@ai-sdk/google-vertex 5.0.0-beta.11 → 5.0.0-beta.112

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

@@ -1,6 +1,6 @@
 import {
-  EmbeddingModelV4,
   TooManyEmbeddingValuesForCallError,
+  type EmbeddingModelV4,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
@@ -8,27 +8,49 @@ 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';
 import {
-  GoogleVertexEmbeddingModelId,
   googleVertexEmbeddingModelOptions,
-} from './google-vertex-embedding-options';
-import { GoogleVertexConfig } from './google-vertex-config';
+  type GoogleVertexEmbeddingModelId,
+} from './google-vertex-embedding-model-options';
+import type { GoogleVertexConfig } from './google-vertex-config';
 
 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;
   }
 
+  // gemini-embedding-2 models only support :embedContent (one value per call),
+  // not the :predict batch endpoint. https://github.com/vercel/ai/issues/15853
+  get maxEmbeddingsPerCall(): number {
+    return usesEmbedContentEndpoint(this.modelId) ? 1 : 2048;
+  }
+
   constructor(
     modelId: GoogleVertexEmbeddingModelId,
     config: GoogleVertexConfig,
@@ -46,11 +68,19 @@ export class GoogleVertexEmbeddingModel implements EmbeddingModelV4 {
     Awaited<ReturnType<EmbeddingModelV4['doEmbed']>>
   > {
     let googleOptions = await parseProviderOptions({
-      provider: 'vertex',
+      provider: 'googleVertex',
       providerOptions,
       schema: googleVertexEmbeddingModelOptions,
     });
 
+    if (googleOptions == null) {
+      googleOptions = await parseProviderOptions({
+        provider: 'vertex',
+        providerOptions,
+        schema: googleVertexEmbeddingModelOptions,
+      });
+    }
+
     if (googleOptions == null) {
       googleOptions = await parseProviderOptions({
         provider: 'google',
@@ -71,10 +101,46 @@ export class GoogleVertexEmbeddingModel implements EmbeddingModelV4 {
     }
 
     const mergedHeaders = combineHeaders(
-      await resolve(this.config.headers),
+      this.config.headers ? await resolve(this.config.headers) : undefined,
       headers,
     );
 
+    if (usesEmbedContentEndpoint(this.modelId)) {
+      const {
+        responseHeaders,
+        value: response,
+        rawValue,
+      } = await postJsonToApi({
+        url: `${this.config.baseURL}/models/${this.modelId}:embedContent`,
+        headers: mergedHeaders,
+        body: {
+          content: { parts: [{ text: values[0] }] },
+          embedContentConfig: {
+            outputDimensionality: googleOptions.outputDimensionality,
+            taskType: googleOptions.taskType,
+            title: googleOptions.title,
+            autoTruncate: googleOptions.autoTruncate,
+          },
+        },
+        failedResponseHandler: googleVertexFailedResponseHandler,
+        successfulResponseHandler: createJsonResponseHandler(
+          googleVertexEmbedContentResponseSchema,
+        ),
+        abortSignal,
+        fetch: this.config.fetch,
+      });
+
+      return {
+        warnings: [],
+        embeddings: [response.embedding.values],
+        usage:
+          response.usageMetadata?.promptTokenCount == null
+            ? undefined
+            : { tokens: response.usageMetadata.promptTokenCount },
+        response: { headers: responseHeaders, body: rawValue },
+      };
+    }
+
     const url = `${this.config.baseURL}/models/${this.modelId}:predict`;
     const {
       responseHeaders,
@@ -133,3 +199,20 @@ const googleVertexTextEmbeddingResponseSchema = z.object({
     }),
   ),
 });
+
+const googleVertexEmbedContentResponseSchema = z.object({
+  embedding: z.object({
+    values: z.array(z.number()),
+  }),
+  usageMetadata: z
+    .object({
+      promptTokenCount: z.number().nullish(),
+    })
+    .nullish(),
+});
+
+function usesEmbedContentEndpoint(modelId: GoogleVertexEmbeddingModelId) {
+  return (
+    modelId === 'gemini-embedding-2' || modelId === 'gemini-embedding-2-preview'
+  );
+}

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

@@ -0,0 +1,74 @@
+import { z } from 'zod/v4';
+
+export const googleVertexImageModelOptionsSchema = z.object({
+  negativePrompt: z.string().nullish(),
+  personGeneration: z
+    .enum(['dont_allow', 'allow_adult', 'allow_all'])
+    .nullish(),
+  safetySetting: z
+    .enum([
+      'block_low_and_above',
+      'block_medium_and_above',
+      'block_only_high',
+      'block_none',
+    ])
+    .nullish(),
+  addWatermark: z.boolean().nullish(),
+  storageUri: z.string().nullish(),
+  sampleImageSize: z.enum(['1K', '2K']).nullish(),
+  /**
+   * Configuration for image editing operations
+   */
+  edit: z
+    .object({
+      /**
+       * An integer that represents the number of sampling steps.
+       * A higher value offers better image quality, a lower value offers better latency.
+       * Try 35 steps to start. If the quality doesn't meet your requirements,
+       * increase the value towards an upper limit of 75.
+       */
+      baseSteps: z.number().nullish(),
+
+      // Edit mode options
+      // https://cloud.google.com/vertex-ai/generative-ai/docs/image/edit-insert-objects
+      mode: z
+        .enum([
+          'EDIT_MODE_INPAINT_INSERTION',
+          'EDIT_MODE_INPAINT_REMOVAL',
+          'EDIT_MODE_OUTPAINT',
+          'EDIT_MODE_CONTROLLED_EDITING',
+          'EDIT_MODE_PRODUCT_IMAGE',
+          'EDIT_MODE_BGSWAP',
+        ])
+        .nullish(),
+
+      /**
+       * The mask mode to use.
+       * - `MASK_MODE_DEFAULT` - Default value for mask mode.
+       * - `MASK_MODE_USER_PROVIDED` - User provided mask. No segmentation needed.
+       * - `MASK_MODE_DETECTION_BOX` - Mask from detected bounding boxes.
+       * - `MASK_MODE_CLOTHING_AREA` - Masks from segmenting the clothing area with open-vocab segmentation.
+       * - `MASK_MODE_PARSED_PERSON` - Masks from segmenting the person body and clothing using the person-parsing model.
+       */
+      maskMode: z
+        .enum([
+          'MASK_MODE_DEFAULT',
+          'MASK_MODE_USER_PROVIDED',
+          'MASK_MODE_DETECTION_BOX',
+          'MASK_MODE_CLOTHING_AREA',
+          'MASK_MODE_PARSED_PERSON',
+        ])
+        .nullish(),
+
+      /**
+       * Optional. A float value between 0 and 1, inclusive, that represents the
+       * percentage of the image width to grow the mask by. Using dilation helps
+       * compensate for imprecise masks. We recommend a value of 0.01.
+       */
+      maskDilation: z.number().nullish(),
+    })
+    .nullish(),
+});
+export type GoogleVertexImageModelOptions = z.infer<
+  typeof googleVertexImageModelOptionsSchema
+>;

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

@@ -1,13 +1,12 @@
 import type { GoogleLanguageModelOptions } from '@ai-sdk/google';
-import { GoogleGenerativeAILanguageModel } from '@ai-sdk/google/internal';
-import {
+import { GoogleLanguageModel } from '@ai-sdk/google/internal';
+import type {
   ImageModelV4,
   ImageModelV4File,
   LanguageModelV4Prompt,
   SharedV4Warning,
 } from '@ai-sdk/provider';
 import {
-  Resolvable,
   combineHeaders,
   convertToBase64,
   convertUint8ArrayToBase64,
@@ -16,10 +15,15 @@ import {
   parseProviderOptions,
   postJsonToApi,
   resolve,
+  serializeModelOptions,
+  WORKFLOW_SERIALIZE,
+  WORKFLOW_DESERIALIZE,
+  type Resolvable,
 } from '@ai-sdk/provider-utils';
 import { z } from 'zod/v4';
 import { googleVertexFailedResponseHandler } from './google-vertex-error';
-import { GoogleVertexImageModelId } from './google-vertex-image-settings';
+import { googleVertexImageModelOptionsSchema } from './google-vertex-image-model-options';
+import type { GoogleVertexImageModelId } from './google-vertex-image-settings';
 
 interface GoogleVertexImageModelConfig {
   provider: string;
@@ -36,6 +40,20 @@ interface GoogleVertexImageModelConfig {
 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)) {
       return 10;
@@ -86,14 +104,20 @@ export class GoogleVertexImageModel implements ImageModelV4 {
       });
     }
 
-    const vertexImageOptions = await parseProviderOptions({
-      provider: 'vertex',
-      providerOptions,
-      schema: googleVertexImageModelOptionsSchema,
-    });
+    const googleVertexImageOptions =
+      (await parseProviderOptions({
+        provider: 'googleVertex',
+        providerOptions,
+        schema: googleVertexImageModelOptionsSchema,
+      })) ??
+      (await parseProviderOptions({
+        provider: 'vertex',
+        providerOptions,
+        schema: googleVertexImageModelOptionsSchema,
+      }));
 
     // Extract edit-specific options from provider options
-    const { edit, ...otherOptions } = vertexImageOptions ?? {};
+    const { edit, ...otherOptions } = googleVertexImageOptions ?? {};
     const { mode: editMode, baseSteps, maskMode, maskDilation } = edit ?? {};
 
     // Build the request body based on whether we're editing or generating
@@ -164,11 +188,14 @@ export class GoogleVertexImageModel implements ImageModelV4 {
     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(
-        vertexImageResponseSchema,
+        googleVertexImageResponseSchema,
       ),
       abortSignal,
       fetch: this.config.fetch,
@@ -185,8 +212,8 @@ export class GoogleVertexImageModel implements ImageModelV4 {
         modelId: this.modelId,
         headers: responseHeaders,
       },
-      providerMetadata: {
-        vertex: {
+      providerMetadata: (() => {
+        const payload = {
           images:
             response.predictions?.map(prediction => {
               const {
@@ -196,8 +223,9 @@ export class GoogleVertexImageModel implements ImageModelV4 {
 
               return { ...(revisedPrompt != null && { revisedPrompt }) };
             }) ?? [],
-        },
-      },
+        };
+        return { googleVertex: payload, vertex: payload };
+      })(),
     };
   }
 
@@ -240,7 +268,13 @@ export class GoogleVertexImageModel implements ImageModelV4 {
 
     const userContent: Array<
       | { type: 'text'; text: string }
-      | { type: 'file'; data: string | Uint8Array | URL; mediaType: string }
+      | {
+          type: 'file';
+          data:
+            | { type: 'data'; data: string | Uint8Array }
+            | { type: 'url'; url: URL };
+          mediaType: string;
+        }
     > = [];
 
     if (prompt != null) {
@@ -252,16 +286,19 @@ export class GoogleVertexImageModel implements ImageModelV4 {
         if (file.type === 'url') {
           userContent.push({
             type: 'file',
-            data: new URL(file.url),
+            data: { type: 'url', url: new URL(file.url) },
             mediaType: 'image/*',
           });
         } else {
           userContent.push({
             type: 'file',
-            data:
-              typeof file.data === 'string'
-                ? file.data
-                : new Uint8Array(file.data),
+            data: {
+              type: 'data',
+              data:
+                typeof file.data === 'string'
+                  ? file.data
+                  : new Uint8Array(file.data),
+            },
             mediaType: file.mediaType,
           });
         }
@@ -272,7 +309,7 @@ export class GoogleVertexImageModel implements ImageModelV4 {
       { role: 'user', content: userContent },
     ];
 
-    const languageModel = new GoogleGenerativeAILanguageModel(this.modelId, {
+    const languageModel = new GoogleLanguageModel(this.modelId, {
       provider: this.config.provider,
       baseURL: this.config.baseURL,
       headers: this.config.headers ?? {},
@@ -283,24 +320,27 @@ export class GoogleVertexImageModel implements ImageModelV4 {
       }),
     });
 
+    const userVertexOptions = (providerOptions?.googleVertex ??
+      providerOptions?.vertex) as
+      | Omit<GoogleLanguageModelOptions, 'responseModalities' | 'imageConfig'>
+      | undefined;
+    const innerVertexOptions: GoogleLanguageModelOptions = {
+      responseModalities: ['IMAGE'],
+      imageConfig: aspectRatio
+        ? {
+            aspectRatio: aspectRatio as NonNullable<
+              GoogleLanguageModelOptions['imageConfig']
+            >['aspectRatio'],
+          }
+        : undefined,
+      ...(userVertexOptions ?? {}),
+    };
     const result = await languageModel.doGenerate({
       prompt: languageModelPrompt,
       seed,
       providerOptions: {
-        vertex: {
-          responseModalities: ['IMAGE'],
-          imageConfig: aspectRatio
-            ? {
-                aspectRatio: aspectRatio as NonNullable<
-                  GoogleLanguageModelOptions['imageConfig']
-                >['aspectRatio'],
-              }
-            : undefined,
-          ...((providerOptions?.vertex as Omit<
-            GoogleLanguageModelOptions,
-            'responseModalities' | 'imageConfig'
-          >) ?? {}),
-        } satisfies GoogleLanguageModelOptions,
+        googleVertex: innerVertexOptions,
+        vertex: innerVertexOptions,
       },
       headers,
       abortSignal,
@@ -310,18 +350,24 @@ export class GoogleVertexImageModel implements ImageModelV4 {
 
     const images: string[] = [];
     for (const part of result.content) {
-      if (part.type === 'file' && part.mediaType.startsWith('image/')) {
-        images.push(convertToBase64(part.data));
+      if (
+        part.type === 'file' &&
+        part.mediaType.startsWith('image/') &&
+        part.data.type === 'data'
+      ) {
+        images.push(convertToBase64(part.data.data));
       }
     }
 
+    const geminiPayload = {
+      images: images.map(() => ({})),
+    };
     return {
       images,
       warnings,
       providerMetadata: {
-        vertex: {
-          images: images.map(() => ({})),
-        },
+        googleVertex: geminiPayload,
+        vertex: geminiPayload,
       },
       response: {
         timestamp: currentDate,
@@ -347,7 +393,7 @@ function isGeminiModel(modelId: string): boolean {
 
 // minimal version of the schema, focussed on what is needed for the implementation
 // this approach limits breakages when the API changes and increases efficiency
-const vertexImageResponseSchema = z.object({
+const googleVertexImageResponseSchema = z.object({
   predictions: z
     .array(
       z.object({
@@ -359,79 +405,6 @@ const vertexImageResponseSchema = z.object({
     .nullish(),
 });
 
-const googleVertexImageModelOptionsSchema = z.object({
-  negativePrompt: z.string().nullish(),
-  personGeneration: z
-    .enum(['dont_allow', 'allow_adult', 'allow_all'])
-    .nullish(),
-  safetySetting: z
-    .enum([
-      'block_low_and_above',
-      'block_medium_and_above',
-      'block_only_high',
-      'block_none',
-    ])
-    .nullish(),
-  addWatermark: z.boolean().nullish(),
-  storageUri: z.string().nullish(),
-  sampleImageSize: z.enum(['1K', '2K']).nullish(),
-  /**
-   * Configuration for image editing operations
-   */
-  edit: z
-    .object({
-      /**
-       * An integer that represents the number of sampling steps.
-       * A higher value offers better image quality, a lower value offers better latency.
-       * Try 35 steps to start. If the quality doesn't meet your requirements,
-       * increase the value towards an upper limit of 75.
-       */
-      baseSteps: z.number().nullish(),
-
-      // Edit mode options
-      // https://cloud.google.com/vertex-ai/generative-ai/docs/image/edit-insert-objects
-      mode: z
-        .enum([
-          'EDIT_MODE_INPAINT_INSERTION',
-          'EDIT_MODE_INPAINT_REMOVAL',
-          'EDIT_MODE_OUTPAINT',
-          'EDIT_MODE_CONTROLLED_EDITING',
-          'EDIT_MODE_PRODUCT_IMAGE',
-          'EDIT_MODE_BGSWAP',
-        ])
-        .nullish(),
-
-      /**
-       * The mask mode to use.
-       * - `MASK_MODE_DEFAULT` - Default value for mask mode.
-       * - `MASK_MODE_USER_PROVIDED` - User provided mask. No segmentation needed.
-       * - `MASK_MODE_DETECTION_BOX` - Mask from detected bounding boxes.
-       * - `MASK_MODE_CLOTHING_AREA` - Masks from segmenting the clothing area with open-vocab segmentation.
-       * - `MASK_MODE_PARSED_PERSON` - Masks from segmenting the person body and clothing using the person-parsing model.
-       */
-      maskMode: z
-        .enum([
-          'MASK_MODE_DEFAULT',
-          'MASK_MODE_USER_PROVIDED',
-          'MASK_MODE_DETECTION_BOX',
-          'MASK_MODE_CLOTHING_AREA',
-          'MASK_MODE_PARSED_PERSON',
-        ])
-        .nullish(),
-
-      /**
-       * Optional. A float value between 0 and 1, inclusive, that represents the
-       * percentage of the image width to grow the mask by. Using dilation helps
-       * compensate for imprecise masks. We recommend a value of 0.01.
-       */
-      maskDilation: z.number().nullish(),
-    })
-    .nullish(),
-});
-export type GoogleVertexImageModelOptions = z.infer<
-  typeof googleVertexImageModelOptionsSchema
->;
-
 /**
  * Helper to convert ImageModelV4File data to base64 string
  */

package/src/google-vertex-options.ts

@@ -3,6 +3,7 @@
 // https://console.cloud.google.com/vertex-ai/studio/
 export type GoogleVertexModelId =
   // Stable models
+  | 'gemini-3.5-flash'
   | 'gemini-2.5-pro'
   | 'gemini-2.5-flash'
   | 'gemini-2.5-flash-lite'
@@ -21,7 +22,6 @@ 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'