@ai-sdk/google 4.0.0-beta.13 → 4.0.0-beta.15

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

@@ -2,22 +2,179 @@ import {
   LanguageModelV4Prompt,
   UnsupportedFunctionalityError,
 } from '@ai-sdk/provider';
+import { convertToBase64 } from '@ai-sdk/provider-utils';
 import {
   GoogleGenerativeAIContent,
   GoogleGenerativeAIContentPart,
+  GoogleGenerativeAIFunctionResponsePart,
   GoogleGenerativeAIPrompt,
 } from './google-generative-ai-prompt';
-import { convertToBase64 } from '@ai-sdk/provider-utils';
+
+const dataUrlRegex = /^data:([^;,]+);base64,(.+)$/s;
+
+function parseBase64DataUrl(
+  value: string,
+): { mediaType: string; data: string } | undefined {
+  const match = dataUrlRegex.exec(value);
+  if (match == null) {
+    return undefined;
+  }
+
+  return {
+    mediaType: match[1],
+    data: match[2],
+  };
+}
+
+function convertUrlToolResultPart(
+  url: string,
+): GoogleGenerativeAIFunctionResponsePart | 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;
+  }
+
+  return {
+    inlineData: {
+      mimeType: parsedDataUrl.mediaType,
+      data: parsedDataUrl.data,
+    },
+  };
+}
+
+/*
+ * Appends tool result content parts to the message using the functionResponse
+ * format with support for multimodal parts (e.g. inline images/files alongside
+ * text). This format is supported by Gemini 3+ models.
+ */
+function appendToolResultParts(
+  parts: GoogleGenerativeAIContentPart[],
+  toolName: string,
+  outputValue: Array<{
+    type: string;
+    [key: string]: unknown;
+  }>,
+): void {
+  const functionResponseParts: GoogleGenerativeAIFunctionResponsePart[] = [];
+  const responseTextParts: string[] = [];
+
+  for (const contentPart of outputValue) {
+    switch (contentPart.type) {
+      case 'text': {
+        responseTextParts.push(contentPart.text as string);
+        break;
+      }
+      case 'image-data':
+      case 'file-data': {
+        functionResponseParts.push({
+          inlineData: {
+            mimeType: contentPart.mediaType as string,
+            data: contentPart.data as string,
+          },
+        });
+        break;
+      }
+      case 'image-url':
+      case 'file-url': {
+        const functionResponsePart = convertUrlToolResultPart(
+          contentPart.url as string,
+        );
+
+        if (functionResponsePart != null) {
+          functionResponseParts.push(functionResponsePart);
+        } else {
+          responseTextParts.push(JSON.stringify(contentPart));
+        }
+        break;
+      }
+      default: {
+        responseTextParts.push(JSON.stringify(contentPart));
+        break;
+      }
+    }
+  }
+
+  parts.push({
+    functionResponse: {
+      name: toolName,
+      response: {
+        name: toolName,
+        content:
+          responseTextParts.length > 0
+            ? responseTextParts.join('\n')
+            : 'Tool executed successfully.',
+      },
+      ...(functionResponseParts.length > 0
+        ? { parts: functionResponseParts }
+        : {}),
+    },
+  });
+}
+
+/*
+ * Appends tool result content parts using a legacy format for pre-Gemini 3
+ * models that do not support multimodal parts within functionResponse. Instead,
+ * non-text content like images is sent as separate top-level inlineData parts.
+ */
+function appendLegacyToolResultParts(
+  parts: GoogleGenerativeAIContentPart[],
+  toolName: string,
+  outputValue: Array<{
+    type: string;
+    [key: string]: unknown;
+  }>,
+): void {
+  for (const contentPart of outputValue) {
+    switch (contentPart.type) {
+      case 'text':
+        parts.push({
+          functionResponse: {
+            name: toolName,
+            response: {
+              name: toolName,
+              content: contentPart.text,
+            },
+          },
+        });
+        break;
+      case 'image-data':
+        parts.push(
+          {
+            inlineData: {
+              mimeType: String(contentPart.mediaType),
+              data: String(contentPart.data),
+            },
+          },
+          {
+            text: 'Tool executed successfully and returned this image as a response',
+          },
+        );
+        break;
+      default:
+        parts.push({ text: JSON.stringify(contentPart) });
+        break;
+    }
+  }
+}
 
 export function convertToGoogleGenerativeAIMessages(
   prompt: LanguageModelV4Prompt,
-  options?: { isGemmaModel?: boolean; providerOptionsName?: string },
+  options?: {
+    isGemmaModel?: boolean;
+    providerOptionsName?: string;
+    supportsFunctionResponseParts?: boolean;
+  },
 ): GoogleGenerativeAIPrompt {
   const systemInstructionParts: Array<{ text: string }> = [];
   const contents: Array<GoogleGenerativeAIContent> = [];
   let systemMessagesAllowed = true;
   const isGemmaModel = options?.isGemmaModel ?? false;
   const providerOptionsName = options?.providerOptionsName ?? 'google';
+  const supportsFunctionResponseParts =
+    options?.supportsFunctionResponseParts ?? true;
 
   for (const { role, content } of prompt) {
     switch (role) {
@@ -178,36 +335,10 @@ export function convertToGoogleGenerativeAIMessages(
           const output = part.output;
 
           if (output.type === 'content') {
-            for (const contentPart of output.value) {
-              switch (contentPart.type) {
-                case 'text':
-                  parts.push({
-                    functionResponse: {
-                      name: part.toolName,
-                      response: {
-                        name: part.toolName,
-                        content: contentPart.text,
-                      },
-                    },
-                  });
-                  break;
-                case 'image-data':
-                  parts.push(
-                    {
-                      inlineData: {
-                        mimeType: contentPart.mediaType,
-                        data: contentPart.data,
-                      },
-                    },
-                    {
-                      text: 'Tool executed successfully and returned this image as a response',
-                    },
-                  );
-                  break;
-                default:
-                  parts.push({ text: JSON.stringify(contentPart) });
-                  break;
-              }
+            if (supportsFunctionResponseParts) {
+              appendToolResultParts(parts, part.toolName, output.value);
+            } else {
+              appendLegacyToolResultParts(parts, part.toolName, output.value);
             }
           } else {
             parts.push({

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

@@ -17,7 +17,10 @@ import {
   FetchFunction,
   generateId,
   InferSchema,
+  isCustomReasoning,
   lazySchema,
+  mapReasoningToProviderBudget,
+  mapReasoningToProviderEffort,
   parseProviderOptions,
   ParseResult,
   postJsonToApi,
@@ -96,6 +99,7 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV4 {
     seed,
     tools,
     toolChoice,
+    reasoning,
     providerOptions,
   }: LanguageModelV4CallOptions) {
     const warnings: SharedV4Warning[] = [];
@@ -135,10 +139,15 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV4 {
     }
 
     const isGemmaModel = this.modelId.toLowerCase().startsWith('gemma-');
+    const supportsFunctionResponseParts = this.modelId.startsWith('gemini-3');
 
     const { contents, systemInstruction } = convertToGoogleGenerativeAIMessages(
       prompt,
-      { isGemmaModel, providerOptionsName },
+      {
+        isGemmaModel,
+        providerOptionsName,
+        supportsFunctionResponseParts,
+      },
     );
 
     const {
@@ -151,6 +160,16 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV4 {
       modelId: this.modelId,
     });
 
+    const resolvedThinking = resolveThinkingConfig({
+      reasoning,
+      modelId: this.modelId,
+      warnings,
+    });
+    const thinkingConfig =
+      googleOptions?.thinkingConfig || resolvedThinking
+        ? { ...resolvedThinking, ...googleOptions?.thinkingConfig }
+        : undefined;
+
     return {
       args: {
         generationConfig: {
@@ -182,7 +201,7 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV4 {
 
           // provider options:
           responseModalities: googleOptions?.responseModalities,
-          thinkingConfig: googleOptions?.thinkingConfig,
+          thinkingConfig,
           ...(googleOptions?.mediaResolution && {
             mediaResolution: googleOptions.mediaResolution,
           }),
@@ -706,6 +725,109 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV4 {
   }
 }
 
+function isGemini3Model(modelId: string): boolean {
+  return /gemini-3[\.\-]/i.test(modelId) || /gemini-3$/i.test(modelId);
+}
+
+function getMaxOutputTokensForGemini25Model(): number {
+  return 65536;
+}
+
+function getMaxThinkingTokensForGemini25Model(modelId: string): number {
+  const id = modelId.toLowerCase();
+  if (id.includes('2.5-pro') || id.includes('gemini-3-pro-image')) {
+    return 32768;
+  }
+  return 24576;
+}
+
+type GoogleThinkingConfig = NonNullable<
+  InferSchema<typeof googleLanguageModelOptions>['thinkingConfig']
+>;
+
+function resolveThinkingConfig({
+  reasoning,
+  modelId,
+  warnings,
+}: {
+  reasoning: LanguageModelV4CallOptions['reasoning'];
+  modelId: string;
+  warnings: SharedV4Warning[];
+}): Omit<GoogleThinkingConfig, 'includeThoughts'> | undefined {
+  if (!isCustomReasoning(reasoning)) {
+    return undefined;
+  }
+
+  if (isGemini3Model(modelId) && !modelId.includes('gemini-3-pro-image')) {
+    return resolveGemini3ThinkingConfig({ reasoning, warnings });
+  }
+
+  return resolveGemini25ThinkingConfig({ reasoning, modelId, warnings });
+}
+
+function resolveGemini3ThinkingConfig({
+  reasoning,
+  warnings,
+}: {
+  reasoning: Exclude<
+    LanguageModelV4CallOptions['reasoning'],
+    'provider-default' | undefined
+  >;
+  warnings: SharedV4Warning[];
+}): Pick<GoogleThinkingConfig, 'thinkingLevel'> | undefined {
+  if (reasoning === 'none') {
+    // It's not possible to fully disable thinking with Gemini 3.
+    return { thinkingLevel: 'minimal' };
+  }
+
+  const thinkingLevel = mapReasoningToProviderEffort({
+    reasoning,
+    effortMap: {
+      minimal: 'minimal',
+      low: 'low',
+      medium: 'medium',
+      high: 'high',
+      xhigh: 'high',
+    },
+    warnings,
+  });
+
+  if (thinkingLevel == null) {
+    return undefined;
+  }
+
+  return { thinkingLevel };
+}
+
+function resolveGemini25ThinkingConfig({
+  reasoning,
+  modelId,
+  warnings,
+}: {
+  reasoning: Exclude<
+    LanguageModelV4CallOptions['reasoning'],
+    'provider-default' | undefined
+  >;
+  modelId: string;
+  warnings: SharedV4Warning[];
+}): Pick<GoogleThinkingConfig, 'thinkingBudget'> | undefined {
+  if (reasoning === 'none') {
+    return { thinkingBudget: 0 };
+  }
+
+  const thinkingBudget = mapReasoningToProviderBudget({
+    reasoning,
+    maxOutputTokens: getMaxOutputTokensForGemini25Model(),
+    maxReasoningBudget: getMaxThinkingTokensForGemini25Model(modelId),
+    minReasoningBudget: 0,
+    warnings,
+  });
+  if (thinkingBudget == null) {
+    return undefined;
+  }
+  return { thinkingBudget };
+}
+
 function getToolCallsFromParts({
   parts,
   generateId,

package/src/google-generative-ai-prompt.ts

@@ -2,9 +2,9 @@ import {
   GroundingMetadataSchema,
   PromptFeedbackSchema,
   UrlContextMetadataSchema,
+  type SafetyRatingSchema,
   UsageMetadataSchema,
 } from './google-generative-ai-language-model';
-import { type SafetyRatingSchema } from './google-generative-ai-language-model';
 
 export type GoogleGenerativeAIPrompt = {
   systemInstruction?: GoogleGenerativeAISystemInstruction;
@@ -24,9 +24,19 @@ export type GoogleGenerativeAIContentPart =
   | { text: string; thought?: boolean; thoughtSignature?: string }
   | { inlineData: { mimeType: string; data: string } }
   | { functionCall: { name: string; args: unknown }; thoughtSignature?: string }
-  | { functionResponse: { name: string; response: unknown } }
+  | {
+      functionResponse: {
+        name: string;
+        response: unknown;
+        parts?: Array<GoogleGenerativeAIFunctionResponsePart>;
+      };
+    }
   | { fileData: { mimeType: string; fileUri: string } };
 
+export type GoogleGenerativeAIFunctionResponsePart = {
+  inlineData: { mimeType: string; data: string };
+};
+
 export type GoogleGenerativeAIGroundingMetadata = GroundingMetadataSchema;
 
 export type GoogleGenerativeAIUrlContextMetadata = UrlContextMetadataSchema;