fix(vertex_ai): keep all embeddings in multimodal embedding response

An instance that carries both `text` and `image` makes Vertex return
`textEmbedding` and `imageEmbedding` in the same prediction, but
transform_embedding_response_to_openai chained the branches with `elif`,
so only the first match was emitted and every other embedding was
silently dropped.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

litellm/llms/vertex_ai/multimodal_embeddings/transformation.py

@@ -287,14 +287,14 @@ class VertexAIMultimodalEmbeddingConfig(BaseEmbeddingConfig):
                             object="embedding",
                         )
                         openai_embeddings.append(openai_embedding_object)
-                    elif "imageEmbedding" in _prediction:
+                    if "imageEmbedding" in _prediction:
                         openai_embedding_object = Embedding(
                             embedding=_prediction["imageEmbedding"],
                             index=idx,
                             object="embedding",
                         )
                         openai_embeddings.append(openai_embedding_object)
-                    elif "videoEmbeddings" in _prediction:
+                    if "videoEmbeddings" in _prediction:
                         for video_embedding in _prediction["videoEmbeddings"]:
                             openai_embedding_object = Embedding(
                                 embedding=video_embedding["embedding"],

tests/local_testing/test_amazing_vertex_completion.py

@@ -1880,8 +1880,9 @@ async def test_vertexai_multimodal_embedding():
 
         assert args_to_vertexai == expected_payload
         assert response.model == "multimodalembedding@001"
-        assert len(response.data) == 1
-        response_data = response.data[0]
+        assert len(response.data) == 2
+        assert response.data[0]["embedding"] == [0.4, 0.5, 0.6]
+        assert response.data[1]["embedding"] == [0.1, 0.2, 0.3]
 
         # Optional: Print for debugging
         print("Arguments passed to Vertex AI:", args_to_vertexai)

tests/test_litellm/llms/vertex_ai/multimodal_embeddings/test_vertex_ai_multimodal_embedding_transformation.py

@@ -179,3 +179,49 @@ class TestVertexMultimodalEmbedding:
             headers={},
         )
         assert "parameters" not in request
+
+    def test_response_keeps_both_text_and_image_embeddings(self):
+        predictions = {
+            "predictions": [
+                {
+                    "textEmbedding": [0.4, 0.5, 0.6],
+                    "imageEmbedding": [0.1, 0.2, 0.3],
+                }
+            ]
+        }
+        result = self.config.transform_embedding_response_to_openai(predictions)
+        assert [e["embedding"] for e in result] == [[0.4, 0.5, 0.6], [0.1, 0.2, 0.3]]
+        assert [e["index"] for e in result] == [0, 0]
+
+    def test_response_keeps_text_and_video_embeddings(self):
+        predictions = {
+            "predictions": [
+                {
+                    "textEmbedding": [0.4, 0.5, 0.6],
+                    "videoEmbeddings": [
+                        {
+                            "startOffsetSec": 0,
+                            "endOffsetSec": 5,
+                            "embedding": [0.7, 0.8, 0.9],
+                        }
+                    ],
+                }
+            ]
+        }
+        result = self.config.transform_embedding_response_to_openai(predictions)
+        assert [e["embedding"] for e in result] == [[0.4, 0.5, 0.6], [0.7, 0.8, 0.9]]
+
+    @pytest.mark.parametrize(
+        "prediction, expected",
+        [
+            ({"textEmbedding": [0.1, 0.2]}, [[0.1, 0.2]]),
+            ({"imageEmbedding": [0.3, 0.4]}, [[0.3, 0.4]]),
+            (
+                {"videoEmbeddings": [{"startOffsetSec": 0, "endOffsetSec": 5, "embedding": [0.5]}]},
+                [[0.5]],
+            ),
+        ],
+    )
+    def test_response_single_modality_unchanged(self, prediction, expected):
+        result = self.config.transform_embedding_response_to_openai({"predictions": [prediction]})
+        assert [e["embedding"] for e in result] == expected