@ai-sdk/google 3.0.114 → 3.0.118

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/google",
-  "version": "3.0.114",
+  "version": "3.0.118",
   "license": "Apache-2.0",
   "sideEffects": false,
   "main": "./dist/index.js",
@@ -37,7 +37,7 @@
   },
   "dependencies": {
     "@ai-sdk/provider": "3.0.15",
-    "@ai-sdk/provider-utils": "4.0.48"
+    "@ai-sdk/provider-utils": "4.0.49"
   },
   "devDependencies": {
     "@types/node": "20.17.24",

package/src/convert-json-schema-to-openapi-schema.ts

@@ -100,10 +100,6 @@ function convertJSONSchemaDefinition(
   if (required) result.required = required;
   if (format) result.format = format;
 
-  if (constValue !== undefined) {
-    result.enum = [constValue];
-  }
-
   // Handle type
   if (type) {
     if (Array.isArray(type)) {
@@ -125,9 +121,11 @@ function convertJSONSchemaDefinition(
     }
   }
 
-  // Handle enum
-  if (enumValues !== undefined) {
-    result.enum = enumValues;
+  const values =
+    enumValues ?? (constValue !== undefined ? [constValue] : undefined);
+
+  if (values !== undefined) {
+    addEnumToSchema({ values, type, result });
   }
 
   if (properties != null) {
@@ -201,6 +199,123 @@ function convertJSONSchemaDefinition(
   return result;
 }
 
+type EnumValues = NonNullable<JSONSchema7['enum']>;
+type EnumType = 'string' | 'number' | 'integer' | 'boolean';
+type GoogleEnumSchema = {
+  type?: JSONSchema7['type'];
+  enum?: JSONSchema7['enum'];
+  format?: JSONSchema7['format'];
+  anyOf?: JSONSchema7['anyOf'];
+  nullable?: boolean;
+};
+
+function addEnumToSchema({
+  values,
+  type,
+  result,
+}: {
+  values: EnumValues;
+  type: JSONSchema7['type'];
+  result: GoogleEnumSchema;
+}) {
+  const nullable =
+    (Array.isArray(type) && type.includes('null')) ||
+    (type === undefined && values.includes(null));
+
+  // Gemini uses nullable instead of a null enum member.
+  const enumValues = nullable ? values.filter(value => value !== null) : values;
+
+  if (values.length > 0 && values.every(value => value === null)) {
+    const typeAllowsNull =
+      type === undefined ||
+      type === 'null' ||
+      (Array.isArray(type) && type.includes('null'));
+
+    if (typeAllowsNull) {
+      result.type = 'null';
+      if (Array.isArray(type)) {
+        delete result.anyOf;
+      }
+      return;
+    }
+  }
+
+  const enumType = getEnumType({ values: enumValues, type });
+
+  if (enumType === undefined) {
+    throw new UnsupportedFunctionalityError({
+      functionality: 'JSON Schema enum with mixed or unsupported values',
+      message:
+        'Google does not support this JSON Schema enum. Enum values must share one supported primitive type and match the schema type.',
+    });
+  }
+
+  result.type = enumType;
+
+  // The earlier type-array conversion created anyOf. The enum gives us one
+  // concrete value type, so store that type directly.
+  if (Array.isArray(type)) {
+    delete result.anyOf;
+  }
+
+  if (nullable) {
+    result.nullable = true;
+  }
+
+  if (enumType === 'string') {
+    result.enum = enumValues;
+  } else {
+    result.format = 'enum';
+    result.enum = enumValues.map(String);
+  }
+}
+
+function getEnumType({
+  values,
+  type,
+}: {
+  values: EnumValues;
+  type: JSONSchema7['type'];
+}): EnumType | undefined {
+  if (values.length === 0) {
+    return undefined;
+  }
+
+  const typeAllows = (enumType: EnumType) =>
+    type === undefined ||
+    type === enumType ||
+    (Array.isArray(type) && type.includes(enumType));
+
+  if (
+    typeAllows('string') &&
+    values.every(value => typeof value === 'string')
+  ) {
+    return 'string';
+  }
+
+  if (
+    (typeAllows('number') || typeAllows('integer')) &&
+    values.every(value => typeof value === 'number' && Number.isFinite(value))
+  ) {
+    if (typeAllows('number')) {
+      return 'number';
+    }
+
+    if (values.every(value => Number.isInteger(value))) {
+      return 'integer';
+    }
+  }
+
+  if (
+    typeAllows('boolean') &&
+    values.every(value => typeof value === 'boolean')
+  ) {
+    return 'boolean';
+  }
+
+  return undefined;
+}
+
 function convertJSONSchemaReference({
   jsonSchema,
   reference,
@@ -318,6 +433,7 @@ function throwUnsupportedReference(reference: string): never {
       'Google schema conversion only supports references to direct children of root-level $defs or definitions.',
   });
 }
+
 function isEmptyObjectSchema(jsonSchema: JSONSchema7Definition): boolean {
   return (
     jsonSchema != null &&

package/src/google-generative-ai-language-model.ts

@@ -58,6 +58,8 @@ const configurableSafetySettingCategories = [
   'HARM_CATEGORY_SEXUALLY_EXPLICIT',
 ] as const;
 
+const gemini25ModelPattern = /(^|\/)gemini-2\.5(?:[.-]|$)/i;
+
 type GoogleGenerativeAIConfig = {
   provider: string;
   baseURL: string;
@@ -230,6 +232,22 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
     }
 
     const isGemmaModel = this.modelId.toLowerCase().startsWith('gemma-');
+    const isGemini25DeveloperApiModel =
+      !isVertexProvider && gemini25ModelPattern.test(this.modelId);
+
+    if (isGemini25DeveloperApiModel && frequencyPenalty != null) {
+      warnings.push({
+        type: 'unsupported',
+        feature: 'frequencyPenalty',
+      });
+    }
+    if (isGemini25DeveloperApiModel && presencePenalty != null) {
+      warnings.push({
+        type: 'unsupported',
+        feature: 'presencePenalty',
+      });
+    }
+
     const { usesGemini3Features } = getGoogleModelCapabilities(this.modelId);
 
     const { contents, systemInstruction } = convertToGoogleGenerativeAIMessages(
@@ -296,8 +314,12 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
           temperature,
           topK,
           topP,
-          frequencyPenalty,
-          presencePenalty,
+          frequencyPenalty: isGemini25DeveloperApiModel
+            ? undefined
+            : frequencyPenalty,
+          presencePenalty: isGemini25DeveloperApiModel
+            ? undefined
+            : presencePenalty,
           stopSequences,
           seed,
 
@@ -368,11 +390,16 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
       fetch: this.config.fetch,
     });
 
-    const candidate = response.candidates[0];
+    const candidate = response.candidates?.[0];
+    const promptBlockReason = response.promptFeedback?.blockReason;
+    const isPromptBlocked =
+      candidate?.finishReason == null && promptBlockReason != null;
+    const rawFinishReason =
+      candidate?.finishReason ?? promptBlockReason ?? undefined;
     const content: Array<LanguageModelV3Content> = [];
 
     // map ordered parts to content:
-    const parts = candidate.content?.parts ?? [];
+    const parts = candidate?.content?.parts ?? [];
 
     const usageMetadata = response.usageMetadata;
 
@@ -515,7 +542,7 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
 
     const sources =
       extractSources({
-        groundingMetadata: candidate.groundingMetadata,
+        groundingMetadata: candidate?.groundingMetadata,
         generateId: this.config.generateId,
       }) ?? [];
     for (const source of sources) {
@@ -525,25 +552,27 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
     return {
       content,
       finishReason: {
-        unified: mapGoogleGenerativeAIFinishReason({
-          finishReason: candidate.finishReason,
-          // Only count client-executed tool calls for finish reason determination.
-          hasToolCalls: content.some(
-            part => part.type === 'tool-call' && !part.providerExecuted,
-          ),
-        }),
-        raw: candidate.finishReason ?? undefined,
+        unified: isPromptBlocked
+          ? 'content-filter'
+          : mapGoogleGenerativeAIFinishReason({
+              finishReason: rawFinishReason,
+              // Only count client-executed tool calls for finish reason determination.
+              hasToolCalls: content.some(
+                part => part.type === 'tool-call' && !part.providerExecuted,
+              ),
+            }),
+        raw: rawFinishReason,
       },
       usage: convertGoogleGenerativeAIUsage(usageMetadata),
       warnings,
       providerMetadata: {
         [providerOptionsName]: {
           promptFeedback: response.promptFeedback ?? null,
-          groundingMetadata: candidate.groundingMetadata ?? null,
-          urlContextMetadata: candidate.urlContextMetadata ?? null,
-          safetyRatings: candidate.safetyRatings ?? null,
+          groundingMetadata: candidate?.groundingMetadata ?? null,
+          urlContextMetadata: candidate?.urlContextMetadata ?? null,
+          safetyRatings: candidate?.safetyRatings ?? null,
           usageMetadata: usageMetadata ?? null,
-          finishMessage: candidate.finishMessage ?? null,
+          finishMessage: candidate?.finishMessage ?? null,
           serviceTier: usageMetadata?.serviceTier ?? null,
         } satisfies GoogleGenerativeAIProviderMetadata,
       },
@@ -689,6 +718,24 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
 
             // sometimes the API returns an empty candidates array
             if (candidate == null) {
+              const promptBlockReason = value.promptFeedback?.blockReason;
+              if (promptBlockReason != null) {
+                finishReason = {
+                  unified: 'content-filter',
+                  raw: promptBlockReason,
+                };
+                providerMetadata = {
+                  [providerOptionsName]: {
+                    promptFeedback: value.promptFeedback ?? null,
+                    groundingMetadata: lastGroundingMetadata,
+                    urlContextMetadata: lastUrlContextMetadata,
+                    safetyRatings: null,
+                    usageMetadata: usageMetadata ?? null,
+                    finishMessage: null,
+                    serviceTier: usage?.serviceTier ?? null,
+                  } satisfies GoogleGenerativeAIProviderMetadata,
+                };
+              }
               return;
             }
 
@@ -1078,13 +1125,21 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
               }
             }
 
-            if (candidate.finishReason != null) {
+            const promptBlockReason = value.promptFeedback?.blockReason;
+            const isPromptBlocked =
+              candidate.finishReason == null && promptBlockReason != null;
+            const rawFinishReason =
+              candidate.finishReason ?? promptBlockReason ?? undefined;
+
+            if (rawFinishReason != null) {
               finishReason = {
-                unified: mapGoogleGenerativeAIFinishReason({
-                  finishReason: candidate.finishReason,
-                  hasToolCalls,
-                }),
-                raw: candidate.finishReason,
+                unified: isPromptBlocked
+                  ? 'content-filter'
+                  : mapGoogleGenerativeAIFinishReason({
+                      finishReason: rawFinishReason,
+                      hasToolCalls,
+                    }),
+                raw: rawFinishReason,
               };
 
               providerMetadata = {
@@ -1464,16 +1519,18 @@ const responseSchema = lazySchema(() =>
   zodSchema(
     z.object({
       responseId: z.string().nullish(),
-      candidates: z.array(
-        z.object({
-          content: getContentSchema().nullish().or(z.object({}).strict()),
-          finishReason: z.string().nullish(),
-          finishMessage: z.string().nullish(),
-          safetyRatings: z.array(getSafetyRatingSchema()).nullish(),
-          groundingMetadata: getGroundingMetadataSchema().nullish(),
-          urlContextMetadata: getUrlContextMetadataSchema().nullish(),
-        }),
-      ),
+      candidates: z
+        .array(
+          z.object({
+            content: getContentSchema().nullish().or(z.object({}).strict()),
+            finishReason: z.string().nullish(),
+            finishMessage: z.string().nullish(),
+            safetyRatings: z.array(getSafetyRatingSchema()).nullish(),
+            groundingMetadata: getGroundingMetadataSchema().nullish(),
+            urlContextMetadata: getUrlContextMetadataSchema().nullish(),
+          }),
+        )
+        .nullish(),
       usageMetadata: usageSchema.nullish(),
       promptFeedback: z
         .object({
@@ -1485,11 +1542,14 @@ const responseSchema = lazySchema(() =>
   ),
 );
 
-type ContentSchema = NonNullable<
-  InferSchema<typeof responseSchema>['candidates'][number]['content']
->;
+type CandidateSchema = NonNullable<
+  InferSchema<typeof responseSchema>['candidates']
+>[number];
+
+type ContentSchema = NonNullable<CandidateSchema['content']>;
+
 export type GroundingMetadataSchema = NonNullable<
-  InferSchema<typeof responseSchema>['candidates'][number]['groundingMetadata']
+  CandidateSchema['groundingMetadata']
 >;
 
 type GroundingChunkSchema = NonNullable<
@@ -1497,11 +1557,11 @@ type GroundingChunkSchema = NonNullable<
 >[number];
 
 export type UrlContextMetadataSchema = NonNullable<
-  InferSchema<typeof responseSchema>['candidates'][number]['urlContextMetadata']
+  CandidateSchema['urlContextMetadata']
 >;
 
 export type SafetyRatingSchema = NonNullable<
-  InferSchema<typeof responseSchema>['candidates'][number]['safetyRatings']
+  CandidateSchema['safetyRatings']
 >[number];
 
 export type PromptFeedbackSchema = NonNullable<

package/src/google-provider.ts

@@ -4,6 +4,7 @@ import type {
   ImageModelV3,
   LanguageModelV3,
   ProviderV3,
+  TranscriptionModelV3,
 } from '@ai-sdk/provider';
 import {
   generateId,
@@ -32,6 +33,8 @@ import {
 } from './interactions/google-interactions-language-model';
 import type { GoogleInteractionsModelId } from './interactions/google-interactions-language-model-options';
 import type { GoogleInteractionsAgentName } from './interactions/google-interactions-agent';
+import { GoogleTranscriptionModel } from './transcription/google-transcription-model';
+import type { GoogleTranscriptionModelId } from './transcription/google-transcription-model-options';
 
 export interface GoogleGenerativeAIProvider extends ProviderV3 {
   (modelId: GoogleGenerativeAIModelId): LanguageModelV3;
@@ -75,6 +78,17 @@ export interface GoogleGenerativeAIProvider extends ProviderV3 {
     modelId: GoogleGenerativeAIEmbeddingModelId,
   ): EmbeddingModelV3;
 
+  /**
+   * Creates a model for transcription (speech-to-text), e.g.
+   * `gemini-3.5-transcribe`.
+   */
+  transcription(modelId: GoogleTranscriptionModelId): TranscriptionModelV3;
+
+  /**
+   * Creates a model for transcription (speech-to-text).
+   */
+  transcriptionModel(modelId: GoogleTranscriptionModelId): TranscriptionModelV3;
+
   /**
    * Creates a model for video generation.
    */
@@ -245,6 +259,14 @@ export function createGoogleGenerativeAI(
       fetch: options.fetch,
     });
 
+  const createTranscriptionModel = (modelId: GoogleTranscriptionModelId) =>
+    new GoogleTranscriptionModel(modelId, {
+      provider: `${providerName}.transcription`,
+      baseURL,
+      headers: getHeaders,
+      fetch: options.fetch,
+    });
+
   const createVideoModel = (modelId: GoogleGenerativeAIVideoModelId) =>
     new GoogleGenerativeAIVideoModel(modelId, {
       provider: providerName,
@@ -291,6 +313,8 @@ export function createGoogleGenerativeAI(
   provider.textEmbeddingModel = createEmbeddingModel;
   provider.image = createImageModel;
   provider.imageModel = createImageModel;
+  provider.transcription = createTranscriptionModel;
+  provider.transcriptionModel = createTranscriptionModel;
   provider.video = createVideoModel;
   provider.videoModel = createVideoModel;
   provider.interactions = createInteractionsModel;

package/src/index.ts

@@ -27,6 +27,11 @@ export type {
 } from './interactions/google-interactions-language-model-options';
 export type { GoogleInteractionsProviderMetadata } from './interactions/google-interactions-provider-metadata';
 export type { GoogleInteractionsAgentName } from './interactions/google-interactions-agent';
+export { GoogleTranscriptionModel } from './transcription/google-transcription-model';
+export type {
+  GoogleTranscriptionModelId,
+  GoogleTranscriptionModelOptions,
+} from './transcription/google-transcription-model-options';
 export { createGoogleGenerativeAI, google } from './google-provider';
 export type {
   GoogleGenerativeAIProvider,

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

@@ -0,0 +1,51 @@
+import { z } from 'zod/v4';
+
+export type GoogleTranscriptionModelId =
+  | 'gemini-3.5-transcribe'
+  | 'gemini-3.5-transcribe-live'
+  | (string & {});
+
+/**
+ * Speech recognition options for Gemini transcription models
+ * (`gemini-3.5-transcribe`). Maps onto Google's `AudioTranscriptionConfig`.
+ * The live variant (`gemini-3.5-transcribe-live`) requires streaming
+ * transcription, which is only available in AI SDK v7.
+ */
+export const googleTranscriptionModelOptions = z.object({
+  /**
+   * BCP-47 language codes providing hints about the languages present in the
+   * audio. If omitted or empty, defaults to automatic language detection.
+   */
+  languageCodes: z.array(z.string()).optional(),
+
+  /**
+   * Custom vocabulary phrases, which bias the speech recognition model
+   * toward recognizing specific terms.
+   */
+  customVocabulary: z.array(z.string()).optional(),
+
+  /**
+   * Enables word-level timestamp generation.
+   */
+  wordTimestamp: z.boolean().optional(),
+
+  /**
+   * Enables speaker diarization.
+   */
+  diarization: z.boolean().optional(),
+
+  /**
+   * Transcription output formatting mode.
+   *
+   * - `VERBATIM` (default): exact literal transcript preserving filler
+   *   words, repetitions, and false starts.
+   * - `SMART`: cleans up and structures the transcript in real time —
+   *   disfluency removal, inline self-corrections, structured formatting
+   *   (lists, numbers, dates, paragraph breaks), and grammar/casing polish.
+   */
+  mode: z.enum(['SMART', 'VERBATIM']).optional(),
+});
+
+export type GoogleTranscriptionModelOptions = z.infer<
+  typeof googleTranscriptionModelOptions
+>;