@ai-sdk/google 4.0.79 → 4.0.82

package/docs/15-google.mdx

@@ -923,14 +923,14 @@ The `vertexRagStore` tool accepts the following configuration options:
 
 ### Image Outputs
 
-Gemini models with image generation capabilities (e.g. `gemini-2.5-flash-image`) support generating images as part of a multimodal response. Images are exposed as files in the response.
+Gemini models with image generation capabilities (e.g. `gemini-3.1-flash-image-preview`) support generating images as part of a multimodal response. Images are exposed as files in the response.
 
 ```ts
 import { google } from '@ai-sdk/google';
 import { generateText } from 'ai';
 
 const result = await generateText({
-  model: google('gemini-2.5-flash-image'),
+  model: google('gemini-3.1-flash-image-preview'),
   prompt:
     'Create a picture of a nano banana dish in a fancy restaurant with a Gemini theme',
 });
@@ -1088,7 +1088,7 @@ You can pass a `webhookUrl` to receive a notification when the batch reaches a t
 import { google } from '@ai-sdk/google';
 import { experimental_startBatch as startBatch } from 'ai';
 
-const model = 'gemini-3.6-flash';
+const model = 'gemini-3.8-flash';
 
 const batch = await startBatch({
   provider: google,
@@ -1578,7 +1578,7 @@ import { google, type GoogleInteractionsVideoOptions } from '@ai-sdk/google';
 import { generateText } from 'ai';
 
 const result = await generateText({
-  model: google.interactions('gemini-3.7-flash'),
+  model: google.interactions('gemini-3.8-flash'),
   messages: [
     {
       role: 'user',
@@ -1626,7 +1626,7 @@ These parts are informational and do not require a tool response. They are
 available from both `generateText` and `streamText`.
 
 For static processing, use `'static'` or provide clipping and frame-rate
-options:
+options. `startOffset` and `endOffset` are measured in seconds:
 
 ```ts
 const video = {
@@ -1637,8 +1637,8 @@ const video = {
     google: {
       processing: {
         type: 'static',
-        startOffset: 1200,
-        endOffset: 1500,
+        startOffset: 0,
+        endOffset: 10.5,
         fps: 0.5,
       },
     } satisfies GoogleInteractionsVideoOptions,
@@ -1778,7 +1778,7 @@ import {
 import { generateText } from 'ai';
 
 const result = await generateText({
-  model: google.interactions('gemini-2.5-flash-image'),
+  model: google.interactions('gemini-3.1-flash-image-preview'),
   prompt:
     'Tell me a three sentence bedtime story about a unicorn, accompanied by a suitable illustration.',
   providerOptions: {
@@ -2042,7 +2042,7 @@ You can create models that call the [Google Generative AI embeddings API](https:
 using the `.embedding()` factory method.
 
 ```ts
-const model = google.embedding('gemini-embedding-001');
+const model = google.embedding('gemini-embedding-2');
 ```
 
 The Google provider sends API calls to the right endpoint based on the type of embedding:
@@ -2078,7 +2078,7 @@ import { google, type GoogleEmbeddingModelOptions } from '@ai-sdk/google';
 import { embedMany } from 'ai';
 
 const { embeddings } = await embedMany({
-  model: google.embedding('gemini-embedding-2-preview'),
+  model: google.embedding('gemini-embedding-2'),
   values: ['sunny day at the beach', 'rainy afternoon in the city'],
   providerOptions: {
     google: {
@@ -2132,14 +2132,14 @@ output capabilities through the `:generateContent` API.
 
 ### Gemini Image Models
 
-[Gemini image models](https://ai.google.dev/gemini-api/docs/image-generation) (e.g. `gemini-2.5-flash-image`) are technically multimodal output language models, but they can be used with the `generateImage()` function for a simpler image generation experience. Internally, the provider calls the language model API with `responseModalities: ['IMAGE']`.
+[Gemini image models](https://ai.google.dev/gemini-api/docs/image-generation) (e.g. `gemini-3.1-flash-image-preview`) are technically multimodal output language models, but they can be used with the `generateImage()` function for a simpler image generation experience. Internally, the provider calls the language model API with `responseModalities: ['IMAGE']`.
 
 ```ts
 import { google } from '@ai-sdk/google';
 import { generateImage } from 'ai';
 
 const { image } = await generateImage({
-  model: google.image('gemini-2.5-flash-image'),
+  model: google.image('gemini-3.1-flash-image-preview'),
   prompt: 'A photorealistic image of a cat wearing a wizard hat',
   aspectRatio: '1:1',
 });
@@ -2155,7 +2155,7 @@ import fs from 'node:fs';
 const sourceImage = fs.readFileSync('./cat.png');
 
 const { image } = await generateImage({
-  model: google.image('gemini-2.5-flash-image'),
+  model: google.image('gemini-3.1-flash-image-preview'),
   prompt: {
     text: 'Add a small wizard hat to this cat',
     images: [sourceImage],
@@ -2170,7 +2170,7 @@ import { google } from '@ai-sdk/google';
 import { generateImage } from 'ai';
 
 const { image } = await generateImage({
-  model: google.image('gemini-2.5-flash-image'),
+  model: google.image('gemini-3.1-flash-image-preview'),
   prompt: {
     text: 'Add a small wizard hat to this cat',
     images: ['https://example.com/cat.png'],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/google",
-  "version": "4.0.79",
+  "version": "4.0.82",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -36,10 +36,10 @@
   },
   "dependencies": {
     "@ai-sdk/provider": "4.0.18",
-    "@ai-sdk/provider-utils": "5.0.47"
+    "@ai-sdk/provider-utils": "5.0.49"
   },
   "devDependencies": {
-    "@ai-sdk/test-server": "2.0.1",
+    "@ai-sdk/test-server": "2.0.2",
     "@types/node": "22.19.19",
     "@vercel/ai-tsconfig": "0.0.0",
     "tsup": "^8.5.1",

package/src/convert-to-google-messages.ts

@@ -1,5 +1,6 @@
 import {
   UnsupportedFunctionalityError,
+  type JSONValue,
   type LanguageModelV4Prompt,
   type LanguageModelV4ToolResultOutput,
   type SharedV4Warning,
@@ -8,6 +9,7 @@ import {
   convertToBase64,
   getTopLevelMediaType,
   isFullMediaType,
+  isUrlSupported,
   resolveFullMediaType,
   resolveProviderReference,
   secureJsonParse,
@@ -51,9 +53,6 @@ function parseBase64DataUrl(
 function convertUrlToolResultPart(
   url: string,
 ): GoogleFunctionResponsePart | undefined {
-  // Per https://ai.google.dev/api/caching#FunctionResponsePart, only inline data is supported.
-  // https://docs.cloud.google.com/vertex-ai/generative-ai/docs/model-reference/function-calling#functionresponsepart suggests that this
-  // may be different for Vertex, but this needs to be confirmed and further tested for both APIs.
   const parsedDataUrl = parseBase64DataUrl(url);
   if (parsedDataUrl == null) {
     return undefined;
@@ -67,6 +66,30 @@ function convertUrlToolResultPart(
   };
 }
 
+function containsJSONSchemaReference(value: JSONValue | undefined): boolean {
+  if (Array.isArray(value)) {
+    return value.some(containsJSONSchemaReference);
+  }
+
+  if (typeof value !== 'object' || value === null) {
+    return false;
+  }
+
+  return Object.entries(value).some(
+    ([key, nestedValue]) =>
+      key === '$ref' || containsJSONSchemaReference(nestedValue),
+  );
+}
+
+function serializeFunctionResponseContent(
+  value: JSONValue,
+): JSONValue | string {
+  // Google reserves { $ref: displayName } in structured function responses for
+  // multimodal parts. This conflicts with JSON Schema $ref, so serialize the
+  // result to preserve it without triggering Google's reference handling.
+  return containsJSONSchemaReference(value) ? JSON.stringify(value) : value;
+}
+
 /*
  * Appends tool result content parts to the message using the functionResponse
  * format with support for multimodal parts (e.g. inline images/files alongside
@@ -81,6 +104,7 @@ function appendToolResultParts(
   >['value'],
   toolCallId?: string,
   includeFunctionCallIds = true,
+  supportedUrls: Record<string, RegExp[]> = {},
 ): void {
   const functionResponseParts: GoogleFunctionResponsePart[] = [];
   const responseTextParts: string[] = [];
@@ -100,12 +124,30 @@ function appendToolResultParts(
             },
           });
         } else if (contentPart.data.type === 'url') {
-          const functionResponsePart = convertUrlToolResultPart(
-            contentPart.data.url.toString(),
-          );
-
-          if (functionResponsePart != null) {
-            functionResponseParts.push(functionResponsePart);
+          const url = contentPart.data.url.toString();
+          const convertedUrlPart = convertUrlToolResultPart(url);
+          const supportedUrl =
+            contentPart.data.url.protocol === 'gs:' &&
+            contentPart.data.originalUrl != null
+              ? contentPart.data.originalUrl
+              : url;
+
+          if (convertedUrlPart != null) {
+            functionResponseParts.push(convertedUrlPart);
+          } else if (
+            isFullMediaType(contentPart.mediaType) &&
+            isUrlSupported({
+              url: supportedUrl,
+              mediaType: contentPart.mediaType,
+              supportedUrls,
+            })
+          ) {
+            functionResponseParts.push({
+              fileData: {
+                mimeType: contentPart.mediaType,
+                fileUri: supportedUrl,
+              },
+            });
           } else {
             responseTextParts.push(JSON.stringify(contentPart));
           }
@@ -217,6 +259,7 @@ export function convertToGoogleMessages(
     providerOptionsNames?: readonly string[];
     supportsFunctionResponseParts?: boolean;
     includeFunctionCallIds?: boolean;
+    supportedFunctionResponseUrls?: Record<string, RegExp[]>;
   },
 ): GooglePrompt {
   const systemInstructionParts: Array<{ text: string }> = [];
@@ -230,6 +273,8 @@ export function convertToGoogleMessages(
   const supportsFunctionResponseParts =
     options?.supportsFunctionResponseParts ?? true;
   const includeFunctionCallIds = options?.includeFunctionCallIds ?? true;
+  const supportedFunctionResponseUrls =
+    options?.supportedFunctionResponseUrls ?? {};
 
   let sentinelInjected = false;
   const missingSignatureToolNames: string[] = [];
@@ -632,6 +677,7 @@ export function convertToGoogleMessages(
                 output.value,
                 part.toolCallId,
                 includeFunctionCallIds,
+                supportedFunctionResponseUrls,
               );
             } else {
               appendLegacyToolResultParts(
@@ -654,7 +700,7 @@ export function convertToGoogleMessages(
                   content:
                     output.type === 'execution-denied'
                       ? (output.reason ?? 'Tool call execution denied.')
-                      : output.value,
+                      : serializeFunctionResponseContent(output.value),
                 },
               },
             });

package/src/download-tool-result-files.ts

@@ -6,20 +6,23 @@ import {
   detectMediaType,
   downloadBlob,
   isFullMediaType,
+  isUrlSupported,
 } from '@ai-sdk/provider-utils';
 
 /**
- * Vertex function responses only accept inline file data. Download remote tool
- * result files before converting the prompt to the Google request format.
+ * Download tool result file URLs that Vertex cannot reference directly before
+ * converting the prompt to the Google request format.
  */
 export async function downloadToolResultFiles(
   prompt: LanguageModelV4Prompt,
   {
     abortSignal,
     maxBytes,
+    supportedUrls = {},
   }: {
     abortSignal: AbortSignal | undefined;
     maxBytes: number;
+    supportedUrls?: Record<string, RegExp[]>;
   },
 ): Promise<LanguageModelV4Prompt> {
   const result: LanguageModelV4Prompt = [];
@@ -36,6 +39,7 @@ export async function downloadToolResultFiles(
                 output: await downloadToolResultOutput(part.output, {
                   abortSignal,
                   maxBytes,
+                  supportedUrls,
                 }),
               }
             : part,
@@ -60,6 +64,7 @@ export async function downloadToolResultFiles(
           output: await downloadToolResultOutput(part.output, {
             abortSignal,
             maxBytes,
+            supportedUrls,
           }),
         });
       }
@@ -79,9 +84,11 @@ async function downloadToolResultOutput(
   {
     abortSignal,
     maxBytes,
+    supportedUrls,
   }: {
     abortSignal: AbortSignal | undefined;
     maxBytes: number;
+    supportedUrls: Record<string, RegExp[]>;
   },
 ): Promise<LanguageModelV4ToolResultOutput> {
   if (output.type !== 'content') {
@@ -96,6 +103,17 @@ async function downloadToolResultOutput(
       continue;
     }
 
+    if (
+      isUrlSupported({
+        url: part.data.url.toString(),
+        mediaType: part.mediaType,
+        supportedUrls,
+      })
+    ) {
+      value.push(part);
+      continue;
+    }
+
     const blob = await downloadBlob(part.data.url.toString(), {
       abortSignal,
       maxBytes,

package/src/google-language-model.ts

@@ -66,6 +66,14 @@ const configurableSafetySettingCategories = [
 
 const gemini25ModelPattern = /(^|\/)gemini-2\.5(?:[.-]|$)/i;
 
+const googleCloudStorageFunctionResponseUrls = {
+  'image/png': [/^gs:\/\/.*$/],
+  'image/jpeg': [/^gs:\/\/.*$/],
+  'image/webp': [/^gs:\/\/.*$/],
+  'application/pdf': [/^gs:\/\/.*$/],
+  'text/plain': [/^gs:\/\/.*$/],
+} satisfies Record<string, RegExp[]>;
+
 export type GoogleLanguageModelConfig = {
   provider: string;
   baseURL: string;
@@ -83,6 +91,7 @@ export type GoogleLanguageModelConfig = {
    */
   downloadToolResultFiles?: {
     maxBytes: number;
+    supportsGoogleCloudStorageUrls?: boolean;
   };
 };
 
@@ -295,11 +304,17 @@ export class GoogleLanguageModel implements LanguageModelV4 {
     }
 
     const { usesGemini3Features } = getGoogleModelCapabilities(modelId);
+    const supportedFunctionResponseUrls =
+      usesGemini3Features &&
+      config.downloadToolResultFiles?.supportsGoogleCloudStorageUrls
+        ? googleCloudStorageFunctionResponseUrls
+        : undefined;
 
     const promptWithDownloadedToolResultFiles = config.downloadToolResultFiles
       ? await downloadToolResultFiles(prompt, {
           abortSignal,
           maxBytes: config.downloadToolResultFiles.maxBytes,
+          supportedUrls: supportedFunctionResponseUrls,
         })
       : prompt;
 
@@ -312,6 +327,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
         providerOptionsNames,
         supportsFunctionResponseParts: usesGemini3Features,
         includeFunctionCallIds: !isVertexProvider,
+        supportedFunctionResponseUrls,
       },
     );
 

package/src/google-prompt.ts

@@ -73,9 +73,13 @@ export type GoogleContentPart =
       };
     };
 
-export type GoogleFunctionResponsePart = {
-  inlineData: { mimeType: string; data: string };
-};
+export type GoogleFunctionResponsePart =
+  | {
+      inlineData: { mimeType: string; data: string };
+    }
+  | {
+      fileData: { mimeType: string; fileUri: string };
+    };
 
 export type GoogleGroundingMetadata = GroundingMetadataSchema;
 

package/src/interactions/convert-to-google-interactions-input.ts

@@ -401,10 +401,10 @@ function getVideoProcessingField({
       processing: {
         type: 'static',
         ...(typeof config.startOffset === 'number'
-          ? { start_offset: config.startOffset }
+          ? { start_offset: `${config.startOffset}s` }
           : {}),
         ...(typeof config.endOffset === 'number'
-          ? { end_offset: config.endOffset }
+          ? { end_offset: `${config.endOffset}s` }
           : {}),
         ...(typeof config.fps === 'number' ? { fps: config.fps } : {}),
       },

package/src/interactions/google-interactions-prompt.ts

@@ -40,8 +40,8 @@ export type GoogleInteractionsVideoProcessing =
   | 'static'
   | {
       type: 'static';
-      start_offset?: number;
-      end_offset?: number;
+      start_offset?: string;
+      end_offset?: string;
       fps?: number;
     };