chore(litellm): CI Failures For LiteLLM

python/instrumentation/openinference-instrumentation-litellm/src/openinference/instrumentation/litellm/__init__.py

@@ -334,23 +334,33 @@ def _finalize_span(span: trace_api.Span, result: Any) -> None:
         if model_name := getattr(result, "model", None):
             _set_span_attribute(span, SpanAttributes.EMBEDDING_MODEL_NAME, model_name)
 
-        if result_data := result.data:
-            # Extract embedding vectors directly (litellm uses enumeration, not explicit index)
-            for index, embedding_item in enumerate(result_data):
-                # LiteLLM returns dicts with 'embedding' key
-                raw_vector = (
-                    embedding_item.get("embedding") if hasattr(embedding_item, "get") else None
-                )
-                if not raw_vector:
+        if result_data := getattr(result, "data", None):
+            # Extract embedding vectors directly
+            for embedding_item in result_data:
+                # LiteLLM may return embedding items as dicts or OpenAIObject instances
+                if isinstance(embedding_item, dict):
+                    raw_vector = embedding_item.get("embedding")
+                    index = embedding_item.get("index")
+                else:
+                    raw_vector = getattr(embedding_item, "embedding", None)
+                    index = getattr(embedding_item, "index", None)
+
+                # Skip entries without a usable vector or explicit index
+                if raw_vector is None or index is None:
                     continue
 
-                vector = None
-                # Check if it's a list of floats
-                if isinstance(raw_vector, (list, tuple)) and raw_vector:
+                vector: Union[Tuple[Any, ...], str, None] = None
+                # Record numeric vectors as tuples
+                if isinstance(raw_vector, (list, tuple)):
                     if all(isinstance(x, (int, float)) for x in raw_vector):
                         vector = tuple(raw_vector)
+                # Record base64-encoded vectors directly
+                elif isinstance(raw_vector, str):
+                    vector = raw_vector
+                else:
+                    continue
 
-                if vector:
+                if vector is not None:
                     _set_span_attribute(
                         span,
                         f"{SpanAttributes.EMBEDDING_EMBEDDINGS}.{index}.{EmbeddingAttributes.EMBEDDING_VECTOR}",

python/instrumentation/openinference-instrumentation-litellm/tests/test_embeddings.py

@@ -61,7 +61,7 @@ def test_batch_embedding(
             f"{SpanAttributes.EMBEDDING_EMBEDDINGS}.{i}.{EmbeddingAttributes.EMBEDDING_VECTOR}"
         )
         assert vector is not None
-        assert isinstance(vector, tuple)
+        assert isinstance(vector, (tuple, str))
         assert len(vector) > 0
 
     # Check token counts
@@ -126,7 +126,7 @@ def test_single_string_embedding(
         f"{SpanAttributes.EMBEDDING_EMBEDDINGS}.0.{EmbeddingAttributes.EMBEDDING_VECTOR}"
     )
     assert vector is not None
-    assert isinstance(vector, tuple)
+    assert isinstance(vector, (tuple, str))
     assert len(vector) > 0
 
     # Check token counts
@@ -194,7 +194,7 @@ def test_batch_embedding_with_different_model(
             f"{SpanAttributes.EMBEDDING_EMBEDDINGS}.{i}.{EmbeddingAttributes.EMBEDDING_VECTOR}"
         )
         assert vector is not None
-        assert isinstance(vector, tuple)
+        assert isinstance(vector, (tuple, str))
         assert len(vector) > 0
 
     # Check token counts

python/instrumentation/openinference-instrumentation-litellm/tests/test_token_counts.py

@@ -1,5 +1,6 @@
 import json
-from typing import Iterator
+from typing import Generator, Iterator
+from unittest.mock import MagicMock, patch
 
 import litellm
 import pytest
@@ -18,6 +19,17 @@ def instrument(
     yield
 
 
+@pytest.fixture
+def patch_tiktoken_encoding() -> Generator[None, None, None]:
+    """Patch `tiktoken.get_encoding` for LiteLLM to avoid network calls."""
+
+    with patch("tiktoken.get_encoding") as mock_get_encoding:
+        mock_encoding = MagicMock()
+        mock_encoding.encode.return_value = [1, 2, 3]
+        mock_get_encoding.return_value = mock_encoding
+        yield
+
+
 class TestTokenCounts:
     @pytest.mark.vcr(
         decode_compressed_response=True,
@@ -79,6 +91,7 @@ def test_openai(
     def test_anthropic(
         self,
         in_memory_span_exporter: InMemorySpanExporter,
+        patch_tiktoken_encoding: None,
     ) -> None:
         messages = [{"role": "user", "content": "Hello!"}]
         resp = litellm.completion(