@ai-sdk/google 3.0.50 → 3.0.52

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

@@ -2,22 +2,179 @@ import {
   LanguageModelV3Prompt,
   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: LanguageModelV3Prompt,
-  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) {
@@ -160,36 +317,10 @@ 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

@@ -38,7 +38,10 @@ import {
   GoogleGenerativeAIModelId,
   googleLanguageModelOptions,
 } from './google-generative-ai-options';
-import { GoogleGenerativeAIContentPart } from './google-generative-ai-prompt';
+import {
+  GoogleGenerativeAIContentPart,
+  GoogleGenerativeAIProviderMetadata,
+} from './google-generative-ai-prompt';
 import { prepareTools } from './google-prepare-tools';
 import { mapGoogleGenerativeAIFinishReason } from './map-google-generative-ai-finish-reason';
 
@@ -132,10 +135,15 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
     }
 
     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 {
@@ -355,7 +363,8 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
           urlContextMetadata: candidate.urlContextMetadata ?? null,
           safetyRatings: candidate.safetyRatings ?? null,
           usageMetadata: usageMetadata ?? null,
-        },
+          finishMessage: candidate.finishMessage ?? null,
+        } satisfies GoogleGenerativeAIProviderMetadata,
       },
       request: { body: args },
       response: {
@@ -674,16 +683,10 @@ export class GoogleGenerativeAILanguageModel implements LanguageModelV3 {
                   groundingMetadata: lastGroundingMetadata,
                   urlContextMetadata: lastUrlContextMetadata,
                   safetyRatings: candidate.safetyRatings ?? null,
-                },
+                  usageMetadata: usageMetadata ?? null,
+                  finishMessage: candidate.finishMessage ?? null,
+                } satisfies GoogleGenerativeAIProviderMetadata,
               };
-              if (usageMetadata != null) {
-                (
-                  providerMetadata[providerOptionsName] as Record<
-                    string,
-                    unknown
-                  >
-                ).usageMetadata = usageMetadata;
-              }
             }
           },
 
@@ -1012,6 +1015,7 @@ const responseSchema = lazySchema(() =>
         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(),
@@ -1047,6 +1051,14 @@ export type SafetyRatingSchema = NonNullable<
   InferSchema<typeof responseSchema>['candidates'][number]['safetyRatings']
 >[number];
 
+export type PromptFeedbackSchema = NonNullable<
+  InferSchema<typeof responseSchema>['promptFeedback']
+>;
+
+export type UsageMetadataSchema = NonNullable<
+  InferSchema<typeof responseSchema>['usageMetadata']
+>;
+
 // limited version of the schema, focussed on what is needed for the implementation
 // this approach limits breakages when the API changes and increases efficiency
 const chunkSchema = lazySchema(() =>
@@ -1057,6 +1069,7 @@ const chunkSchema = lazySchema(() =>
           z.object({
             content: getContentSchema().nullish(),
             finishReason: z.string().nullish(),
+            finishMessage: z.string().nullish(),
             safetyRatings: z.array(getSafetyRatingSchema()).nullish(),
             groundingMetadata: getGroundingMetadataSchema().nullish(),
             urlContextMetadata: getUrlContextMetadataSchema().nullish(),

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

@@ -1,8 +1,10 @@
 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;
@@ -22,17 +24,34 @@ 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;
 
 export type GoogleGenerativeAISafetyRating = SafetyRatingSchema;
 
+export type GoogleGenerativeAIPromptFeedback = PromptFeedbackSchema;
+
+export type GoogleGenerativeAIUsageMetadata = UsageMetadataSchema;
+
 export interface GoogleGenerativeAIProviderMetadata {
+  promptFeedback: GoogleGenerativeAIPromptFeedback | null;
   groundingMetadata: GoogleGenerativeAIGroundingMetadata | null;
   urlContextMetadata: GoogleGenerativeAIUrlContextMetadata | null;
   safetyRatings: GoogleGenerativeAISafetyRating[] | null;
+  usageMetadata: GoogleGenerativeAIUsageMetadata | null;
+  finishMessage: string | null;
 }