@ai-sdk/google 4.0.73 → 4.0.75

package/docs/15-google.mdx

@@ -2304,3 +2304,54 @@ const result = await generateSpeech({
 | `gemini-2.5-flash-preview-tts` | <Check />     | <Check />              |
 | `gemini-2.5-pro-preview-tts`   | <Check />     | <Check />              |
 | `gemini-3.1-flash-tts-preview` | <Check />     | <Check />              |
+
+## Evaluation Models
+
+Create an experimental evaluation model with `google.evaluationModel(modelId)`.
+It uses Gemini structured output for Choice, Score, and Boolean questions. Choose a model
+that supports structured output, such as `gemini-3.5-flash-lite`.
+
+```ts
+import { google } from '@ai-sdk/google';
+import { experimental_evaluate } from 'ai';
+
+const { answers } = await experimental_evaluate({
+  model: google.evaluationModel('gemini-3.5-flash-lite'),
+  state: 'I was charged twice.',
+  questions: {
+    requestsRefund: {
+      type: 'boolean',
+      instructions: 'Is the customer requesting money back?',
+    },
+    department: {
+      type: 'choice',
+      instructions: 'Which team should handle this?',
+      criteria: { billing: 'Charges and refunds', support: 'Other requests' },
+    },
+  },
+  providerOptions: {
+    google: { thinkingConfig: { thinkingLevel: 'minimal' } },
+  },
+});
+```
+
+The factory respects `createGoogle` settings and forwards `providerOptions.google`,
+including thinking configuration. Thinking controls depend on the selected model;
+use options supported by that model. The existing Gemini implementation sends the
+answer schema as `generationConfig.responseJsonSchema`.
+
+Choice labels are returned exactly, and Scores are finite fractional positions
+within the ordered rubric. The adapter validates complete answers after parsing
+and rejects safety-blocked, truncated, or invalid output. It preserves usage
+(including reasoning tokens), warnings, response data, and provider metadata.
+
+Choice and Score answers do not include probability distributions. Boolean
+answers contain prompted estimates of P(true), validated to be finite and in
+`[0, 1]`. These estimates are not guaranteed to be calibrated. Apply thresholds
+in application code, for example `answers.requestsRefund.probability >= 0.5`. See [Evaluation](/docs/ai-sdk-core/evaluation).
+
+Evaluation models can also be accessed through `customProvider` aliases or
+`createProviderRegistry().evaluationModel('provider:model')`. Direct string IDs
+use Gateway by default, or an explicitly configured default provider with an
+`evaluationModel` method. See
+[model aliases and registries](/docs/ai-sdk-core/evaluation#model-aliases-and-registries).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/google",
-  "version": "4.0.73",
+  "version": "4.0.75",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -35,8 +35,8 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/provider": "4.0.16",
-    "@ai-sdk/provider-utils": "5.0.42"
+    "@ai-sdk/provider": "4.0.17",
+    "@ai-sdk/provider-utils": "5.0.44"
   },
   "devDependencies": {
     "@ai-sdk/test-server": "2.0.1",

package/src/convert-google-usage.ts

@@ -30,13 +30,15 @@ export function convertGoogleUsage(
 
   const promptTokens = usage.promptTokenCount ?? 0;
   const candidatesTokens = usage.candidatesTokenCount ?? 0;
+  const toolUsePromptTokens = usage.toolUsePromptTokenCount ?? 0;
   const cachedContentTokens = usage.cachedContentTokenCount ?? 0;
   const thoughtsTokens = usage.thoughtsTokenCount ?? 0;
+  const inputTokens = promptTokens + toolUsePromptTokens;
 
   return {
     inputTokens: {
-      total: promptTokens,
-      noCache: promptTokens - cachedContentTokens,
+      total: inputTokens,
+      noCache: inputTokens - cachedContentTokens,
       cacheRead: cachedContentTokens,
       cacheWrite: undefined,
     },

package/src/google-image-model.ts

@@ -53,7 +53,7 @@ export class GoogleImageModel implements ImageModelV4 {
     if (this.settings.maxImagesPerCall != null) {
       return this.settings.maxImagesPerCall;
     }
-    return 10;
+    return 1;
   }
 
   get provider(): string {
@@ -77,7 +77,6 @@ export class GoogleImageModel implements ImageModelV4 {
 
     const {
       prompt,
-      n,
       size,
       aspectRatio,
       seed,
@@ -96,13 +95,6 @@ export class GoogleImageModel implements ImageModelV4 {
       );
     }
 
-    // Gemini does not support generating multiple images per call via n parameter
-    if (n != null && n > 1) {
-      throw new Error(
-        'Gemini image models do not support generating a set number of images per call. Use n=1 or omit the n parameter.',
-      );
-    }
-
     if (size != null) {
       warnings.push({
         type: 'unsupported',

package/src/google-image-settings.ts

@@ -7,7 +7,7 @@ export type GoogleImageModelId =
 
 export interface GoogleImageSettings {
   /**
-   * Override the maximum number of images per call (default 10)
+   * Override the maximum number of images per call (default 1)
    */
   maxImagesPerCall?: number;
 }

package/src/google-language-model.ts

@@ -463,10 +463,15 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       ) as SharedV4ProviderMetadata;
     const candidate = response.candidates?.[0];
     const promptBlockReason = response.promptFeedback?.blockReason;
+    const confirmedPromptBlockReason = isConfirmedPromptBlockReason(
+      promptBlockReason,
+    )
+      ? promptBlockReason
+      : undefined;
     const isPromptBlocked =
-      candidate?.finishReason == null && promptBlockReason != null;
+      candidate?.finishReason == null && confirmedPromptBlockReason != null;
     const rawFinishReason =
-      candidate?.finishReason ?? promptBlockReason ?? undefined;
+      candidate?.finishReason ?? confirmedPromptBlockReason;
     const content: Array<LanguageModelV4Content> = [];
 
     // map ordered parts to content:

package/src/google-provider.ts

@@ -1,6 +1,7 @@
 import type {
   EmbeddingModelV4,
   Experimental_BatchV4 as BatchV4,
+  Experimental_EvaluationModelV4 as EvaluationModelV4,
   Experimental_VideoModelV4,
   FilesV4,
   ImageModelV4,
@@ -20,6 +21,7 @@ import {
   type FetchFunction,
   type WebSocketConstructor,
 } from '@ai-sdk/provider-utils';
+import { Experimental_EvaluationLanguageModel as EvaluationLanguageModel } from '@ai-sdk/provider-utils/experimental-evaluation';
 import { VERSION } from './version';
 import { GoogleEmbeddingModel } from './google-embedding-model';
 import type { GoogleEmbeddingModelId } from './google-embedding-model-options';
@@ -61,6 +63,9 @@ export interface GoogleProvider extends ProviderV4 {
 
   chat(modelId: GoogleModelId): LanguageModelV4;
 
+  /** Creates an experimental Choice/Score/Boolean evaluation model using Gemini. */
+  evaluationModel(modelId: GoogleModelId): EvaluationModelV4;
+
   experimental_batch(): BatchV4<{
     text: GoogleModelId;
     image: GoogleImageModelId;
@@ -430,6 +435,11 @@ export function createGoogle(
   provider.languageModel = createChatModel;
   provider.chat = createChatModel;
   provider.generativeAI = createChatModel;
+  provider.evaluationModel = (modelId: GoogleModelId) =>
+    new EvaluationLanguageModel({
+      model: createChatModel(modelId),
+      provider: `${providerName.replace(/\.generative-ai$/, '')}.evaluation`,
+    });
   provider.experimental_batch = createBatch;
   provider.embedding = createEmbeddingModel;
   provider.embeddingModel = createEmbeddingModel;