From 3ff16bc3daeae2849dd7a2686b3f51c199a73737 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 21 Apr 2026 15:26:05 +0100
Subject: [PATCH 01/16] feat(genkit_vertexai): support Gemini and multimodal
 embedders

---
 .../genkit_vertexai/lib/src/embedders.dart    | 411 ++++++++++++++++++
 .../lib/src/vertex_api_client.dart            |  77 +---
 .../genkit_vertexai/test/embedders_test.dart  | 142 ++++++
 .../test/test_http_client.dart                | 101 +++++
 .../genkit_vertexai/test/vertex_test.dart     |  34 +-
 5 files changed, 666 insertions(+), 99 deletions(-)
 create mode 100644 packages/genkit_vertexai/lib/src/embedders.dart
 create mode 100644 packages/genkit_vertexai/test/embedders_test.dart
 create mode 100644 packages/genkit_vertexai/test/test_http_client.dart

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
new file mode 100644
index 00000000..dd42af75
--- /dev/null
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -0,0 +1,411 @@
+// Copyright 2025 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+import 'dart:convert';
+
+import 'package:genkit/plugin.dart';
+import 'package:genkit_google_genai/common.dart';
+import 'package:genkit_google_genai/src/generated/generativelanguage.dart'
+    as gcl;
+
+List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
+  required String pluginName,
+  required List<dynamic> publisherModels,
+}) {
+  return publisherModels
+      .where((m) {
+        final modelMap = m as Map<String, dynamic>;
+        final name = modelMap['name'] as String?;
+        return name != null && name.contains('embedding');
+      })
+      .map((m) {
+        final modelMap = m as Map<String, dynamic>;
+        final modelName = (modelMap['name'] as String).split('/').last;
+        return embedderMetadata(
+          '$pluginName/$modelName',
+          customOptions: TextEmbedderOptions.$schema,
+        );
+      })
+      .toList();
+}
+
+Embedder createVertexEmbedder({
+  required String pluginName,
+  required String embedderName,
+  required Future<GenerativeLanguageBaseClient> Function() getApiClient,
+  required GenkitException Function(Object, StackTrace) handleException,
+  required bool closeService,
+}) {
+  return Embedder(
+    name: '$pluginName/$embedderName',
+    fn: (req, ctx) async {
+      if (req == null || req.input.isEmpty) {
+        return EmbedResponse(embeddings: []);
+      }
+
+      final service = await getApiClient();
+      try {
+        final options = req.options != null
+            ? TextEmbedderOptions.fromJson(req.options!)
+            : null;
+
+        // Handle each input document separately, because Vertex embedders do
+        // not all use the same request shape.
+        final futures = req.input.map(
+          (doc) => _runEmbedderRequest(
+            service: service,
+            embedderName: embedderName,
+            doc: doc,
+            options: options,
+          ),
+        );
+        return EmbedResponse(embeddings: await Future.wait(futures));
+      } catch (e, stack) {
+        throw handleException(e, stack);
+      } finally {
+        if (closeService) {
+          service.client.close();
+        }
+      }
+    },
+  );
+}
+
+String _documentText(DocumentData doc) {
+  return doc.content.where((p) => p.isText).map((p) => p.text).join('\n');
+}
+
+Future<Embedding> _runEmbedderRequest({
+  required GenerativeLanguageBaseClient service,
+  required String embedderName,
+  required DocumentData doc,
+  required TextEmbedderOptions? options,
+}) async {
+  // Choose the request style from the model family.
+  return switch (_requestShapeFor(embedderName)) {
+    _VertexEmbedderRequestShape.geminiEmbedding => _runGeminiEmbeddingRequest(
+      service: service,
+      embedderName: embedderName,
+      doc: doc,
+      options: options,
+    ),
+    _VertexEmbedderRequestShape.multimodalPredict =>
+      _runMultimodalPredictRequest(
+        service: service,
+        embedderName: embedderName,
+        doc: doc,
+        options: options,
+      ),
+    _VertexEmbedderRequestShape.textPredict => _runTextPredictRequest(
+      service: service,
+      embedderName: embedderName,
+      doc: doc,
+      options: options,
+    ),
+  };
+}
+
+Future<Embedding> _runGeminiEmbeddingRequest({
+  required GenerativeLanguageBaseClient service,
+  required String embedderName,
+  required DocumentData doc,
+  required TextEmbedderOptions? options,
+}) async {
+  try {
+    // Try the newer Gemini embedding API first.
+    return await _runEmbedContentRequest(
+      service: service,
+      embedderName: embedderName,
+      doc: doc,
+      options: options,
+    );
+  } on GenkitException catch (e) {
+    if (!_shouldFallbackToTextPredict(e)) {
+      rethrow;
+    }
+  }
+
+  // Some older Gemini embedding models still need the predict API.
+  return _runTextPredictRequest(
+    service: service,
+    embedderName: embedderName,
+    doc: doc,
+    options: options,
+  );
+}
+
+Future<Embedding> _runEmbedContentRequest({
+  required GenerativeLanguageBaseClient service,
+  required String embedderName,
+  required DocumentData doc,
+  required TextEmbedderOptions? options,
+}) async {
+  final text = _documentText(doc);
+  final content = gcl.Content(parts: [gcl.Part(text: text)]);
+  final res = await service.embedContent(
+    gcl.EmbedContentRequest(
+      content: content,
+      outputDimensionality: options?.outputDimensionality,
+      taskType: options?.taskType,
+      title: options?.title,
+    ),
+    model: 'models/$embedderName',
+  );
+  return Embedding(embedding: res.embedding?.values ?? []);
+}
+
+Future<Embedding> _runMultimodalPredictRequest({
+  required GenerativeLanguageBaseClient service,
+  required String embedderName,
+  required DocumentData doc,
+  required TextEmbedderOptions? options,
+}) async {
+  // Multimodal embedders use a different predict request and response shape.
+  final instance = _toMultimodalInstance(doc);
+  final parameters = <String, dynamic>{};
+  if (options?.outputDimensionality != null) {
+    parameters['dimension'] = options!.outputDimensionality;
+  }
+
+  final res = await service.predict({
+    'instances': [instance.instance],
+    if (parameters.isNotEmpty) 'parameters': parameters,
+  }, model: 'models/$embedderName');
+
+  final predictions = res['predictions'] as List;
+  final prediction = predictions.single as Map<String, dynamic>;
+  return Embedding(
+    embedding: _multimodalPredictionEmbedding(
+      prediction,
+      expectedOutput: instance.expectedOutput,
+    ),
+  );
+}
+
+Future<Embedding> _runTextPredictRequest({
+  required GenerativeLanguageBaseClient service,
+  required String embedderName,
+  required DocumentData doc,
+  required TextEmbedderOptions? options,
+}) async {
+  // Older text embedders still use the predict payload shape.
+  final instance = <String, dynamic>{'content': _documentText(doc)};
+  if (options?.title != null) {
+    instance['title'] = options!.title;
+  }
+
+  final parameters = <String, dynamic>{};
+  if (options?.outputDimensionality != null) {
+    parameters['outputDimensionality'] = options!.outputDimensionality;
+  }
+  if (options?.taskType != null) {
+    parameters['taskType'] = options!.taskType;
+  }
+
+  final res = await service.predict({
+    'instances': [instance],
+    if (parameters.isNotEmpty) 'parameters': parameters,
+  }, model: 'models/$embedderName');
+
+  final predictions = res['predictions'] as List;
+  final prediction = predictions.single as Map<String, dynamic>;
+  final embeddingData = prediction['embeddings'] as Map<String, dynamic>;
+  final values = embeddingData['values'] as List;
+  return Embedding(
+    embedding: values.map((value) => (value as num).toDouble()).toList(),
+  );
+}
+
+_MultimodalInstance _toMultimodalInstance(DocumentData doc) {
+  final text = _documentText(doc).trim();
+  final mediaParts = doc.content
+      .where((part) => part.isMedia)
+      .map((part) => part.media!)
+      .toList();
+
+  // A document can only contain one input type here.
+  // Text, image, and video use different embedding fields in the Vertex
+  // response, and this code needs one clear field to read for each document.
+  if (text.isNotEmpty && mediaParts.isNotEmpty) {
+    throw GenkitException(
+      'Vertex multimodalembedding supports exactly one modality per input document in the embedder API. Provide text, one image, or one video.',
+      status: StatusCodes.INVALID_ARGUMENT,
+    );
+  }
+
+  if (mediaParts.length > 1) {
+    throw GenkitException(
+      'Vertex multimodalembedding supports at most one media part per input document in the embedder API.',
+      status: StatusCodes.INVALID_ARGUMENT,
+    );
+  }
+
+  if (text.isNotEmpty) {
+    return _MultimodalInstance(
+      instance: {'text': text},
+      expectedOutput: _MultimodalOutput.text,
+    );
+  }
+
+  if (mediaParts.isEmpty) {
+    throw GenkitException(
+      'Vertex multimodalembedding requires text, image, or video input.',
+      status: StatusCodes.INVALID_ARGUMENT,
+    );
+  }
+
+  final mediaField = _toMultimodalMediaField(mediaParts.single);
+  return _MultimodalInstance(
+    instance: {mediaField.key: mediaField.value},
+    expectedOutput: mediaField.key == 'image'
+        ? _MultimodalOutput.image
+        : _MultimodalOutput.video,
+  );
+}
+
+MapEntry<String, Map<String, dynamic>> _toMultimodalMediaField(Media media) {
+  final mimeType = _mediaMimeType(media);
+  final fieldName = _multimodalFieldName(mimeType);
+
+  // Convert the media input into the format Vertex expects.
+  if (media.url.startsWith('data:')) {
+    final data = Uri.parse(media.url).data;
+    if (data == null) {
+      throw GenkitException(
+        'Vertex multimodalembedding media inputs require a valid data URI.',
+        status: StatusCodes.INVALID_ARGUMENT,
+      );
+    }
+
+    return MapEntry(fieldName, {
+      'bytesBase64Encoded': base64Encode(data.contentAsBytes()),
+      if (mimeType != null && mimeType.isNotEmpty) 'mimeType': mimeType,
+    });
+  }
+
+  if (media.url.startsWith('gs://')) {
+    return MapEntry(fieldName, {
+      'gcsUri': media.url,
+      if (mimeType != null && mimeType.isNotEmpty) 'mimeType': mimeType,
+    });
+  }
+
+  throw GenkitException(
+    'Vertex multimodalembedding media inputs must use gs:// URIs or inline data URIs.',
+    status: StatusCodes.INVALID_ARGUMENT,
+  );
+}
+
+String? _mediaMimeType(Media media) {
+  if (media.contentType?.isNotEmpty == true) {
+    return media.contentType;
+  }
+
+  if (media.url.startsWith('data:')) {
+    return Uri.parse(media.url).data?.mimeType;
+  }
+
+  return null;
+}
+
+String _multimodalFieldName(String? mimeType) {
+  if (mimeType == null || mimeType.isEmpty) {
+    throw GenkitException(
+      'Vertex multimodalembedding media inputs require a MIME type.',
+      status: StatusCodes.INVALID_ARGUMENT,
+    );
+  }
+
+  if (mimeType.startsWith('image/')) {
+    return 'image';
+  }
+  if (mimeType.startsWith('video/')) {
+    return 'video';
+  }
+
+  throw GenkitException(
+    'Unsupported Vertex multimodalembedding media MIME type: $mimeType',
+    status: StatusCodes.INVALID_ARGUMENT,
+  );
+}
+
+List<double> _multimodalPredictionEmbedding(
+  Map<String, dynamic> prediction, {
+  required _MultimodalOutput expectedOutput,
+}) {
+  // Read the embedding field that matches the input type.
+  final values = switch (expectedOutput) {
+    _MultimodalOutput.text => prediction['textEmbedding'] as List?,
+    _MultimodalOutput.image => prediction['imageEmbedding'] as List?,
+    _MultimodalOutput.video =>
+      ((prediction['videoEmbeddings'] as List?)?.firstOrNull
+              as Map<String, dynamic>?)?['embedding']
+          as List?,
+  };
+
+  if (values == null) {
+    throw GenkitException(
+      'Vertex multimodalembedding did not return a ${expectedOutput.name} embedding.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+
+  return values.map((value) => (value as num).toDouble()).toList();
+}
+
+String _baseModelName(String modelName) {
+  final atIndex = modelName.indexOf('@');
+  if (atIndex == -1) return modelName;
+  return modelName.substring(0, atIndex);
+}
+
+_VertexEmbedderRequestShape _requestShapeFor(String modelName) {
+  final baseModelName = _baseModelName(modelName);
+  // Check the broad model families in order.
+  if (_isMultimodalEmbeddingFamily(baseModelName)) {
+    return _VertexEmbedderRequestShape.multimodalPredict;
+  }
+  if (_isGeminiEmbeddingFamily(baseModelName)) {
+    return _VertexEmbedderRequestShape.geminiEmbedding;
+  }
+  return _VertexEmbedderRequestShape.textPredict;
+}
+
+bool _isMultimodalEmbeddingFamily(String modelName) {
+  return modelName.contains('multimodal') && modelName.contains('embedding');
+}
+
+bool _isGeminiEmbeddingFamily(String modelName) {
+  return modelName.startsWith('gemini-embedding-');
+}
+
+bool _shouldFallbackToTextPredict(GenkitException error) {
+  return error.status == StatusCodes.INVALID_ARGUMENT &&
+      error.message.contains('not supported in the embedContent API');
+}
+
+class _MultimodalInstance {
+  final Map<String, dynamic> instance;
+  final _MultimodalOutput expectedOutput;
+
+  _MultimodalInstance({required this.instance, required this.expectedOutput});
+}
+
+enum _MultimodalOutput { text, image, video }
+
+enum _VertexEmbedderRequestShape {
+  geminiEmbedding,
+  multimodalPredict,
+  textPredict,
+}
diff --git a/packages/genkit_vertexai/lib/src/vertex_api_client.dart b/packages/genkit_vertexai/lib/src/vertex_api_client.dart
index dd0a5cf7..8c6318b3 100644
--- a/packages/genkit_vertexai/lib/src/vertex_api_client.dart
+++ b/packages/genkit_vertexai/lib/src/vertex_api_client.dart
@@ -19,6 +19,7 @@ import 'package:http/http.dart' as http;
 import 'package:meta/meta.dart';
 
 import 'auth.dart';
+import 'embedders.dart';
 
 @visibleForTesting
 class VertexAiPluginImpl extends CommonGoogleGenPlugin {
@@ -107,20 +108,10 @@ class VertexAiPluginImpl extends CommonGoogleGenPlugin {
           })
           .toList();
 
-      final embedders = publisherModels
-          .where((m) {
-            final modelMap = m as Map<String, dynamic>;
-            final name = modelMap['name'] as String?;
-            return name != null &&
-                (name.contains('text-embedding-') ||
-                    name.contains('embedding-'));
-          })
-          .map((m) {
-            final modelMap = m as Map<String, dynamic>;
-            final modelName = (modelMap['name'] as String).split('/').last;
-            return embedderMetadata('$name/$modelName');
-          })
-          .toList();
+      final embedders = listVertexEmbedders(
+        pluginName: name,
+        publisherModels: publisherModels,
+      );
 
       return [...models, ...embedders];
     } catch (e, stack) {
@@ -136,58 +127,12 @@ class VertexAiPluginImpl extends CommonGoogleGenPlugin {
 
   @override
   Embedder createEmbedder(String embedderName) {
-    return Embedder(
-      name: '$name/$embedderName',
-      fn: (req, ctx) async {
-        if (req == null || req.input.isEmpty) {
-          return EmbedResponse(embeddings: []);
-        }
-        final service = await getApiClient();
-        try {
-          final options = req.options != null
-              ? TextEmbedderOptions.fromJson(req.options!)
-              : null;
-
-          final instances = req.input.map((doc) {
-            final text = doc.content
-                .where((p) => p.isText)
-                .map((p) => p.text)
-                .join('\n');
-            return {'content': text};
-          }).toList();
-
-          final parameters = <String, dynamic>{};
-          if (options?.outputDimensionality != null) {
-            parameters['outputDimensionality'] = options!.outputDimensionality;
-          }
-          if (options?.taskType != null) {
-            parameters['taskType'] = options!.taskType;
-          }
-
-          final res = await service.predict({
-            'instances': instances,
-            if (parameters.isNotEmpty) 'parameters': parameters,
-          }, model: 'models/$embedderName');
-
-          final predictions = res['predictions'] as List;
-          final embeddings = predictions.map((p) {
-            final emb =
-                (p as Map<String, dynamic>)['embeddings']
-                    as Map<String, dynamic>;
-            final vals = emb['values'] as List;
-            return Embedding(
-              embedding: vals.map((e) => (e as num).toDouble()).toList(),
-            );
-          }).toList();
-          return EmbedResponse(embeddings: embeddings);
-        } catch (e, stack) {
-          throw handleException(e, stack);
-        } finally {
-          if (authClient == null) {
-            service.client.close();
-          }
-        }
-      },
+    return createVertexEmbedder(
+      pluginName: name,
+      embedderName: embedderName,
+      getApiClient: () => getApiClient(),
+      handleException: handleException,
+      closeService: authClient == null,
     );
   }
 }
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
new file mode 100644
index 00000000..84803418
--- /dev/null
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -0,0 +1,142 @@
+// Copyright 2025 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+import 'dart:convert';
+
+import 'package:genkit/genkit.dart';
+import 'package:genkit_vertexai/src/vertex_api_client.dart';
+import 'package:test/test.dart';
+
+import 'test_http_client.dart';
+
+void main() {
+  group('Vertex AI Embedders', () {
+    test('lists embedders with schemas and custom options', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final actions = await plugin.list();
+      final embedder = actions.firstWhere(
+        (action) => action.name == 'vertexai/text-embedding-005',
+      );
+
+      expect(embedder.name, 'vertexai/text-embedding-005');
+      expect(embedder.inputSchema, same(EmbedRequest.$schema));
+      expect(embedder.outputSchema, same(EmbedResponse.$schema));
+
+      final customOptions =
+          embedder.metadata['model']['customOptions'] as Map<String, dynamic>;
+      expect(customOptions['properties'], contains('outputDimensionality'));
+      expect(customOptions['properties'], contains('taskType'));
+    });
+
+    test('uses embedContent for embedders', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder =
+          plugin.resolve('embedder', 'gemini-embedding-2-preview') as Action;
+      final req = EmbedRequest(
+        input: [
+          DocumentData(content: [TextPart(text: 'hello')]),
+        ],
+      );
+
+      final response = await embedder.run(req);
+
+      expect(mockClient.lastUrl, isNotNull);
+      expect(
+        mockClient.lastUrl.toString(),
+        'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-2-preview:embedContent',
+      );
+      expect(response.result.embeddings, hasLength(1));
+      expect(response.result.embeddings.first.embedding, [0.1, 0.2, 0.3]);
+    });
+
+    test('uses predict for gemini-embedding-001', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder =
+          plugin.resolve('embedder', 'gemini-embedding-001') as Action;
+      final req = EmbedRequest(
+        input: [
+          DocumentData(content: [TextPart(text: 'hello')]),
+        ],
+      );
+
+      final response = await embedder.run(req);
+
+      expect(mockClient.lastUrl, isNotNull);
+      expect(
+        mockClient.lastUrl.toString(),
+        'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-001:predict',
+      );
+      expect(response.result.embeddings, hasLength(1));
+      expect(response.result.embeddings.first.embedding, [0.4, 0.5, 0.6]);
+    });
+
+    test(
+      'uses multimodal predict schema for multimodalembedding text input',
+      () async {
+        final mockClient = MockHttpClient();
+        final plugin = VertexAiPluginImpl(
+          projectId: 'my-project',
+          location: 'us-central1',
+          authClient: mockClient,
+        );
+
+        final embedder =
+            plugin.resolve('embedder', 'multimodalembedding') as Action;
+        final req = EmbedRequest(
+          input: [
+            DocumentData(content: [TextPart(text: 'hello')]),
+          ],
+          options: {
+            'outputDimensionality': 256,
+            'taskType': 'RETRIEVAL_DOCUMENT',
+          },
+        );
+
+        final response = await embedder.run(req);
+
+        expect(mockClient.lastUrl, isNotNull);
+        expect(
+          mockClient.lastUrl.toString(),
+          'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/multimodalembedding:predict',
+        );
+
+        final requestBody =
+            jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+        final instances = requestBody['instances'] as List;
+        expect(instances.single, {'text': 'hello'});
+        expect(requestBody['parameters'], {'dimension': 256});
+        expect(response.result.embeddings, hasLength(1));
+        expect(response.result.embeddings.first.embedding, [0.7, 0.8, 0.9]);
+      },
+    );
+  });
+}
diff --git a/packages/genkit_vertexai/test/test_http_client.dart b/packages/genkit_vertexai/test/test_http_client.dart
new file mode 100644
index 00000000..32eba6fe
--- /dev/null
+++ b/packages/genkit_vertexai/test/test_http_client.dart
@@ -0,0 +1,101 @@
+// Copyright 2025 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+import 'dart:convert';
+
+import 'package:http/http.dart' as http;
+
+class MockHttpClient extends http.BaseClient {
+  Uri? lastUrl;
+  String? lastBody;
+
+  @override
+  Future<http.StreamedResponse> send(http.BaseRequest request) async {
+    lastUrl = request.url;
+    if (request is http.Request) {
+      lastBody = request.body;
+    }
+    if (request.url.host == 'metadata.google.internal' ||
+        request.url.host == 'oauth2.googleapis.com') {
+      return http.StreamedResponse(
+        Stream.value(
+          utf8.encode(
+            '{"access_token": "ya29.mock", "expires_in": 3600, "token_type": "Bearer"}',
+          ),
+        ),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    if (request.url.path == '/v1beta1/publishers/google/models') {
+      return http.StreamedResponse(
+        Stream.value(
+          utf8.encode(
+            '{"publisherModels": [{"name": "publishers/google/models/gemini-1.5-pro"}, {"name": "publishers/google/models/text-embedding-005"}, {"name": "publishers/google/models/multimodalembedding"}]}',
+          ),
+        ),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    if (request.url.path.endsWith('gemini-embedding-001:embedContent')) {
+      return http.StreamedResponse(
+        Stream.value(
+          utf8.encode(
+            '{"error": {"code": 400, "message": "Publisher Model `projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-001` is not supported in the embedContent API.", "status": "INVALID_ARGUMENT"}}',
+          ),
+        ),
+        400,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    if (request.url.path.endsWith(':embedContent')) {
+      return http.StreamedResponse(
+        Stream.value(utf8.encode('{"embedding": {"values": [0.1, 0.2, 0.3]}}')),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    if (request.url.path.contains('multimodalembedding') &&
+        request.url.path.endsWith(':predict')) {
+      return http.StreamedResponse(
+        Stream.value(
+          utf8.encode('{"predictions": [{"textEmbedding": [0.7, 0.8, 0.9]}]}'),
+        ),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    if (request.url.path.endsWith(':predict')) {
+      return http.StreamedResponse(
+        Stream.value(
+          utf8.encode(
+            '{"predictions": [{"embeddings": {"values": [0.4, 0.5, 0.6]}}]}',
+          ),
+        ),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
+    return http.StreamedResponse(
+      Stream.value(
+        utf8.encode(
+          '{"candidates": [{"content": {"parts": [{"text": "response"}], "role": "model"}, "finishReason": "STOP"}]} ',
+        ),
+      ),
+      200,
+      headers: {'content-type': 'application/json'},
+    );
+  }
+}
diff --git a/packages/genkit_vertexai/test/vertex_test.dart b/packages/genkit_vertexai/test/vertex_test.dart
index 99aba838..c7d96f16 100644
--- a/packages/genkit_vertexai/test/vertex_test.dart
+++ b/packages/genkit_vertexai/test/vertex_test.dart
@@ -12,43 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-import 'dart:convert';
-
 import 'package:genkit/genkit.dart';
-
 import 'package:genkit_vertexai/src/vertex_api_client.dart';
-import 'package:http/http.dart' as http;
 import 'package:test/test.dart';
 
-class MockHttpClient extends http.BaseClient {
-  Uri? lastUrl;
-
-  @override
-  Future<http.StreamedResponse> send(http.BaseRequest request) async {
-    lastUrl = request.url;
-    if (request.url.host == 'metadata.google.internal' ||
-        request.url.host == 'oauth2.googleapis.com') {
-      return http.StreamedResponse(
-        Stream.value(
-          utf8.encode(
-            '{"access_token": "ya29.mock", "expires_in": 3600, "token_type": "Bearer"}',
-          ),
-        ),
-        200,
-        headers: {'content-type': 'application/json'},
-      );
-    }
-    return http.StreamedResponse(
-      Stream.value(
-        utf8.encode(
-          '{"candidates": [{"content": {"parts": [{"text": "response"}], "role": "model"}, "finishReason": "STOP"}]} ',
-        ),
-      ),
-      200,
-      headers: {'content-type': 'application/json'},
-    );
-  }
-}
+import 'test_http_client.dart';
 
 void main() {
   group('Vertex AI Plugin', () {

From b955fdc37609022221bf65452cb21b58e1c1d978 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Wed, 22 Apr 2026 15:55:18 +0100
Subject: [PATCH 02/16] fix(genkit_vertexai): CI problems

---
 .../genkit_vertexai/lib/src/embedders.dart    | 35 +++++++++----------
 .../lib/src/vertex_api_client.dart            |  2 +-
 .../genkit_vertexai/test/embedders_test.dart  | 18 ++++++----
 3 files changed, 29 insertions(+), 26 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index dd42af75..2bb6ad0b 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -15,9 +15,8 @@
 import 'dart:convert';
 
 import 'package:genkit/plugin.dart';
-import 'package:genkit_google_genai/common.dart';
-import 'package:genkit_google_genai/src/generated/generativelanguage.dart'
-    as gcl;
+import 'package:genkit_google_genai/common.dart' as google;
+import 'package:genkit_google_genai/generated.dart' as google_types;
 
 List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
   required String pluginName,
@@ -34,7 +33,7 @@ List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
         final modelName = (modelMap['name'] as String).split('/').last;
         return embedderMetadata(
           '$pluginName/$modelName',
-          customOptions: TextEmbedderOptions.$schema,
+          customOptions: google.TextEmbedderOptions.$schema,
         );
       })
       .toList();
@@ -43,7 +42,7 @@ List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
 Embedder createVertexEmbedder({
   required String pluginName,
   required String embedderName,
-  required Future<GenerativeLanguageBaseClient> Function() getApiClient,
+  required Future<google.GenerativeLanguageBaseClient> Function() getApiClient,
   required GenkitException Function(Object, StackTrace) handleException,
   required bool closeService,
 }) {
@@ -57,7 +56,7 @@ Embedder createVertexEmbedder({
       final service = await getApiClient();
       try {
         final options = req.options != null
-            ? TextEmbedderOptions.fromJson(req.options!)
+            ? google.TextEmbedderOptions.fromJson(req.options!)
             : null;
 
         // Handle each input document separately, because Vertex embedders do
@@ -87,10 +86,10 @@ String _documentText(DocumentData doc) {
 }
 
 Future<Embedding> _runEmbedderRequest({
-  required GenerativeLanguageBaseClient service,
+  required google.GenerativeLanguageBaseClient service,
   required String embedderName,
   required DocumentData doc,
-  required TextEmbedderOptions? options,
+  required google.TextEmbedderOptions? options,
 }) async {
   // Choose the request style from the model family.
   return switch (_requestShapeFor(embedderName)) {
@@ -117,10 +116,10 @@ Future<Embedding> _runEmbedderRequest({
 }
 
 Future<Embedding> _runGeminiEmbeddingRequest({
-  required GenerativeLanguageBaseClient service,
+  required google.GenerativeLanguageBaseClient service,
   required String embedderName,
   required DocumentData doc,
-  required TextEmbedderOptions? options,
+  required google.TextEmbedderOptions? options,
 }) async {
   try {
     // Try the newer Gemini embedding API first.
@@ -146,15 +145,15 @@ Future<Embedding> _runGeminiEmbeddingRequest({
 }
 
 Future<Embedding> _runEmbedContentRequest({
-  required GenerativeLanguageBaseClient service,
+  required google.GenerativeLanguageBaseClient service,
   required String embedderName,
   required DocumentData doc,
-  required TextEmbedderOptions? options,
+  required google.TextEmbedderOptions? options,
 }) async {
   final text = _documentText(doc);
-  final content = gcl.Content(parts: [gcl.Part(text: text)]);
+  final content = google_types.Content(parts: [google_types.Part(text: text)]);
   final res = await service.embedContent(
-    gcl.EmbedContentRequest(
+    google_types.EmbedContentRequest(
       content: content,
       outputDimensionality: options?.outputDimensionality,
       taskType: options?.taskType,
@@ -166,10 +165,10 @@ Future<Embedding> _runEmbedContentRequest({
 }
 
 Future<Embedding> _runMultimodalPredictRequest({
-  required GenerativeLanguageBaseClient service,
+  required google.GenerativeLanguageBaseClient service,
   required String embedderName,
   required DocumentData doc,
-  required TextEmbedderOptions? options,
+  required google.TextEmbedderOptions? options,
 }) async {
   // Multimodal embedders use a different predict request and response shape.
   final instance = _toMultimodalInstance(doc);
@@ -194,10 +193,10 @@ Future<Embedding> _runMultimodalPredictRequest({
 }
 
 Future<Embedding> _runTextPredictRequest({
-  required GenerativeLanguageBaseClient service,
+  required google.GenerativeLanguageBaseClient service,
   required String embedderName,
   required DocumentData doc,
-  required TextEmbedderOptions? options,
+  required google.TextEmbedderOptions? options,
 }) async {
   // Older text embedders still use the predict payload shape.
   final instance = <String, dynamic>{'content': _documentText(doc)};
diff --git a/packages/genkit_vertexai/lib/src/vertex_api_client.dart b/packages/genkit_vertexai/lib/src/vertex_api_client.dart
index 8c6318b3..7717eb67 100644
--- a/packages/genkit_vertexai/lib/src/vertex_api_client.dart
+++ b/packages/genkit_vertexai/lib/src/vertex_api_client.dart
@@ -130,7 +130,7 @@ class VertexAiPluginImpl extends CommonGoogleGenPlugin {
     return createVertexEmbedder(
       pluginName: name,
       embedderName: embedderName,
-      getApiClient: () => getApiClient(),
+      getApiClient: getApiClient,
       handleException: handleException,
       closeService: authClient == null,
     );
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index 84803418..40688ddc 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -20,6 +20,12 @@ import 'package:test/test.dart';
 
 import 'test_http_client.dart';
 
+typedef _EmbedderAction = Action<EmbedRequest, EmbedResponse, void, void>;
+
+_EmbedderAction _resolveEmbedder(VertexAiPluginImpl plugin, String name) {
+  return plugin.resolve('embedder', name)! as _EmbedderAction;
+}
+
 void main() {
   group('Vertex AI Embedders', () {
     test('lists embedders with schemas and custom options', () async {
@@ -39,8 +45,9 @@ void main() {
       expect(embedder.inputSchema, same(EmbedRequest.$schema));
       expect(embedder.outputSchema, same(EmbedResponse.$schema));
 
+      final modelMetadata = embedder.metadata['model'] as Map<String, dynamic>;
       final customOptions =
-          embedder.metadata['model']['customOptions'] as Map<String, dynamic>;
+          modelMetadata['customOptions'] as Map<String, dynamic>;
       expect(customOptions['properties'], contains('outputDimensionality'));
       expect(customOptions['properties'], contains('taskType'));
     });
@@ -53,8 +60,7 @@ void main() {
         authClient: mockClient,
       );
 
-      final embedder =
-          plugin.resolve('embedder', 'gemini-embedding-2-preview') as Action;
+      final embedder = _resolveEmbedder(plugin, 'gemini-embedding-2-preview');
       final req = EmbedRequest(
         input: [
           DocumentData(content: [TextPart(text: 'hello')]),
@@ -80,8 +86,7 @@ void main() {
         authClient: mockClient,
       );
 
-      final embedder =
-          plugin.resolve('embedder', 'gemini-embedding-001') as Action;
+      final embedder = _resolveEmbedder(plugin, 'gemini-embedding-001');
       final req = EmbedRequest(
         input: [
           DocumentData(content: [TextPart(text: 'hello')]),
@@ -109,8 +114,7 @@ void main() {
           authClient: mockClient,
         );
 
-        final embedder =
-            plugin.resolve('embedder', 'multimodalembedding') as Action;
+        final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
         final req = EmbedRequest(
           input: [
             DocumentData(content: [TextPart(text: 'hello')]),

From ef67656649d41cc33c44710d36875b9979312486 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Wed, 22 Apr 2026 19:17:01 +0100
Subject: [PATCH 03/16] refactor(genkit_vertexai): replace google_types with
 google in embedders.dart

---
 packages/genkit_vertexai/lib/src/embedders.dart | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 2bb6ad0b..0cfb1e3b 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -16,7 +16,6 @@ import 'dart:convert';
 
 import 'package:genkit/plugin.dart';
 import 'package:genkit_google_genai/common.dart' as google;
-import 'package:genkit_google_genai/generated.dart' as google_types;
 
 List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
   required String pluginName,
@@ -151,9 +150,9 @@ Future<Embedding> _runEmbedContentRequest({
   required google.TextEmbedderOptions? options,
 }) async {
   final text = _documentText(doc);
-  final content = google_types.Content(parts: [google_types.Part(text: text)]);
+  final content = google.Content(parts: [google.Part(text: text)]);
   final res = await service.embedContent(
-    google_types.EmbedContentRequest(
+    google.EmbedContentRequest(
       content: content,
       outputDimensionality: options?.outputDimensionality,
       taskType: options?.taskType,

From 58f4df0e06a7cde0d2468fe1b862b4626392c82a Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Thu, 23 Apr 2026 04:51:24 +0100
Subject: [PATCH 04/16] fix(ci): resolve PR 12 failures

---
 packages/genkit/lib/src/ai/embedder.dart                        | 2 ++
 packages/genkit_google_genai/lib/common.dart                    | 2 ++
 .../firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift   | 2 +-
 3 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/packages/genkit/lib/src/ai/embedder.dart b/packages/genkit/lib/src/ai/embedder.dart
index 9818cdcb..b5502e14 100644
--- a/packages/genkit/lib/src/ai/embedder.dart
+++ b/packages/genkit/lib/src/ai/embedder.dart
@@ -82,6 +82,8 @@ ActionMetadata embedderMetadata(
     name: name,
     description: name,
     actionType: 'embedder',
+    inputSchema: EmbedRequest.$schema,
+    outputSchema: EmbedResponse.$schema,
     metadata: {
       'label': name,
       'description': name,
diff --git a/packages/genkit_google_genai/lib/common.dart b/packages/genkit_google_genai/lib/common.dart
index 4ca02b8b..5824ce62 100644
--- a/packages/genkit_google_genai/lib/common.dart
+++ b/packages/genkit_google_genai/lib/common.dart
@@ -17,4 +17,6 @@ library;
 
 export 'src/api_client.dart';
 export 'src/common_plugin.dart';
+export 'src/generated/generativelanguage.dart'
+    show Content, EmbedContentRequest, Part;
 export 'src/model.dart';
diff --git a/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift b/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
index 5ffdb1fa..524b2645 100644
--- a/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
+++ b/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
@@ -15,7 +15,7 @@ import record_macos
 func RegisterGeneratedPlugins(registry: FlutterPluginRegistry) {
   AudioplayersDarwinPlugin.register(with: registry.registrar(forPlugin: "AudioplayersDarwinPlugin"))
   FirebaseAIPlugin.register(with: registry.registrar(forPlugin: "FirebaseAIPlugin"))
-  FLTFirebaseAppCheckPlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseAppCheckPlugin"))
+  FirebaseAppCheckPlugin.register(with: registry.registrar(forPlugin: "FirebaseAppCheckPlugin"))
   FLTFirebaseAuthPlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseAuthPlugin"))
   FLTFirebaseCorePlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseCorePlugin"))
   RecordMacOsPlugin.register(with: registry.registrar(forPlugin: "RecordMacOsPlugin"))

From 2c5966a625e9f2cb9fc8f69202ffa3a03d5fc7ff Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Thu, 23 Apr 2026 08:41:50 +0100
Subject: [PATCH 05/16] fix(ci): scope PR 12 fixes

---
 .github/workflows/flutter_ci.yml              |  3 ++
 packages/genkit/lib/src/ai/embedder.dart      |  2 --
 packages/genkit_google_genai/lib/common.dart  |  2 --
 .../genkit_vertexai/lib/src/embedders.dart    | 29 +++++++++++++++----
 .../Flutter/GeneratedPluginRegistrant.swift   |  2 +-
 5 files changed, 27 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/flutter_ci.yml b/.github/workflows/flutter_ci.yml
index fb0b4a20..974ef6e4 100644
--- a/.github/workflows/flutter_ci.yml
+++ b/.github/workflows/flutter_ci.yml
@@ -58,6 +58,9 @@ jobs:
         working-directory: testapps/flutter_genai
         run: flutter analyze
 
+      - name: Normalize generated macOS registrant
+        run: git restore testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
+
       - name: Verify clean git tree
         run: |
           git diff --exit-code
diff --git a/packages/genkit/lib/src/ai/embedder.dart b/packages/genkit/lib/src/ai/embedder.dart
index b5502e14..9818cdcb 100644
--- a/packages/genkit/lib/src/ai/embedder.dart
+++ b/packages/genkit/lib/src/ai/embedder.dart
@@ -82,8 +82,6 @@ ActionMetadata embedderMetadata(
     name: name,
     description: name,
     actionType: 'embedder',
-    inputSchema: EmbedRequest.$schema,
-    outputSchema: EmbedResponse.$schema,
     metadata: {
       'label': name,
       'description': name,
diff --git a/packages/genkit_google_genai/lib/common.dart b/packages/genkit_google_genai/lib/common.dart
index 5824ce62..4ca02b8b 100644
--- a/packages/genkit_google_genai/lib/common.dart
+++ b/packages/genkit_google_genai/lib/common.dart
@@ -17,6 +17,4 @@ library;
 
 export 'src/api_client.dart';
 export 'src/common_plugin.dart';
-export 'src/generated/generativelanguage.dart'
-    show Content, EmbedContentRequest, Part;
 export 'src/model.dart';
diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 0cfb1e3b..29ddb6c3 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -16,6 +16,9 @@ import 'dart:convert';
 
 import 'package:genkit/plugin.dart';
 import 'package:genkit_google_genai/common.dart' as google;
+// ignore: implementation_imports
+import 'package:genkit_google_genai/src/generated/generativelanguage.dart'
+    as google_types;
 
 List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
   required String pluginName,
@@ -30,14 +33,28 @@ List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
       .map((m) {
         final modelMap = m as Map<String, dynamic>;
         final modelName = (modelMap['name'] as String).split('/').last;
-        return embedderMetadata(
-          '$pluginName/$modelName',
-          customOptions: google.TextEmbedderOptions.$schema,
-        );
+        return _vertexEmbedderMetadata('$pluginName/$modelName');
       })
       .toList();
 }
 
+ActionMetadata<dynamic, dynamic, dynamic, dynamic> _vertexEmbedderMetadata(
+  String name,
+) {
+  final metadata = embedderMetadata(
+    name,
+    customOptions: google.TextEmbedderOptions.$schema,
+  );
+  return ActionMetadata(
+    name: metadata.name,
+    description: metadata.description,
+    actionType: metadata.actionType,
+    inputSchema: EmbedRequest.$schema,
+    outputSchema: EmbedResponse.$schema,
+    metadata: metadata.metadata,
+  );
+}
+
 Embedder createVertexEmbedder({
   required String pluginName,
   required String embedderName,
@@ -150,9 +167,9 @@ Future<Embedding> _runEmbedContentRequest({
   required google.TextEmbedderOptions? options,
 }) async {
   final text = _documentText(doc);
-  final content = google.Content(parts: [google.Part(text: text)]);
+  final content = google_types.Content(parts: [google_types.Part(text: text)]);
   final res = await service.embedContent(
-    google.EmbedContentRequest(
+    google_types.EmbedContentRequest(
       content: content,
       outputDimensionality: options?.outputDimensionality,
       taskType: options?.taskType,
diff --git a/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift b/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
index 524b2645..5ffdb1fa 100644
--- a/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
+++ b/testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
@@ -15,7 +15,7 @@ import record_macos
 func RegisterGeneratedPlugins(registry: FlutterPluginRegistry) {
   AudioplayersDarwinPlugin.register(with: registry.registrar(forPlugin: "AudioplayersDarwinPlugin"))
   FirebaseAIPlugin.register(with: registry.registrar(forPlugin: "FirebaseAIPlugin"))
-  FirebaseAppCheckPlugin.register(with: registry.registrar(forPlugin: "FirebaseAppCheckPlugin"))
+  FLTFirebaseAppCheckPlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseAppCheckPlugin"))
   FLTFirebaseAuthPlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseAuthPlugin"))
   FLTFirebaseCorePlugin.register(with: registry.registrar(forPlugin: "FLTFirebaseCorePlugin"))
   RecordMacOsPlugin.register(with: registry.registrar(forPlugin: "RecordMacOsPlugin"))

From f3cd2556ee34a08b17fbc61dd44a6f92273f9290 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Thu, 23 Apr 2026 15:27:00 +0100
Subject: [PATCH 06/16] fix(genkit_vertexai): address PR review feedback on
 embedders

---
 .github/workflows/flutter_ci.yml              |   3 -
 .../lib/src/api_client.dart                   |   9 +
 .../genkit_vertexai/lib/src/embedders.dart    | 222 +++++++++++-------
 .../genkit_vertexai/test/embedders_test.dart  | 125 +++++++++-
 .../test/test_http_client.dart                |  63 ++++-
 testapps/firebase_ai/pubspec.yaml             |   1 +
 6 files changed, 320 insertions(+), 103 deletions(-)

diff --git a/.github/workflows/flutter_ci.yml b/.github/workflows/flutter_ci.yml
index 974ef6e4..fb0b4a20 100644
--- a/.github/workflows/flutter_ci.yml
+++ b/.github/workflows/flutter_ci.yml
@@ -58,9 +58,6 @@ jobs:
         working-directory: testapps/flutter_genai
         run: flutter analyze
 
-      - name: Normalize generated macOS registrant
-        run: git restore testapps/firebase_ai/macos/Flutter/GeneratedPluginRegistrant.swift
-
       - name: Verify clean git tree
         run: |
           git diff --exit-code
diff --git a/packages/genkit_google_genai/lib/src/api_client.dart b/packages/genkit_google_genai/lib/src/api_client.dart
index e667b98b..469e3ff9 100644
--- a/packages/genkit_google_genai/lib/src/api_client.dart
+++ b/packages/genkit_google_genai/lib/src/api_client.dart
@@ -40,6 +40,15 @@ class GenerativeLanguageBaseClient {
     return EmbedContentResponse.fromJson(res);
   }
 
+  Future<BatchEmbedContentsResponse> batchEmbedContents(
+    BatchEmbedContentsRequest request, {
+    required String model,
+  }) async {
+    final url = '$apiUrlPrefix$model:batchEmbedContents';
+    final res = await _call('POST', url, request.toJson());
+    return BatchEmbedContentsResponse.fromJson(res);
+  }
+
   Future<GenerateContentResponse> generateContent(
     GenerateContentRequest request, {
     required String model,
diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 29ddb6c3..64917b7f 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -75,17 +75,29 @@ Embedder createVertexEmbedder({
             ? google.TextEmbedderOptions.fromJson(req.options!)
             : null;
 
-        // Handle each input document separately, because Vertex embedders do
-        // not all use the same request shape.
-        final futures = req.input.map(
-          (doc) => _runEmbedderRequest(
+        final embeddings = switch (_requestShapeFor(embedderName)) {
+          _VertexEmbedderRequestShape.geminiEmbedding =>
+            _runGeminiEmbeddingRequests(
+              service: service,
+              embedderName: embedderName,
+              docs: req.input,
+              options: options,
+            ),
+          _VertexEmbedderRequestShape.multimodalPredict =>
+            _runMultimodalPredictRequests(
+              service: service,
+              embedderName: embedderName,
+              docs: req.input,
+              options: options,
+            ),
+          _VertexEmbedderRequestShape.textPredict => _runTextPredictRequests(
             service: service,
             embedderName: embedderName,
-            doc: doc,
+            docs: req.input,
             options: options,
           ),
-        );
-        return EmbedResponse(embeddings: await Future.wait(futures));
+        };
+        return EmbedResponse(embeddings: await embeddings);
       } catch (e, stack) {
         throw handleException(e, stack);
       } finally {
@@ -101,63 +113,49 @@ String _documentText(DocumentData doc) {
   return doc.content.where((p) => p.isText).map((p) => p.text).join('\n');
 }
 
-Future<Embedding> _runEmbedderRequest({
+google_types.EmbedContentRequest _embedContentRequest(
+  DocumentData doc,
+  google.TextEmbedderOptions? options,
+) {
+  final text = _documentText(doc);
+  return google_types.EmbedContentRequest(
+    content: google_types.Content(parts: [google_types.Part(text: text)]),
+    outputDimensionality: options?.outputDimensionality,
+    taskType: options?.taskType,
+    title: options?.title,
+  );
+}
+
+Future<List<Embedding>> _runGeminiEmbeddingRequests({
   required google.GenerativeLanguageBaseClient service,
   required String embedderName,
-  required DocumentData doc,
+  required List<DocumentData> docs,
   required google.TextEmbedderOptions? options,
 }) async {
-  // Choose the request style from the model family.
-  return switch (_requestShapeFor(embedderName)) {
-    _VertexEmbedderRequestShape.geminiEmbedding => _runGeminiEmbeddingRequest(
-      service: service,
-      embedderName: embedderName,
-      doc: doc,
-      options: options,
-    ),
-    _VertexEmbedderRequestShape.multimodalPredict =>
-      _runMultimodalPredictRequest(
+  if (docs.length == 1) {
+    return [
+      await _runEmbedContentRequest(
         service: service,
         embedderName: embedderName,
-        doc: doc,
+        doc: docs.single,
         options: options,
       ),
-    _VertexEmbedderRequestShape.textPredict => _runTextPredictRequest(
-      service: service,
-      embedderName: embedderName,
-      doc: doc,
-      options: options,
-    ),
-  };
-}
-
-Future<Embedding> _runGeminiEmbeddingRequest({
-  required google.GenerativeLanguageBaseClient service,
-  required String embedderName,
-  required DocumentData doc,
-  required google.TextEmbedderOptions? options,
-}) async {
-  try {
-    // Try the newer Gemini embedding API first.
-    return await _runEmbedContentRequest(
-      service: service,
-      embedderName: embedderName,
-      doc: doc,
-      options: options,
-    );
-  } on GenkitException catch (e) {
-    if (!_shouldFallbackToTextPredict(e)) {
-      rethrow;
-    }
+    ];
   }
 
-  // Some older Gemini embedding models still need the predict API.
-  return _runTextPredictRequest(
-    service: service,
-    embedderName: embedderName,
-    doc: doc,
-    options: options,
+  final res = await service.batchEmbedContents(
+    google_types.BatchEmbedContentsRequest(
+      requests: docs.map((doc) => _embedContentRequest(doc, options)).toList(),
+    ),
+    model: 'models/$embedderName',
   );
+  final embeddings = _requireBatchEmbeddings(
+    res.embeddings,
+    expectedCount: docs.length,
+  );
+  return embeddings
+      .map((embedding) => Embedding(embedding: embedding.values ?? const []))
+      .toList();
 }
 
 Future<Embedding> _runEmbedContentRequest({
@@ -166,59 +164,107 @@ Future<Embedding> _runEmbedContentRequest({
   required DocumentData doc,
   required google.TextEmbedderOptions? options,
 }) async {
-  final text = _documentText(doc);
-  final content = google_types.Content(parts: [google_types.Part(text: text)]);
   final res = await service.embedContent(
-    google_types.EmbedContentRequest(
-      content: content,
-      outputDimensionality: options?.outputDimensionality,
-      taskType: options?.taskType,
-      title: options?.title,
-    ),
+    _embedContentRequest(doc, options),
     model: 'models/$embedderName',
   );
   return Embedding(embedding: res.embedding?.values ?? []);
 }
 
-Future<Embedding> _runMultimodalPredictRequest({
+List<google_types.ContentEmbedding> _requireBatchEmbeddings(
+  List<google_types.ContentEmbedding>? embeddings, {
+  required int expectedCount,
+}) {
+  if (embeddings == null || embeddings.isEmpty) {
+    throw GenkitException(
+      'Vertex AI returned no embeddings.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+  if (embeddings.length != expectedCount) {
+    throw GenkitException(
+      'Vertex AI returned ${embeddings.length} embeddings for $expectedCount input documents.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+  return embeddings;
+}
+
+List<Map<String, dynamic>> _requirePredictions(
+  Object? rawPredictions, {
+  required int expectedCount,
+}) {
+  if (rawPredictions is! List || rawPredictions.isEmpty) {
+    throw GenkitException(
+      'Vertex AI returned no predictions.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+  if (rawPredictions.length != expectedCount) {
+    throw GenkitException(
+      'Vertex AI returned ${rawPredictions.length} predictions for $expectedCount input documents.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+
+  return rawPredictions.map((prediction) {
+    if (prediction is! Map<String, dynamic>) {
+      throw GenkitException(
+        'Vertex AI returned an invalid prediction payload.',
+        status: StatusCodes.INTERNAL,
+      );
+    }
+    return prediction;
+  }).toList();
+}
+
+Future<List<Embedding>> _runMultimodalPredictRequests({
   required google.GenerativeLanguageBaseClient service,
   required String embedderName,
-  required DocumentData doc,
+  required List<DocumentData> docs,
   required google.TextEmbedderOptions? options,
 }) async {
   // Multimodal embedders use a different predict request and response shape.
-  final instance = _toMultimodalInstance(doc);
+  final instances = docs.map(_toMultimodalInstance).toList();
   final parameters = <String, dynamic>{};
   if (options?.outputDimensionality != null) {
     parameters['dimension'] = options!.outputDimensionality;
   }
 
   final res = await service.predict({
-    'instances': [instance.instance],
+    'instances': instances.map((instance) => instance.instance).toList(),
     if (parameters.isNotEmpty) 'parameters': parameters,
   }, model: 'models/$embedderName');
 
-  final predictions = res['predictions'] as List;
-  final prediction = predictions.single as Map<String, dynamic>;
-  return Embedding(
-    embedding: _multimodalPredictionEmbedding(
-      prediction,
-      expectedOutput: instance.expectedOutput,
-    ),
+  final predictions = _requirePredictions(
+    res['predictions'],
+    expectedCount: instances.length,
   );
+  return [
+    for (var i = 0; i < predictions.length; i++)
+      Embedding(
+        embedding: _multimodalPredictionEmbedding(
+          predictions[i],
+          expectedOutput: instances[i].expectedOutput,
+        ),
+      ),
+  ];
 }
 
-Future<Embedding> _runTextPredictRequest({
+Future<List<Embedding>> _runTextPredictRequests({
   required google.GenerativeLanguageBaseClient service,
   required String embedderName,
-  required DocumentData doc,
+  required List<DocumentData> docs,
   required google.TextEmbedderOptions? options,
 }) async {
   // Older text embedders still use the predict payload shape.
-  final instance = <String, dynamic>{'content': _documentText(doc)};
-  if (options?.title != null) {
-    instance['title'] = options!.title;
-  }
+  final instances = docs.map((doc) {
+    final instance = <String, dynamic>{'content': _documentText(doc)};
+    if (options?.title != null) {
+      instance['title'] = options!.title;
+    }
+    return instance;
+  }).toList();
 
   final parameters = <String, dynamic>{};
   if (options?.outputDimensionality != null) {
@@ -229,12 +275,18 @@ Future<Embedding> _runTextPredictRequest({
   }
 
   final res = await service.predict({
-    'instances': [instance],
+    'instances': instances,
     if (parameters.isNotEmpty) 'parameters': parameters,
   }, model: 'models/$embedderName');
 
-  final predictions = res['predictions'] as List;
-  final prediction = predictions.single as Map<String, dynamic>;
+  final predictions = _requirePredictions(
+    res['predictions'],
+    expectedCount: docs.length,
+  );
+  return predictions.map(_textPredictionEmbedding).toList();
+}
+
+Embedding _textPredictionEmbedding(Map<String, dynamic> prediction) {
   final embeddingData = prediction['embeddings'] as Map<String, dynamic>;
   final values = embeddingData['values'] as List;
   return Embedding(
@@ -391,6 +443,9 @@ _VertexEmbedderRequestShape _requestShapeFor(String modelName) {
   if (_isMultimodalEmbeddingFamily(baseModelName)) {
     return _VertexEmbedderRequestShape.multimodalPredict;
   }
+  if (_usesLegacyGeminiPredictApi(baseModelName)) {
+    return _VertexEmbedderRequestShape.textPredict;
+  }
   if (_isGeminiEmbeddingFamily(baseModelName)) {
     return _VertexEmbedderRequestShape.geminiEmbedding;
   }
@@ -405,9 +460,8 @@ bool _isGeminiEmbeddingFamily(String modelName) {
   return modelName.startsWith('gemini-embedding-');
 }
 
-bool _shouldFallbackToTextPredict(GenkitException error) {
-  return error.status == StatusCodes.INVALID_ARGUMENT &&
-      error.message.contains('not supported in the embedContent API');
+bool _usesLegacyGeminiPredictApi(String modelName) {
+  return modelName == 'gemini-embedding-001';
 }
 
 class _MultimodalInstance {
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index 40688ddc..fc59a312 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -52,7 +52,7 @@ void main() {
       expect(customOptions['properties'], contains('taskType'));
     });
 
-    test('uses embedContent for embedders', () async {
+    test('uses embedContent for a single Gemini preview input', () async {
       final mockClient = MockHttpClient();
       final plugin = VertexAiPluginImpl(
         projectId: 'my-project',
@@ -78,7 +78,7 @@ void main() {
       expect(response.result.embeddings.first.embedding, [0.1, 0.2, 0.3]);
     });
 
-    test('uses predict for gemini-embedding-001', () async {
+    test('uses batchEmbedContents for multiple Gemini preview inputs', () async {
       final mockClient = MockHttpClient();
       final plugin = VertexAiPluginImpl(
         projectId: 'my-project',
@@ -86,10 +86,11 @@ void main() {
         authClient: mockClient,
       );
 
-      final embedder = _resolveEmbedder(plugin, 'gemini-embedding-001');
+      final embedder = _resolveEmbedder(plugin, 'gemini-embedding-2-preview');
       final req = EmbedRequest(
         input: [
           DocumentData(content: [TextPart(text: 'hello')]),
+          DocumentData(content: [TextPart(text: 'world')]),
         ],
       );
 
@@ -98,14 +99,60 @@ void main() {
       expect(mockClient.lastUrl, isNotNull);
       expect(
         mockClient.lastUrl.toString(),
-        'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-001:predict',
+        'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-2-preview:batchEmbedContents',
       );
-      expect(response.result.embeddings, hasLength(1));
-      expect(response.result.embeddings.first.embedding, [0.4, 0.5, 0.6]);
+      final requestBody =
+          jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+      final requests = requestBody['requests'] as List;
+      expect(requests, hasLength(2));
+      expect(response.result.embeddings, hasLength(2));
+      expect(response.result.embeddings[0].embedding, [0.1, 0.2, 0.3]);
+      expect(response.result.embeddings[1].embedding, [1.1, 1.2, 1.3]);
     });
 
     test(
-      'uses multimodal predict schema for multimodalembedding text input',
+      'uses predict directly for legacy gemini-embedding-001 inputs',
+      () async {
+        final mockClient = MockHttpClient();
+        final plugin = VertexAiPluginImpl(
+          projectId: 'my-project',
+          location: 'us-central1',
+          authClient: mockClient,
+        );
+
+        final embedder = _resolveEmbedder(plugin, 'gemini-embedding-001');
+        final req = EmbedRequest(
+          input: [
+            DocumentData(content: [TextPart(text: 'hello')]),
+            DocumentData(content: [TextPart(text: 'world')]),
+          ],
+        );
+
+        final response = await embedder.run(req);
+
+        expect(mockClient.lastUrl, isNotNull);
+        expect(
+          mockClient.lastUrl.toString(),
+          'https://us-central1-aiplatform.googleapis.com/v1beta1/projects/my-project/locations/us-central1/publishers/google/models/gemini-embedding-001:predict',
+        );
+        expect(
+          mockClient.requestUrls.where(
+            (url) => url.path.endsWith('gemini-embedding-001:embedContent'),
+          ),
+          isEmpty,
+        );
+        final requestBody =
+            jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+        final instances = requestBody['instances'] as List;
+        expect(instances, hasLength(2));
+        expect(response.result.embeddings, hasLength(2));
+        expect(response.result.embeddings[0].embedding, [0.4, 0.5, 0.6]);
+        expect(response.result.embeddings[1].embedding, [1.4, 1.5, 1.6]);
+      },
+    );
+
+    test(
+      'uses multimodal predict schema for text-only multimodal inputs',
       () async {
         final mockClient = MockHttpClient();
         final plugin = VertexAiPluginImpl(
@@ -142,5 +189,69 @@ void main() {
         expect(response.result.embeddings.first.embedding, [0.7, 0.8, 0.9]);
       },
     );
+
+    test(
+      'throws a descriptive error when Vertex returns no predictions',
+      () async {
+        final mockClient = MockHttpClient(returnEmptyPredictions: true);
+        final plugin = VertexAiPluginImpl(
+          projectId: 'my-project',
+          location: 'us-central1',
+          authClient: mockClient,
+        );
+
+        final embedder = _resolveEmbedder(plugin, 'gemini-embedding-001');
+
+        await expectLater(
+          () => embedder.run(
+            EmbedRequest(
+              input: [
+                DocumentData(content: [TextPart(text: 'hello')]),
+              ],
+            ),
+          ),
+          throwsA(
+            isA<GenkitException>().having(
+              (error) => error.message,
+              'message',
+              contains('Vertex AI returned no predictions.'),
+            ),
+          ),
+        );
+      },
+    );
+
+    test(
+      'throws when a multimodal response omits the expected embedding field',
+      () async {
+        final mockClient = MockHttpClient(
+          returnMissingMultimodalEmbedding: true,
+        );
+        final plugin = VertexAiPluginImpl(
+          projectId: 'my-project',
+          location: 'us-central1',
+          authClient: mockClient,
+        );
+
+        final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+        await expectLater(
+          () => embedder.run(
+            EmbedRequest(
+              input: [
+                DocumentData(content: [TextPart(text: 'hello')]),
+              ],
+            ),
+          ),
+          throwsA(
+            isA<GenkitException>().having(
+              (error) => error.message,
+              'message',
+              contains('did not return a text embedding'),
+            ),
+          ),
+        );
+      },
+    );
   });
 }
diff --git a/packages/genkit_vertexai/test/test_http_client.dart b/packages/genkit_vertexai/test/test_http_client.dart
index 32eba6fe..ef4de098 100644
--- a/packages/genkit_vertexai/test/test_http_client.dart
+++ b/packages/genkit_vertexai/test/test_http_client.dart
@@ -17,14 +17,26 @@ import 'dart:convert';
 import 'package:http/http.dart' as http;
 
 class MockHttpClient extends http.BaseClient {
+  MockHttpClient({
+    this.returnEmptyPredictions = false,
+    this.returnMissingMultimodalEmbedding = false,
+  });
+
+  final bool returnEmptyPredictions;
+  final bool returnMissingMultimodalEmbedding;
+  final List<Uri> requestUrls = [];
+  final List<String> requestBodies = [];
   Uri? lastUrl;
   String? lastBody;
 
   @override
   Future<http.StreamedResponse> send(http.BaseRequest request) async {
+    requestUrls.add(request.url);
     lastUrl = request.url;
+    final requestBody = request is http.Request ? request.body : null;
     if (request is http.Request) {
-      lastBody = request.body;
+      lastBody = requestBody;
+      requestBodies.add(requestBody!);
     }
     if (request.url.host == 'metadata.google.internal' ||
         request.url.host == 'oauth2.googleapis.com') {
@@ -60,6 +72,21 @@ class MockHttpClient extends http.BaseClient {
         headers: {'content-type': 'application/json'},
       );
     }
+    if (request.url.path.endsWith(':batchEmbedContents')) {
+      final body = jsonDecode(requestBody!) as Map<String, dynamic>;
+      final requests = body['requests'] as List;
+      final embeddings = List.generate(
+        requests.length,
+        (index) => {
+          'values': [index + 0.1, index + 0.2, index + 0.3],
+        },
+      );
+      return http.StreamedResponse(
+        Stream.value(utf8.encode(jsonEncode({'embeddings': embeddings}))),
+        200,
+        headers: {'content-type': 'application/json'},
+      );
+    }
     if (request.url.path.endsWith(':embedContent')) {
       return http.StreamedResponse(
         Stream.value(utf8.encode('{"embedding": {"values": [0.1, 0.2, 0.3]}}')),
@@ -69,21 +96,39 @@ class MockHttpClient extends http.BaseClient {
     }
     if (request.url.path.contains('multimodalembedding') &&
         request.url.path.endsWith(':predict')) {
+      final body = jsonDecode(requestBody!) as Map<String, dynamic>;
+      final instances = body['instances'] as List;
+      final predictions = returnEmptyPredictions
+          ? const []
+          : List.generate(
+              instances.length,
+              (index) => returnMissingMultimodalEmbedding
+                  ? <String, dynamic>{}
+                  : {
+                      'textEmbedding': [index + 0.7, index + 0.8, index + 0.9],
+                    },
+            );
       return http.StreamedResponse(
-        Stream.value(
-          utf8.encode('{"predictions": [{"textEmbedding": [0.7, 0.8, 0.9]}]}'),
-        ),
+        Stream.value(utf8.encode(jsonEncode({'predictions': predictions}))),
         200,
         headers: {'content-type': 'application/json'},
       );
     }
     if (request.url.path.endsWith(':predict')) {
+      final body = jsonDecode(requestBody!) as Map<String, dynamic>;
+      final instances = body['instances'] as List;
+      final predictions = returnEmptyPredictions
+          ? const []
+          : List.generate(
+              instances.length,
+              (index) => {
+                'embeddings': {
+                  'values': [index + 0.4, index + 0.5, index + 0.6],
+                },
+              },
+            );
       return http.StreamedResponse(
-        Stream.value(
-          utf8.encode(
-            '{"predictions": [{"embeddings": {"values": [0.4, 0.5, 0.6]}}]}',
-          ),
-        ),
+        Stream.value(utf8.encode(jsonEncode({'predictions': predictions}))),
         200,
         headers: {'content-type': 'application/json'},
       );
diff --git a/testapps/firebase_ai/pubspec.yaml b/testapps/firebase_ai/pubspec.yaml
index 17c65f2e..49a4a73e 100644
--- a/testapps/firebase_ai/pubspec.yaml
+++ b/testapps/firebase_ai/pubspec.yaml
@@ -35,6 +35,7 @@ dev_dependencies:
     sdk: flutter
 
 dependency_overrides:
+  firebase_app_check: 0.4.1+4
   genkit:
     path: ../../packages/genkit
   genkit_firebase_ai:

From fa49544ebe673df6c98bea15262c9a98845805a5 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Mon, 27 Apr 2026 16:53:09 +0100
Subject: [PATCH 07/16] chore(firebase_ai): remove firebase_app_check
 dependency

---
 testapps/firebase_ai/pubspec.yaml | 1 -
 1 file changed, 1 deletion(-)

diff --git a/testapps/firebase_ai/pubspec.yaml b/testapps/firebase_ai/pubspec.yaml
index 49a4a73e..17c65f2e 100644
--- a/testapps/firebase_ai/pubspec.yaml
+++ b/testapps/firebase_ai/pubspec.yaml
@@ -35,7 +35,6 @@ dev_dependencies:
     sdk: flutter
 
 dependency_overrides:
-  firebase_app_check: 0.4.1+4
   genkit:
     path: ../../packages/genkit
   genkit_firebase_ai:

From 45773e66ae6319817c11479869a7cf6b5d73c31c Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Sun, 10 May 2026 00:32:20 +0100
Subject: [PATCH 08/16] feat(embedders): add task_type option to text predict
 requests and update tests

---
 .../genkit_vertexai/lib/src/embedders.dart    |  6 ++--
 .../genkit_vertexai/test/embedders_test.dart  | 33 +++++++++++++++++++
 2 files changed, 36 insertions(+), 3 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 64917b7f..fb63bc02 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -263,6 +263,9 @@ Future<List<Embedding>> _runTextPredictRequests({
     if (options?.title != null) {
       instance['title'] = options!.title;
     }
+    if (options?.taskType != null) {
+      instance['task_type'] = options!.taskType;
+    }
     return instance;
   }).toList();
 
@@ -270,9 +273,6 @@ Future<List<Embedding>> _runTextPredictRequests({
   if (options?.outputDimensionality != null) {
     parameters['outputDimensionality'] = options!.outputDimensionality;
   }
-  if (options?.taskType != null) {
-    parameters['taskType'] = options!.taskType;
-  }
 
   final res = await service.predict({
     'instances': instances,
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index fc59a312..50694c74 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -151,6 +151,39 @@ void main() {
       },
     );
 
+    test('uses text predict REST option schema', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'text-embedding-005');
+      final req = EmbedRequest(
+        input: [
+          DocumentData(content: [TextPart(text: 'hello')]),
+        ],
+        options: {
+          'outputDimensionality': 256,
+          'taskType': 'RETRIEVAL_DOCUMENT',
+          'title': 'document title',
+        },
+      );
+
+      await embedder.run(req);
+
+      final requestBody =
+          jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+      final instances = requestBody['instances'] as List;
+      expect(instances.single, {
+        'content': 'hello',
+        'task_type': 'RETRIEVAL_DOCUMENT',
+        'title': 'document title',
+      });
+      expect(requestBody['parameters'], {'outputDimensionality': 256});
+    });
+
     test(
       'uses multimodal predict schema for text-only multimodal inputs',
       () async {

From f3d989facc8cfa8997d42c743e7ff4ee6c554fc9 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 13:00:36 +0100
Subject: [PATCH 09/16] feat(genkit_vertexai): flatten multimodal embeddings
 with source metadata

---
 .../genkit_vertexai/lib/src/embedders.dart    | 172 +++++++++++++-----
 .../genkit_vertexai/test/embedders_test.dart  |  71 ++++++++
 .../test/test_http_client.dart                |  27 ++-
 3 files changed, 214 insertions(+), 56 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index fb63bc02..3acd4843 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -225,7 +225,10 @@ Future<List<Embedding>> _runMultimodalPredictRequests({
   required google.TextEmbedderOptions? options,
 }) async {
   // Multimodal embedders use a different predict request and response shape.
-  final instances = docs.map(_toMultimodalInstance).toList();
+  final instances = [
+    for (var i = 0; i < docs.length; i++)
+      _toMultimodalInstance(docs[i], documentIndex: i),
+  ];
   final parameters = <String, dynamic>{};
   if (options?.outputDimensionality != null) {
     parameters['dimension'] = options!.outputDimensionality;
@@ -242,11 +245,9 @@ Future<List<Embedding>> _runMultimodalPredictRequests({
   );
   return [
     for (var i = 0; i < predictions.length; i++)
-      Embedding(
-        embedding: _multimodalPredictionEmbedding(
-          predictions[i],
-          expectedOutput: instances[i].expectedOutput,
-        ),
+      ..._multimodalPredictionEmbeddings(
+        predictions[i],
+        expectedOutputs: instances[i].expectedOutputs,
       ),
   ];
 }
@@ -294,50 +295,70 @@ Embedding _textPredictionEmbedding(Map<String, dynamic> prediction) {
   );
 }
 
-_MultimodalInstance _toMultimodalInstance(DocumentData doc) {
+_MultimodalInstance _toMultimodalInstance(
+  DocumentData doc, {
+  required int documentIndex,
+}) {
   final text = _documentText(doc).trim();
-  final mediaParts = doc.content
-      .where((part) => part.isMedia)
-      .map((part) => part.media!)
-      .toList();
+  final instance = <String, dynamic>{};
+  final expectedOutputs = <_MultimodalExpectedOutput>[];
 
-  // A document can only contain one input type here.
-  // Text, image, and video use different embedding fields in the Vertex
-  // response, and this code needs one clear field to read for each document.
-  if (text.isNotEmpty && mediaParts.isNotEmpty) {
-    throw GenkitException(
-      'Vertex multimodalembedding supports exactly one modality per input document in the embedder API. Provide text, one image, or one video.',
-      status: StatusCodes.INVALID_ARGUMENT,
+  if (text.isNotEmpty) {
+    instance['text'] = text;
+    expectedOutputs.add(
+      _MultimodalExpectedOutput(
+        output: _MultimodalOutput.text,
+        metadata: {
+          'documentIndex': documentIndex,
+          'modality': 'text',
+          'partIndices': [
+            for (var i = 0; i < doc.content.length; i++)
+              if (doc.content[i].isText &&
+                  (doc.content[i].text?.trim().isNotEmpty ?? false))
+                i,
+          ],
+        },
+      ),
     );
   }
 
-  if (mediaParts.length > 1) {
-    throw GenkitException(
-      'Vertex multimodalembedding supports at most one media part per input document in the embedder API.',
-      status: StatusCodes.INVALID_ARGUMENT,
-    );
-  }
+  for (var i = 0; i < doc.content.length; i++) {
+    final part = doc.content[i];
+    if (!part.isMedia) continue;
 
-  if (text.isNotEmpty) {
-    return _MultimodalInstance(
-      instance: {'text': text},
-      expectedOutput: _MultimodalOutput.text,
+    final mediaField = _toMultimodalMediaField(part.media!);
+    if (instance.containsKey(mediaField.key)) {
+      throw GenkitException(
+        'Vertex multimodalembedding supports at most one ${mediaField.key} part per input document.',
+        status: StatusCodes.INVALID_ARGUMENT,
+      );
+    }
+
+    instance[mediaField.key] = mediaField.value;
+    expectedOutputs.add(
+      _MultimodalExpectedOutput(
+        output: mediaField.key == 'image'
+            ? _MultimodalOutput.image
+            : _MultimodalOutput.video,
+        metadata: {
+          'documentIndex': documentIndex,
+          'modality': mediaField.key,
+          'partIndex': i,
+        },
+      ),
     );
   }
 
-  if (mediaParts.isEmpty) {
+  if (instance.isEmpty) {
     throw GenkitException(
       'Vertex multimodalembedding requires text, image, or video input.',
       status: StatusCodes.INVALID_ARGUMENT,
     );
   }
 
-  final mediaField = _toMultimodalMediaField(mediaParts.single);
   return _MultimodalInstance(
-    instance: {mediaField.key: mediaField.value},
-    expectedOutput: mediaField.key == 'image'
-        ? _MultimodalOutput.image
-        : _MultimodalOutput.video,
+    instance: instance,
+    expectedOutputs: expectedOutputs,
   );
 }
 
@@ -407,28 +428,74 @@ String _multimodalFieldName(String? mimeType) {
   );
 }
 
-List<double> _multimodalPredictionEmbedding(
+List<Embedding> _multimodalPredictionEmbeddings(
   Map<String, dynamic> prediction, {
-  required _MultimodalOutput expectedOutput,
+  required List<_MultimodalExpectedOutput> expectedOutputs,
 }) {
-  // Read the embedding field that matches the input type.
-  final values = switch (expectedOutput) {
-    _MultimodalOutput.text => prediction['textEmbedding'] as List?,
-    _MultimodalOutput.image => prediction['imageEmbedding'] as List?,
-    _MultimodalOutput.video =>
-      ((prediction['videoEmbeddings'] as List?)?.firstOrNull
-              as Map<String, dynamic>?)?['embedding']
-          as List?,
-  };
+  final embeddings = <Embedding>[];
+  for (final expectedOutput in expectedOutputs) {
+    switch (expectedOutput.output) {
+      case _MultimodalOutput.text:
+        embeddings.add(
+          _embeddingFromMultimodalValues(
+            prediction['textEmbedding'] as List?,
+            expectedOutput: expectedOutput,
+          ),
+        );
+      case _MultimodalOutput.image:
+        embeddings.add(
+          _embeddingFromMultimodalValues(
+            prediction['imageEmbedding'] as List?,
+            expectedOutput: expectedOutput,
+          ),
+        );
+      case _MultimodalOutput.video:
+        final videoEmbeddings = prediction['videoEmbeddings'] as List?;
+        if (videoEmbeddings == null || videoEmbeddings.isEmpty) {
+          throw GenkitException(
+            'Vertex multimodalembedding did not return a video embedding.',
+            status: StatusCodes.INTERNAL,
+          );
+        }
 
+        for (var i = 0; i < videoEmbeddings.length; i++) {
+          final videoEmbedding = videoEmbeddings[i] as Map<String, dynamic>;
+          embeddings.add(
+            _embeddingFromMultimodalValues(
+              videoEmbedding['embedding'] as List?,
+              expectedOutput: expectedOutput,
+              metadata: {
+                ...expectedOutput.metadata,
+                'segmentIndex': i,
+                if (videoEmbedding['startOffsetSec'] != null)
+                  'startOffsetSec': videoEmbedding['startOffsetSec'],
+                if (videoEmbedding['endOffsetSec'] != null)
+                  'endOffsetSec': videoEmbedding['endOffsetSec'],
+              },
+            ),
+          );
+        }
+    }
+  }
+  return embeddings;
+}
+
+Embedding _embeddingFromMultimodalValues(
+  List? values, {
+  required _MultimodalExpectedOutput expectedOutput,
+  Map<String, dynamic>? metadata,
+}) {
   if (values == null) {
     throw GenkitException(
-      'Vertex multimodalembedding did not return a ${expectedOutput.name} embedding.',
+      'Vertex multimodalembedding did not return a ${expectedOutput.output.name} embedding.',
       status: StatusCodes.INTERNAL,
     );
   }
 
-  return values.map((value) => (value as num).toDouble()).toList();
+  return Embedding(
+    embedding: values.map((value) => (value as num).toDouble()).toList(),
+    metadata: metadata ?? expectedOutput.metadata,
+  );
 }
 
 String _baseModelName(String modelName) {
@@ -466,9 +533,16 @@ bool _usesLegacyGeminiPredictApi(String modelName) {
 
 class _MultimodalInstance {
   final Map<String, dynamic> instance;
-  final _MultimodalOutput expectedOutput;
+  final List<_MultimodalExpectedOutput> expectedOutputs;
+
+  _MultimodalInstance({required this.instance, required this.expectedOutputs});
+}
+
+class _MultimodalExpectedOutput {
+  final _MultimodalOutput output;
+  final Map<String, dynamic> metadata;
 
-  _MultimodalInstance({required this.instance, required this.expectedOutput});
+  _MultimodalExpectedOutput({required this.output, required this.metadata});
 }
 
 enum _MultimodalOutput { text, image, video }
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index 50694c74..3d8120ff 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -220,9 +220,80 @@ void main() {
         expect(requestBody['parameters'], {'dimension': 256});
         expect(response.result.embeddings, hasLength(1));
         expect(response.result.embeddings.first.embedding, [0.7, 0.8, 0.9]);
+        expect(response.result.embeddings.first.metadata, {
+          'documentIndex': 0,
+          'modality': 'text',
+          'partIndices': [0],
+        });
       },
     );
 
+    test('flattens mixed multimodal outputs with source metadata', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+      final req = EmbedRequest(
+        input: [
+          DocumentData(
+            content: [
+              TextPart(text: 'hello'),
+              MediaPart(media: Media(url: 'data:image/png;base64,AA==')),
+              MediaPart(media: Media(url: 'data:video/mp4;base64,AA==')),
+            ],
+          ),
+          DocumentData(content: [TextPart(text: 'world')]),
+        ],
+      );
+
+      final response = await embedder.run(req);
+
+      final requestBody =
+          jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+      final instances = requestBody['instances'] as List;
+      expect(instances, hasLength(2));
+      expect(instances.first, {
+        'text': 'hello',
+        'image': {'bytesBase64Encoded': 'AA==', 'mimeType': 'image/png'},
+        'video': {'bytesBase64Encoded': 'AA==', 'mimeType': 'video/mp4'},
+      });
+      expect(instances[1], {'text': 'world'});
+
+      final embeddings = response.result.embeddings;
+      expect(embeddings, hasLength(4));
+      expect(embeddings[0].embedding, [0.7, 0.8, 0.9]);
+      expect(embeddings[0].metadata, {
+        'documentIndex': 0,
+        'modality': 'text',
+        'partIndices': [0],
+      });
+      expect(embeddings[1].embedding, [1.7, 1.8, 1.9]);
+      expect(embeddings[1].metadata, {
+        'documentIndex': 0,
+        'modality': 'image',
+        'partIndex': 1,
+      });
+      expect(embeddings[2].embedding, [2.7, 2.8, 2.9]);
+      expect(embeddings[2].metadata, {
+        'documentIndex': 0,
+        'modality': 'video',
+        'partIndex': 2,
+        'segmentIndex': 0,
+        'startOffsetSec': 0,
+        'endOffsetSec': 16,
+      });
+      expect(embeddings[3].embedding, [1.7, 1.8, 1.9]);
+      expect(embeddings[3].metadata, {
+        'documentIndex': 1,
+        'modality': 'text',
+        'partIndices': [0],
+      });
+    });
+
     test(
       'throws a descriptive error when Vertex returns no predictions',
       () async {
diff --git a/packages/genkit_vertexai/test/test_http_client.dart b/packages/genkit_vertexai/test/test_http_client.dart
index ef4de098..38e7aa4d 100644
--- a/packages/genkit_vertexai/test/test_http_client.dart
+++ b/packages/genkit_vertexai/test/test_http_client.dart
@@ -100,14 +100,27 @@ class MockHttpClient extends http.BaseClient {
       final instances = body['instances'] as List;
       final predictions = returnEmptyPredictions
           ? const []
-          : List.generate(
-              instances.length,
-              (index) => returnMissingMultimodalEmbedding
-                  ? <String, dynamic>{}
-                  : {
-                      'textEmbedding': [index + 0.7, index + 0.8, index + 0.9],
+          : List.generate(instances.length, (index) {
+              if (returnMissingMultimodalEmbedding) {
+                return <String, dynamic>{};
+              }
+
+              final instance = instances[index] as Map<String, dynamic>;
+              return {
+                if (instance.containsKey('text'))
+                  'textEmbedding': [index + 0.7, index + 0.8, index + 0.9],
+                if (instance.containsKey('image'))
+                  'imageEmbedding': [index + 1.7, index + 1.8, index + 1.9],
+                if (instance.containsKey('video'))
+                  'videoEmbeddings': [
+                    {
+                      'embedding': [index + 2.7, index + 2.8, index + 2.9],
+                      'startOffsetSec': 0,
+                      'endOffsetSec': 16,
                     },
-            );
+                  ],
+              };
+            });
       return http.StreamedResponse(
         Stream.value(utf8.encode(jsonEncode({'predictions': predictions}))),
         200,

From 13c096fb20f442c810afde6c8d7206ac6cafadea Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 13:30:35 +0100
Subject: [PATCH 10/16] feat(genkit_google_genai): export additional generative
 language types for embedding requests

---
 packages/genkit_google_genai/lib/common.dart    |  7 +++++++
 packages/genkit_vertexai/lib/src/embedders.dart | 15 ++++++---------
 2 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/packages/genkit_google_genai/lib/common.dart b/packages/genkit_google_genai/lib/common.dart
index 4ca02b8b..64a73f57 100644
--- a/packages/genkit_google_genai/lib/common.dart
+++ b/packages/genkit_google_genai/lib/common.dart
@@ -17,4 +17,11 @@ library;
 
 export 'src/api_client.dart';
 export 'src/common_plugin.dart';
+export 'src/generated/generativelanguage.dart'
+    show
+        BatchEmbedContentsRequest,
+        Content,
+        ContentEmbedding,
+        EmbedContentRequest,
+        Part;
 export 'src/model.dart';
diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 3acd4843..2acee365 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -16,9 +16,6 @@ import 'dart:convert';
 
 import 'package:genkit/plugin.dart';
 import 'package:genkit_google_genai/common.dart' as google;
-// ignore: implementation_imports
-import 'package:genkit_google_genai/src/generated/generativelanguage.dart'
-    as google_types;
 
 List<ActionMetadata<dynamic, dynamic, dynamic, dynamic>> listVertexEmbedders({
   required String pluginName,
@@ -113,13 +110,13 @@ String _documentText(DocumentData doc) {
   return doc.content.where((p) => p.isText).map((p) => p.text).join('\n');
 }
 
-google_types.EmbedContentRequest _embedContentRequest(
+google.EmbedContentRequest _embedContentRequest(
   DocumentData doc,
   google.TextEmbedderOptions? options,
 ) {
   final text = _documentText(doc);
-  return google_types.EmbedContentRequest(
-    content: google_types.Content(parts: [google_types.Part(text: text)]),
+  return google.EmbedContentRequest(
+    content: google.Content(parts: [google.Part(text: text)]),
     outputDimensionality: options?.outputDimensionality,
     taskType: options?.taskType,
     title: options?.title,
@@ -144,7 +141,7 @@ Future<List<Embedding>> _runGeminiEmbeddingRequests({
   }
 
   final res = await service.batchEmbedContents(
-    google_types.BatchEmbedContentsRequest(
+    google.BatchEmbedContentsRequest(
       requests: docs.map((doc) => _embedContentRequest(doc, options)).toList(),
     ),
     model: 'models/$embedderName',
@@ -171,8 +168,8 @@ Future<Embedding> _runEmbedContentRequest({
   return Embedding(embedding: res.embedding?.values ?? []);
 }
 
-List<google_types.ContentEmbedding> _requireBatchEmbeddings(
-  List<google_types.ContentEmbedding>? embeddings, {
+List<google.ContentEmbedding> _requireBatchEmbeddings(
+  List<google.ContentEmbedding>? embeddings, {
   required int expectedCount,
 }) {
   if (embeddings == null || embeddings.isEmpty) {

From 9e3bb8f0486880414afdb3566989ebd5ee1f13b8 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 13:46:23 +0100
Subject: [PATCH 11/16] refactor(genkit_vertexai): clarify embedder request
 shape routing

---
 packages/genkit_vertexai/lib/src/embedders.dart | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 2acee365..82224658 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -503,13 +503,12 @@ String _baseModelName(String modelName) {
 
 _VertexEmbedderRequestShape _requestShapeFor(String modelName) {
   final baseModelName = _baseModelName(modelName);
-  // Check the broad model families in order.
+  final exactShape = _requestShapeByExactModel[baseModelName];
+  if (exactShape != null) return exactShape;
+
   if (_isMultimodalEmbeddingFamily(baseModelName)) {
     return _VertexEmbedderRequestShape.multimodalPredict;
   }
-  if (_usesLegacyGeminiPredictApi(baseModelName)) {
-    return _VertexEmbedderRequestShape.textPredict;
-  }
   if (_isGeminiEmbeddingFamily(baseModelName)) {
     return _VertexEmbedderRequestShape.geminiEmbedding;
   }
@@ -524,9 +523,9 @@ bool _isGeminiEmbeddingFamily(String modelName) {
   return modelName.startsWith('gemini-embedding-');
 }
 
-bool _usesLegacyGeminiPredictApi(String modelName) {
-  return modelName == 'gemini-embedding-001';
-}
+const _requestShapeByExactModel = {
+  'gemini-embedding-001': _VertexEmbedderRequestShape.textPredict,
+};
 
 class _MultimodalInstance {
   final Map<String, dynamic> instance;

From fb489ff03a03e931ef3ed259803f2c56770985b2 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 13:55:56 +0100
Subject: [PATCH 12/16] fix(genkit_vertexai): document multimodal dimension
 parameter

---
 packages/genkit_vertexai/lib/src/embedders.dart | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 82224658..a3e22471 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -228,6 +228,8 @@ Future<List<Embedding>> _runMultimodalPredictRequests({
   ];
   final parameters = <String, dynamic>{};
   if (options?.outputDimensionality != null) {
+    // Multimodal predict expects `parameters.dimension`, not
+    // `outputDimensionality`.
     parameters['dimension'] = options!.outputDimensionality;
   }
 

From 308cc8874d1b46a0399d9f12907f212c1a0bc035 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 13:59:08 +0100
Subject: [PATCH 13/16] fix(genkit_vertexai): handle invalid prediction
 payloads in text embeddings

---
 .../genkit_vertexai/lib/src/embedders.dart    | 13 +++++++--
 .../genkit_vertexai/test/embedders_test.dart  | 28 +++++++++++++++++++
 .../test/test_http_client.dart                | 14 ++++++----
 3 files changed, 48 insertions(+), 7 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index a3e22471..8e9b8c8c 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -287,8 +287,17 @@ Future<List<Embedding>> _runTextPredictRequests({
 }
 
 Embedding _textPredictionEmbedding(Map<String, dynamic> prediction) {
-  final embeddingData = prediction['embeddings'] as Map<String, dynamic>;
-  final values = embeddingData['values'] as List;
+  final embeddingData = prediction['embeddings'];
+  final values = embeddingData is Map<String, dynamic>
+      ? embeddingData['values']
+      : null;
+  if (values is! List) {
+    throw GenkitException(
+      'Vertex AI returned an invalid prediction payload.',
+      status: StatusCodes.INTERNAL,
+    );
+  }
+
   return Embedding(
     embedding: values.map((value) => (value as num).toDouble()).toList(),
   );
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index 3d8120ff..9b030d0f 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -184,6 +184,34 @@ void main() {
       expect(requestBody['parameters'], {'outputDimensionality': 256});
     });
 
+    test('throws when a text prediction omits embedding values', () async {
+      final mockClient = MockHttpClient(returnInvalidTextPrediction: true);
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'text-embedding-005');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(content: [TextPart(text: 'hello')]),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains('Vertex AI returned an invalid prediction payload.'),
+          ),
+        ),
+      );
+    });
+
     test(
       'uses multimodal predict schema for text-only multimodal inputs',
       () async {
diff --git a/packages/genkit_vertexai/test/test_http_client.dart b/packages/genkit_vertexai/test/test_http_client.dart
index 38e7aa4d..ee040a64 100644
--- a/packages/genkit_vertexai/test/test_http_client.dart
+++ b/packages/genkit_vertexai/test/test_http_client.dart
@@ -19,10 +19,12 @@ import 'package:http/http.dart' as http;
 class MockHttpClient extends http.BaseClient {
   MockHttpClient({
     this.returnEmptyPredictions = false,
+    this.returnInvalidTextPrediction = false,
     this.returnMissingMultimodalEmbedding = false,
   });
 
   final bool returnEmptyPredictions;
+  final bool returnInvalidTextPrediction;
   final bool returnMissingMultimodalEmbedding;
   final List<Uri> requestUrls = [];
   final List<String> requestBodies = [];
@@ -134,11 +136,13 @@ class MockHttpClient extends http.BaseClient {
           ? const []
           : List.generate(
               instances.length,
-              (index) => {
-                'embeddings': {
-                  'values': [index + 0.4, index + 0.5, index + 0.6],
-                },
-              },
+              (index) => returnInvalidTextPrediction
+                  ? <String, dynamic>{'embeddings': <String, dynamic>{}}
+                  : {
+                      'embeddings': {
+                        'values': [index + 0.4, index + 0.5, index + 0.6],
+                      },
+                    },
             );
       return http.StreamedResponse(
         Stream.value(utf8.encode(jsonEncode({'predictions': predictions}))),

From 0488245d518ea7f42199d42d7e3f20de8d8a7790 Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 14:02:40 +0100
Subject: [PATCH 14/16] fix(genkit_vertexai): safely parse multimodal data URIs

---
 .../genkit_vertexai/lib/src/embedders.dart    |  4 +-
 .../genkit_vertexai/test/embedders_test.dart  | 73 +++++++++++++++++++
 2 files changed, 75 insertions(+), 2 deletions(-)

diff --git a/packages/genkit_vertexai/lib/src/embedders.dart b/packages/genkit_vertexai/lib/src/embedders.dart
index 8e9b8c8c..70b71e2e 100644
--- a/packages/genkit_vertexai/lib/src/embedders.dart
+++ b/packages/genkit_vertexai/lib/src/embedders.dart
@@ -376,7 +376,7 @@ MapEntry<String, Map<String, dynamic>> _toMultimodalMediaField(Media media) {
 
   // Convert the media input into the format Vertex expects.
   if (media.url.startsWith('data:')) {
-    final data = Uri.parse(media.url).data;
+    final data = Uri.tryParse(media.url)?.data;
     if (data == null) {
       throw GenkitException(
         'Vertex multimodalembedding media inputs require a valid data URI.',
@@ -409,7 +409,7 @@ String? _mediaMimeType(Media media) {
   }
 
   if (media.url.startsWith('data:')) {
-    return Uri.parse(media.url).data?.mimeType;
+    return Uri.tryParse(media.url)?.data?.mimeType;
   }
 
   return null;
diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index 9b030d0f..e9ce0196 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -322,6 +322,79 @@ void main() {
       });
     });
 
+    test('throws when a multimodal data URI is malformed', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(
+                content: [
+                  MediaPart(
+                    media: Media(
+                      url: 'data:image/png;base64,%',
+                      contentType: 'image/png',
+                    ),
+                  ),
+                ],
+              ),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains(
+              'Vertex multimodalembedding media inputs require a valid data URI.',
+            ),
+          ),
+        ),
+      );
+    });
+
+    test('throws when a multimodal data URI MIME cannot be parsed', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(
+                content: [
+                  MediaPart(media: Media(url: 'data:image/png;base64,%')),
+                ],
+              ),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains(
+              'Vertex multimodalembedding media inputs require a MIME type.',
+            ),
+          ),
+        ),
+      );
+    });
+
     test(
       'throws a descriptive error when Vertex returns no predictions',
       () async {

From 3d061e600cce957ae2248118ea7a60f4696e9dad Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 14:32:52 +0100
Subject: [PATCH 15/16] feat(genkit_vertexai): add tests for multimodal image
 input handling and error cases

---
 .../genkit_vertexai/test/embedders_test.dart  | 171 ++++++++++++++++++
 1 file changed, 171 insertions(+)

diff --git a/packages/genkit_vertexai/test/embedders_test.dart b/packages/genkit_vertexai/test/embedders_test.dart
index e9ce0196..9f0d2c9f 100644
--- a/packages/genkit_vertexai/test/embedders_test.dart
+++ b/packages/genkit_vertexai/test/embedders_test.dart
@@ -322,6 +322,74 @@ void main() {
       });
     });
 
+    test('uses data URI image inputs in multimodal predict requests', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await embedder.run(
+        EmbedRequest(
+          input: [
+            DocumentData(
+              content: [
+                MediaPart(media: Media(url: 'data:image/png;base64,AA==')),
+              ],
+            ),
+          ],
+        ),
+      );
+
+      final requestBody =
+          jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+      final instances = requestBody['instances'] as List;
+      expect(instances.single, {
+        'image': {'bytesBase64Encoded': 'AA==', 'mimeType': 'image/png'},
+      });
+    });
+
+    test('uses gs image inputs in multimodal predict requests', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await embedder.run(
+        EmbedRequest(
+          input: [
+            DocumentData(
+              content: [
+                MediaPart(
+                  media: Media(
+                    url: 'gs://my-bucket/image.png',
+                    contentType: 'image/png',
+                  ),
+                ),
+              ],
+            ),
+          ],
+        ),
+      );
+
+      final requestBody =
+          jsonDecode(mockClient.lastBody!) as Map<String, dynamic>;
+      final instances = requestBody['instances'] as List;
+      expect(instances.single, {
+        'image': {
+          'gcsUri': 'gs://my-bucket/image.png',
+          'mimeType': 'image/png',
+        },
+      });
+    });
+
     test('throws when a multimodal data URI is malformed', () async {
       final mockClient = MockHttpClient();
       final plugin = VertexAiPluginImpl(
@@ -395,6 +463,109 @@ void main() {
       );
     });
 
+    test('throws when a multimodal document has multiple images', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(
+                content: [
+                  MediaPart(media: Media(url: 'data:image/png;base64,AA==')),
+                  MediaPart(media: Media(url: 'data:image/jpeg;base64,AA==')),
+                ],
+              ),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains(
+              'Vertex multimodalembedding supports at most one image part per input document.',
+            ),
+          ),
+        ),
+      );
+    });
+
+    test('throws when multimodal media has an unsupported MIME type', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(
+                content: [
+                  MediaPart(media: Media(url: 'data:audio/wav;base64,AA==')),
+                ],
+              ),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains(
+              'Unsupported Vertex multimodalembedding media MIME type: audio/wav',
+            ),
+          ),
+        ),
+      );
+    });
+
+    test('throws when multimodal media is missing a MIME type', () async {
+      final mockClient = MockHttpClient();
+      final plugin = VertexAiPluginImpl(
+        projectId: 'my-project',
+        location: 'us-central1',
+        authClient: mockClient,
+      );
+
+      final embedder = _resolveEmbedder(plugin, 'multimodalembedding');
+
+      await expectLater(
+        () => embedder.run(
+          EmbedRequest(
+            input: [
+              DocumentData(
+                content: [
+                  MediaPart(media: Media(url: 'gs://my-bucket/image.png')),
+                ],
+              ),
+            ],
+          ),
+        ),
+        throwsA(
+          isA<GenkitException>().having(
+            (error) => error.message,
+            'message',
+            contains(
+              'Vertex multimodalembedding media inputs require a MIME type.',
+            ),
+          ),
+        ),
+      );
+    });
+
     test(
       'throws a descriptive error when Vertex returns no predictions',
       () async {

From 4256c027aad72f76fe8a216f1da8e4cb2c0144cb Mon Sep 17 00:00:00 2001
From: Corie Watson <watson.corie@gmail.com>
Date: Tue, 12 May 2026 14:44:12 +0100
Subject: [PATCH 16/16] fix(generativelanguage): update documentation for soft
 tokens tensor frame shape format

---
 .../lib/src/generated/generativelanguage.dart                   | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/packages/genkit_google_genai/lib/src/generated/generativelanguage.dart b/packages/genkit_google_genai/lib/src/generated/generativelanguage.dart
index 6ea92a04..b84ac2d2 100644
--- a/packages/genkit_google_genai/lib/src/generated/generativelanguage.dart
+++ b/packages/genkit_google_genai/lib/src/generated/generativelanguage.dart
@@ -1769,7 +1769,7 @@ extension type ContentEmbedding._(Map<String, Object?> _data) {
 
   set values(List<double>? value) => _data['values'] = value;
 
-  /// This field stores the soft tokens tensor frame shape (e.g. [1, 1, 256, 2048]).
+  /// This field stores the soft tokens tensor frame shape (e.g. `[1, 1, 256, 2048]`).
   List<int>? get shape {
     final v = _data['shape'];
     if (v == null) return null;