fix: keep embedding string batches separate

google/genai/models.py

@@ -1226,6 +1226,16 @@ def _FunctionCallingConfig_to_mldev(
   return to_object
 
 
+def _is_text_embedding_batch(
+    contents: Union[types.ContentListUnion, types.ContentListUnionDict],
+) -> bool:
+  return (
+      isinstance(contents, list)
+      and bool(contents)
+      and all(isinstance(content, str) for content in contents)
+  )
+
+
 def _GenerateContentConfig_to_mldev(
     api_client: BaseApiClient,
     from_object: Union[dict[str, Any], object],
@@ -6335,7 +6345,9 @@ def embed_content(
       )
     """
     if not self._api_client.vertexai:
-      if 'gemini-embedding-2' in model:
+      if 'gemini-embedding-2' in model and not _is_text_embedding_batch(
+          contents
+      ):
         contents = t.t_contents(contents)  # type: ignore[assignment]
       return self._embed_content(model=model, contents=contents, config=config)
 
@@ -9296,7 +9308,9 @@ async def embed_content(
       )
     """
     if not self._api_client.vertexai:
-      if 'gemini-embedding-2' in model:
+      if 'gemini-embedding-2' in model and not _is_text_embedding_batch(
+          contents
+      ):
         contents = t.t_contents(contents)  # type: ignore[assignment]
       return await self._embed_content(
           model=model, contents=contents, config=config

google/genai/tests/models/test_embed_content.py

@@ -21,6 +21,7 @@
 import pytest
 
 from ... import _transformers as t
+from ... import models
 from ... import types
 from .. import pytest_helper
 
@@ -227,6 +228,78 @@ def _get_bytes_from_file(relative_path: str) -> bytes:
 )
 
 
+class _FakeApiClient:
+  vertexai = False
+
+
+def test_gemini_embedding_2_text_list_stays_batched(
+    monkeypatch, use_vertex, replays_prefix, http_options
+):
+  module = models.Models(_FakeApiClient())
+  captured = {}
+
+  def fake_embed_content(**kwargs):
+    captured.update(kwargs)
+    return types.EmbedContentResponse()
+
+  monkeypatch.setattr(module, '_embed_content', fake_embed_content)
+
+  module.embed_content(
+      model='gemini-embedding-2-preview',
+      contents=['first text', 'second text'],
+  )
+
+  assert captured['contents'] == ['first text', 'second text']
+
+
+def test_gemini_embedding_2_mixed_content_still_combines_parts(
+    monkeypatch, use_vertex, replays_prefix, http_options
+):
+  module = models.Models(_FakeApiClient())
+  captured = {}
+
+  def fake_embed_content(**kwargs):
+    captured.update(kwargs)
+    return types.EmbedContentResponse()
+
+  monkeypatch.setattr(module, '_embed_content', fake_embed_content)
+
+  module.embed_content(
+      model='gemini-embedding-2-preview',
+      contents=[
+          'Similar things to the following image:',
+          types.Part.from_uri(
+              file_uri='gs://generativeai-downloads/images/scones.jpg',
+              mime_type='image/jpeg',
+          ),
+      ],
+  )
+
+  assert len(captured['contents']) == 1
+  assert len(captured['contents'][0].parts) == 2
+
+
+@pytest.mark.asyncio
+async def test_async_gemini_embedding_2_text_list_stays_batched(
+    monkeypatch, use_vertex, replays_prefix, http_options
+):
+  module = models.AsyncModels(_FakeApiClient())
+  captured = {}
+
+  async def fake_embed_content(**kwargs):
+    captured.update(kwargs)
+    return types.EmbedContentResponse()
+
+  monkeypatch.setattr(module, '_embed_content', fake_embed_content)
+
+  await module.embed_content(
+      model='gemini-embedding-2-preview',
+      contents=['first text', 'second text'],
+  )
+
+  assert captured['contents'] == ['first text', 'second text']
+
+
 def test_gemini_embedding_2_content_combination(client):
   response = client.models.embed_content(
       model='gemini-embedding-2-preview',