@ai-sdk/google 4.0.86 → 4.0.88

package/docs/15-google.mdx

@@ -2393,18 +2393,18 @@ Older models continue to use labelled transcripts, such as
 | `gemini-3.8-flash-tts`         | <Check />     | <Check />              |
 | `gemini-3.8-flash-lite-tts`    | <Check />     | <Check />              |
 
-## Evaluation Models
+## Decision Models
 
-Create an experimental evaluation model with `google.evaluationModel(modelId)`.
+Create an experimental decision model with `google.decisionModel(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';
+import { experimental_decide } from 'ai';
 
-const { answers } = await experimental_evaluate({
-  model: google.evaluationModel('gemini-3.5-flash-lite'),
+const { answers } = await experimental_decide({
+  model: google.decisionModel('gemini-3.5-flash-lite'),
   state: 'I was charged twice.',
   questions: {
     requestsRefund: {
@@ -2436,10 +2436,10 @@ and rejects safety-blocked, truncated, or invalid output. It preserves usage
 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).
+in application code, for example `answers.requestsRefund.probability >= 0.5`. See [Decisions](/docs/ai-sdk-core/decisions).
 
-Evaluation models can also be accessed through `customProvider` aliases or
-`createProviderRegistry().evaluationModel('provider:model')`. Direct string IDs
+Decision models can also be accessed through `customProvider` aliases or
+`createProviderRegistry().decisionModel('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).
+`decisionModel` method. See
+[model aliases and registries](/docs/ai-sdk-core/decisions#model-aliases-and-registries).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/google",
-  "version": "4.0.86",
+  "version": "4.0.88",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -35,8 +35,8 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/provider": "4.0.20",
-    "@ai-sdk/provider-utils": "5.0.52"
+    "@ai-sdk/provider": "4.0.22",
+    "@ai-sdk/provider-utils": "5.0.54"
   },
   "devDependencies": {
     "@ai-sdk/test-server": "2.0.3",

package/src/google-language-model.ts

@@ -2,6 +2,7 @@ import type {
   LanguageModelV4,
   LanguageModelV4CallOptions,
   LanguageModelV4Content,
+  LanguageModelV4FunctionTool,
   LanguageModelV4FinishReason,
   LanguageModelV4GenerateResult,
   LanguageModelV4Source,
@@ -13,6 +14,7 @@ import type {
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
+  convertJsonResponseToolStream,
   createToolNameMapping,
   createEventSourceResponseHandler,
   createJsonResponseHandler,
@@ -448,16 +450,57 @@ export class GoogleLanguageModel implements LanguageModelV4 {
     };
   }
 
-  private getArgs(
+  private async getArgs(
     options: LanguageModelV4CallOptions,
     { isStreaming = false }: { isStreaming?: boolean } = {},
   ) {
-    return GoogleLanguageModel.prepareRequest({
+    const { responseFormat, tools, toolChoice } = options;
+    let jsonResponseTool: LanguageModelV4FunctionTool | undefined;
+    // Gemini models before Gemini 3 reject JSON response formatting together
+    // with function tools. Gemini 3 supports the combination only with automatic
+    // tool choice (AUTO). Forced tool choice (ANY) requires a JSON response tool.
+    if (
+      responseFormat?.type === 'json' &&
+      responseFormat.schema != null &&
+      tools?.some(tool => tool.type === 'function') &&
+      /(^|\/)gemini-/i.test(this.modelId) &&
+      (!getGoogleModelCapabilities(this.modelId).usesGemini3Features ||
+        toolChoice?.type === 'required' ||
+        toolChoice?.type === 'tool')
+    ) {
+      let name = 'json';
+      for (let suffix = 1; tools.some(tool => tool.name === name); suffix++) {
+        name = `json_${suffix}`;
+      }
+      jsonResponseTool = {
+        type: 'function',
+        name,
+        description:
+          responseFormat.description ?? 'Respond with a JSON object.',
+        inputSchema: responseFormat.schema,
+      };
+    }
+
+    const prepared = await GoogleLanguageModel.prepareRequest({
       modelId: this.modelId,
       config: this.config,
-      options,
+      options:
+        jsonResponseTool == null
+          ? options
+          : {
+              ...options,
+              responseFormat: undefined,
+              tools: [...(tools ?? []), jsonResponseTool],
+              toolChoice:
+                toolChoice?.type === 'tool'
+                  ? toolChoice
+                  : toolChoice?.type === 'none'
+                    ? { type: 'tool', toolName: jsonResponseTool.name }
+                    : { type: 'required' },
+            },
       isStreaming,
     });
+    return { ...prepared, jsonResponseToolName: jsonResponseTool?.name };
   }
 
   static convertGenerateContentResponse({
@@ -466,12 +509,14 @@ export class GoogleLanguageModel implements LanguageModelV4 {
     warnings,
     providerOptionsNames,
     toolNameMapping,
+    jsonResponseToolName,
   }: {
     config: GoogleLanguageModelConfig;
     response: InferSchema<typeof responseSchema>;
     warnings: SharedV4Warning[];
     providerOptionsNames: readonly string[];
     toolNameMapping?: ReturnType<typeof createToolNameMapping>;
+    jsonResponseToolName?: string;
   }): LanguageModelV4GenerateResult {
     const wrapProviderMetadata = (payload: Record<string, unknown>) =>
       Object.fromEntries(
@@ -529,6 +574,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
           },
         });
       } else if ('text' in part && part.text != null) {
+        if (jsonResponseToolName != null && part.thought !== true) continue;
         const thoughtSignatureMetadata = part.thoughtSignature
           ? wrapProviderMetadata({
               thoughtSignature: part.thoughtSignature,
@@ -548,6 +594,23 @@ export class GoogleLanguageModel implements LanguageModelV4 {
           });
         }
       } else if ('functionCall' in part && part.functionCall.name != null) {
+        if (part.functionCall.name === jsonResponseToolName) {
+          content.push({
+            type: 'text',
+            text:
+              typeof part.functionCall.args === 'string'
+                ? part.functionCall.args
+                : JSON.stringify(part.functionCall.args ?? {}),
+            ...(part.thoughtSignature
+              ? {
+                  providerMetadata: wrapProviderMetadata({
+                    thoughtSignature: part.thoughtSignature,
+                  }),
+                }
+              : {}),
+          });
+          continue;
+        }
         content.push({
           type: 'tool-call' as const,
           toolCallId: part.functionCall.id || config.generateId(),
@@ -666,6 +729,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       providerOptionsNames,
       extraHeaders,
       toolNameMapping,
+      jsonResponseToolName,
     } = await this.getArgs(options);
 
     const mergedHeaders = combineHeaders(
@@ -696,6 +760,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       warnings,
       providerOptionsNames,
       toolNameMapping,
+      jsonResponseToolName,
     });
 
     return {
@@ -718,6 +783,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       providerOptionsNames,
       extraHeaders,
       toolNameMapping,
+      jsonResponseToolName,
     } = await this.getArgs(options, { isStreaming: true });
     const wrapProviderMetadata = (payload: Record<string, unknown>) =>
       Object.fromEntries(
@@ -813,7 +879,7 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       hasToolCalls = true;
     };
 
-    return {
+    const result = {
       stream: response.pipeThrough(
         new TransformStream<
           ParseResult<ChunkSchema>,
@@ -1301,6 +1367,15 @@ export class GoogleLanguageModel implements LanguageModelV4 {
       response: { headers: responseHeaders },
       request: { body: args },
     };
+    return {
+      ...result,
+      stream:
+        jsonResponseToolName == null
+          ? result.stream
+          : result.stream.pipeThrough(
+              convertJsonResponseToolStream(jsonResponseToolName),
+            ),
+    };
   }
 }
 

package/src/google-provider.ts

@@ -1,7 +1,7 @@
 import type {
   EmbeddingModelV4,
   Experimental_BatchV4 as BatchV4,
-  Experimental_EvaluationModelV4 as EvaluationModelV4,
+  Experimental_DecisionModelV4 as DecisionModelV4,
   Experimental_VideoModelV4,
   FilesV4,
   ImageModelV4,
@@ -21,7 +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 { Experimental_DecisionLanguageModel as DecisionLanguageModel } from '@ai-sdk/provider-utils/experimental-decision';
 import { VERSION } from './version';
 import { GoogleEmbeddingModel } from './google-embedding-model';
 import type { GoogleEmbeddingModelId } from './google-embedding-model-options';
@@ -63,8 +63,12 @@ export interface GoogleProvider extends ProviderV4 {
 
   chat(modelId: GoogleModelId): LanguageModelV4;
 
-  /** Creates an experimental Choice/Score/Boolean evaluation model using Gemini. */
-  evaluationModel(modelId: GoogleModelId): EvaluationModelV4;
+  /** Creates an experimental Choice/Score/Boolean decision model using Gemini. */
+  decisionModel(modelId: GoogleModelId): DecisionModelV4;
+  /** @deprecated Use `decisionModel` instead. */
+  evaluationModel(modelId: GoogleModelId): DecisionModelV4 & {
+    doEvaluate: DecisionModelV4['doDecide'];
+  };
 
   experimental_batch(): BatchV4<{
     text: GoogleModelId;
@@ -435,11 +439,13 @@ export function createGoogle(
   provider.languageModel = createChatModel;
   provider.chat = createChatModel;
   provider.generativeAI = createChatModel;
-  provider.evaluationModel = (modelId: GoogleModelId) =>
-    new EvaluationLanguageModel({
+  provider.decisionModel = (modelId: GoogleModelId) =>
+    new DecisionLanguageModel({
       model: createChatModel(modelId),
-      provider: `${providerName.replace(/\.generative-ai$/, '')}.evaluation`,
+      provider: `${providerName.replace(/\.generative-ai$/, '')}.decision`,
     });
+  provider.evaluationModel =
+    provider.decisionModel as GoogleProvider['evaluationModel'];
   provider.experimental_batch = createBatch;
   provider.embedding = createEmbeddingModel;
   provider.embeddingModel = createEmbeddingModel;

package/src/google-speech-api.ts

@@ -32,6 +32,7 @@ export const googleSpeechResponseSchema = lazySchema(() =>
           }),
         )
         .nullish(),
+      usageMetadata: z.record(z.string(), z.json()).nullish(),
     }),
   ),
 );

package/src/google-speech-model.ts

@@ -353,6 +353,9 @@ export class GoogleSpeechModel implements SpeechModelV4 {
     return {
       audio,
       warnings,
+      ...(response.usageMetadata != null && {
+        usage: response.usageMetadata,
+      }),
       request: {
         body: JSON.stringify(requestBody),
       },

package/src/transcription/google-transcription-model.ts

@@ -217,6 +217,7 @@ export class GoogleTranscriptionModel implements TranscriptionModelV4 {
       language: undefined,
       durationInSeconds: undefined,
       warnings,
+      ...(response.usage != null && { usage: response.usage }),
       response: {
         timestamp: currentDate,
         modelId: this.modelId,
@@ -226,7 +227,7 @@ export class GoogleTranscriptionModel implements TranscriptionModelV4 {
       ...(response.usage != null
         ? {
             providerMetadata: {
-              google: { usage: response.usage as JSONObject },
+              google: { usage: response.usage },
             },
           }
         : {}),
@@ -434,6 +435,7 @@ function createGoogleLiveTranscriptionStream({
           segments: [],
           language,
           durationInSeconds: undefined,
+          usage: usageMetadata,
           ...(usageMetadata != null
             ? { providerMetadata: { google: { usageMetadata } } }
             : {}),
@@ -710,5 +712,5 @@ const googleInteractionsTranscriptionResponseSchema = z.object({
       }),
     )
     .nullish(),
-  usage: z.record(z.string(), z.unknown()).nullish(),
+  usage: z.record(z.string(), z.json()).nullish(),
 });