fix(openai): count base64 embedding vector dimensions

langfuse/openai.py

@@ -23,6 +23,8 @@
 See docs for more details: https://langfuse.com/docs/integrations/openai
 """
 
+import base64
+import binascii
 import json
 import os
 import types
@@ -963,8 +965,21 @@ def _get_langfuse_data_from_default_response(
                 if hasattr(first_embedding, "embedding")
                 else first_embedding.get("embedding", [])
             )
+            if isinstance(embedding_vector, str):
+                try:
+                    # Base64 embeddings contain one float32 per dimension.
+                    decoded = base64.b64decode(embedding_vector, validate=True)
+                    dimensions = (
+                        len(decoded) // 4
+                        if len(decoded) % 4 == 0
+                        else len(embedding_vector)
+                    )
+                except (binascii.Error, ValueError):
+                    dimensions = len(embedding_vector)
+            else:
+                dimensions = len(embedding_vector) if embedding_vector else 0
             completion = {
-                "dimensions": len(embedding_vector) if embedding_vector else 0,
+                "dimensions": dimensions,
                 "count": len(data),
             }
 

tests/unit/test_openai.py

@@ -3,6 +3,7 @@
 from types import SimpleNamespace
 from unittest.mock import patch
 
+import httpx
 import pytest
 from openai.types.chat import ChatCompletion
 from openai.types.responses import ParsedResponseOutputMessage, ParsedResponseOutputText
@@ -1015,6 +1016,55 @@ def test_embedding_exports_dimensions_and_count(
     }
 
 
+@pytest.mark.parametrize(
+    "encoded_embedding, dimensions",
+    [
+        ("AAAAAA==", 1),
+        ("AAAAAAAAAAA=", 2),
+        ("AAAAAAAAAAAAAAAA", 3),
+        pytest.param("!invalid!", 9, id="invalid-base64"),
+        pytest.param("!!!!", 4, id="invalid-base64-characters"),
+        pytest.param("AAAA", 4, id="incomplete-float32"),
+        pytest.param("\u2603", 1, id="non-ascii"),
+    ],
+)
+def test_base64_embedding_exports_vector_dimensions(
+    langfuse_memory_client, get_span, json_attr, encoded_embedding, dimensions
+):
+
+    def respond(request):
+        return httpx.Response(
+            200,
+            json={
+                "model": "text-embedding-3-small",
+                "object": "list",
+                "data": [
+                    {"index": 0, "object": "embedding", "embedding": encoded_embedding}
+                ],
+                "usage": {"prompt_tokens": 2, "total_tokens": 2},
+            },
+        )
+
+    with lf_openai.OpenAI(
+        api_key="test",
+        http_client=httpx.Client(transport=httpx.MockTransport(respond)),
+    ) as openai_client:
+        result = openai_client.embeddings.create(
+            name="unit-openai-base64-embedding",
+            model="text-embedding-3-small",
+            input="hello world",
+            encoding_format="base64",
+        )
+
+    assert result.data[0].embedding == encoded_embedding
+    langfuse_memory_client.flush()
+    span = get_span("unit-openai-base64-embedding")
+    assert json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_OUTPUT) == {
+        "dimensions": dimensions,
+        "count": 1,
+    }
+
+
 def _make_chat_response(**extra_response_fields):
     return SimpleNamespace(
         model="gpt-4o-mini",