderSettings = {},
+): GoogleVertexMaasProvider {
+  // Create a custom fetch wrapper that adds auth headers
+  const customFetch: FetchFunction = async (url, init) => {
+    const token = await generateAuthToken(options.googleAuthOptions);
+    const resolvedHeaders = await resolve(options.headers);
+    const authHeaders = {
+      ...resolvedHeaders,
+      Authorization: `Bearer ${token}`,
+    };
+
+    // Merge auth headers with existing headers from init
+    const fetchInit = {
+      ...init,
+      headers: {
+        ...init?.headers,
+        ...authHeaders,
+      },
+    };
+
+    // Call the original fetch or user's custom fetch
+    return (options.fetch ?? fetch)(url, fetchInit);
+  };
+
+  return createVertexMaasOriginal({
+    ...options,
+    fetch: customFetch,
+    headers: undefined, // Don't pass headers, we handle them in fetch
+  });
+}
+
+/**
+ * Default Google Vertex AI MaaS provider instance for Node.js.
+ */
+export const vertexMaas = createVertexMaas();