@ai-sdk/google 4.0.72 → 4.0.74

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
+require an explicitly configured default provider with an `evaluationModel`
+method; they do not automatically use Gateway. 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": "4.0.72",
+  "version": "4.0.74",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -35,8 +35,8 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/provider": "4.0.15",
-    "@ai-sdk/provider-utils": "5.0.41"
+    "@ai-sdk/provider": "4.0.17",
+    "@ai-sdk/provider-utils": "5.0.43"
   },
   "devDependencies": {
     "@ai-sdk/test-server": "2.0.1",

package/src/google-language-model.ts

@@ -726,9 +726,12 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       raw: undefined,
     };
     let usage: GoogleUsageMetadata | undefined = undefined;
-    let providerMetadata: SharedV4ProviderMetadata | undefined = undefined;
+    let promptFeedback: PromptFeedbackSchema | null = null;
     let lastGroundingMetadata: GroundingMetadataSchema | null = null;
     let lastUrlContextMetadata: UrlContextMetadataSchema | null = null;
+    let lastSafetyRatings: SafetyRatingSchema[] | null = null;
+    let lastFinishMessage: string | null = null;
+    let confirmedPromptBlockReason: string | undefined;
 
     const generateId = this.config.generateId;
     let hasToolCalls = false;
@@ -825,38 +828,48 @@ export class GoogleLanguageModel implements LanguageModelV4 {
               usage = usageMetadata;
             }
 
-            const candidate = value.candidates?.[0];
+            if (
+              value.promptFeedback != null &&
+              confirmedPromptBlockReason == null
+            ) {
+              promptFeedback = value.promptFeedback;
 
-            // sometimes the API returns an empty candidates array
-            if (candidate == null) {
-              const promptBlockReason = value.promptFeedback?.blockReason;
-              if (promptBlockReason != null) {
+              if (
+                isConfirmedPromptBlockReason(value.promptFeedback.blockReason)
+              ) {
+                confirmedPromptBlockReason = value.promptFeedback.blockReason;
                 finishReason = {
                   unified: 'content-filter',
-                  raw: promptBlockReason,
+                  raw: confirmedPromptBlockReason,
                 };
-                providerMetadata = wrapProviderMetadata({
-                  promptFeedback: value.promptFeedback ?? null,
-                  groundingMetadata: lastGroundingMetadata,
-                  urlContextMetadata: lastUrlContextMetadata,
-                  safetyRatings: null,
-                  usageMetadata: usageMetadata ?? null,
-                  finishMessage: null,
-                  serviceTier: usage?.serviceTier ?? null,
-                } satisfies GoogleProviderMetadata);
               }
-              return;
             }
 
-            const content = candidate.content;
+            const candidate = value.candidates?.[0];
 
-            if (candidate.groundingMetadata != null) {
-              lastGroundingMetadata = candidate.groundingMetadata;
+            if (candidate != null) {
+              if (candidate.groundingMetadata != null) {
+                lastGroundingMetadata = candidate.groundingMetadata;
+              }
+              if (candidate.urlContextMetadata != null) {
+                lastUrlContextMetadata = candidate.urlContextMetadata;
+              }
+              if (candidate.safetyRatings != null) {
+                lastSafetyRatings = candidate.safetyRatings;
+              }
+              if (candidate.finishMessage != null) {
+                lastFinishMessage = candidate.finishMessage;
+              }
             }
-            if (candidate.urlContextMetadata != null) {
-              lastUrlContextMetadata = candidate.urlContextMetadata;
+
+            // A confirmed prompt block is terminal for generated content, but
+            // later chunks can still contribute usage and provider metadata.
+            if (confirmedPromptBlockReason != null || candidate == null) {
+              return;
             }
 
+            const content = candidate.content;
+
             const sources = extractSources({
               groundingMetadata: candidate.groundingMetadata,
               generateId,
@@ -1222,32 +1235,14 @@ export class GoogleLanguageModel implements LanguageModelV4 {
               }
             }
 
-            const promptBlockReason = value.promptFeedback?.blockReason;
-            const isPromptBlocked =
-              candidate.finishReason == null && promptBlockReason != null;
-            const rawFinishReason =
-              candidate.finishReason ?? promptBlockReason ?? undefined;
-
-            if (rawFinishReason != null) {
+            if (candidate.finishReason != null) {
               finishReason = {
-                unified: isPromptBlocked
-                  ? 'content-filter'
-                  : mapGoogleFinishReason({
-                      finishReason: rawFinishReason,
-                      hasToolCalls,
-                    }),
-                raw: rawFinishReason,
+                unified: mapGoogleFinishReason({
+                  finishReason: candidate.finishReason,
+                  hasToolCalls,
+                }),
+                raw: candidate.finishReason,
               };
-
-              providerMetadata = wrapProviderMetadata({
-                promptFeedback: value.promptFeedback ?? null,
-                groundingMetadata: lastGroundingMetadata,
-                urlContextMetadata: lastUrlContextMetadata,
-                safetyRatings: candidate.safetyRatings ?? null,
-                usageMetadata: usageMetadata ?? null,
-                finishMessage: candidate.finishMessage ?? null,
-                serviceTier: usage?.serviceTier ?? null,
-              } satisfies GoogleProviderMetadata);
             }
           },
 
@@ -1269,7 +1264,15 @@ export class GoogleLanguageModel implements LanguageModelV4 {
               type: 'finish',
               finishReason,
               usage: convertGoogleUsage(usage),
-              providerMetadata,
+              providerMetadata: wrapProviderMetadata({
+                promptFeedback,
+                groundingMetadata: lastGroundingMetadata,
+                urlContextMetadata: lastUrlContextMetadata,
+                safetyRatings: lastSafetyRatings,
+                usageMetadata: usage ?? null,
+                finishMessage: lastFinishMessage,
+                serviceTier: usage?.serviceTier ?? null,
+              } satisfies GoogleProviderMetadata),
             });
           },
         }),
@@ -1794,3 +1797,14 @@ const chunkSchema = lazySchema(() =>
 );
 
 type ChunkSchema = InferSchema<typeof chunkSchema>;
+
+function isConfirmedPromptBlockReason(
+  blockReason: string | null | undefined,
+): blockReason is string {
+  return (
+    blockReason != null &&
+    blockReason !== '' &&
+    blockReason !== 'BLOCK_REASON_UNSPECIFIED' &&
+    blockReason !== 'BLOCKED_REASON_UNSPECIFIED'
+  );
+}

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;