fix(llmobs): handle multimodal content parts in Chat Completions input messages

ddtrace/llmobs/_integrations/utils.py

@@ -327,7 +327,11 @@ def openai_set_meta_tags_from_chat(
     """Extract prompt/response tags from a chat completion and set them as temporary "_ml_obs.meta.*" tags."""
     input_messages: list[Message] = []
     for m in kwargs.get("messages", []):
-        content = str(_get_attr(m, "content", ""))
+        raw_content = _get_attr(m, "content", "")
+        if isinstance(raw_content, list):
+            content = _extract_chat_content_parts(raw_content)
+        else:
+            content = str(raw_content) if raw_content else ""
         role = str(_get_attr(m, "role", ""))
         processed_message: Message = Message(content=content, role=role)
         tool_call_id = _get_attr(m, "tool_call_id", None)
@@ -821,6 +825,28 @@ def _extract_content_item_text(content_item: Any) -> str:
     return ""
 
 
+def _extract_chat_content_parts(content_parts: list) -> str:
+    """Extract text from OpenAI Chat API multimodal content parts.
+
+    Handles content part types used by the Chat Completions API:
+    - {"type": "text", "text": "..."}
+    - {"type": "image_url", "image_url": {"url": "..."}}
+    - {"type": "input_audio", ...}
+    """
+    texts = []
+    for part in content_parts:
+        part_type = _get_attr(part, "type", "")
+        if part_type == "text":
+            text = _get_attr(part, "text", "")
+            if text:
+                texts.append(str(text))
+        elif part_type == "image_url":
+            texts.append(IMAGE_FALLBACK_MARKER)
+        elif part_type == "input_audio":
+            texts.append("[audio]")
+    return "\n".join(texts)
+
+
 def _normalize_prompt_variables(variables: dict[str, Any]) -> dict[str, Any]:
     """Converts OpenAI SDK response objects or dicts into simple key-value pairs.
 

releasenotes/notes/fix-multimodal-chat-content-cf726c0b53dbf68f.yaml

@@ -0,0 +1,3 @@
+fixes:
+  - |
+    LLM Observability: Fix handling of multimodal content parts in Chat Completions input messages. Messages with list-type content (e.g. text + image_url parts) are now correctly extracted instead of being serialized as a raw list string.

tests/llmobs/test_integrations_utils.py

@@ -1,3 +1,4 @@
+from ddtrace.llmobs._integrations.utils import _extract_chat_content_parts
 from ddtrace.llmobs._integrations.utils import _extract_chat_template_from_instructions
 from ddtrace.llmobs._integrations.utils import _normalize_prompt_variables
 
@@ -163,6 +164,58 @@ def __init__(self, file_url=None, file_id=None, filename=None, file_data=None):
     assert result["file_fallback"] == "[file]"
 
 
+class TestExtractChatContentParts:
+    def test_text_only(self):
+        parts = [{"type": "text", "text": "Hello world"}]
+        assert _extract_chat_content_parts(parts) == "Hello world"
+
+    def test_multiple_text_parts(self):
+        parts = [
+            {"type": "text", "text": "First part"},
+            {"type": "text", "text": "Second part"},
+        ]
+        assert _extract_chat_content_parts(parts) == "First part\nSecond part"
+
+    def test_text_and_image_url(self):
+        parts = [
+            {"type": "text", "text": "Describe this image"},
+            {"type": "image_url", "image_url": {"url": "https://example.com/img.png"}},
+        ]
+        assert _extract_chat_content_parts(parts) == "Describe this image\n[image]"
+
+    def test_image_url_only(self):
+        parts = [{"type": "image_url", "image_url": {"url": "https://example.com/img.png"}}]
+        assert _extract_chat_content_parts(parts) == "[image]"
+
+    def test_input_audio(self):
+        parts = [
+            {"type": "text", "text": "Transcribe this"},
+            {"type": "input_audio", "input_audio": {"data": "base64data", "format": "wav"}},
+        ]
+        assert _extract_chat_content_parts(parts) == "Transcribe this\n[audio]"
+
+    def test_empty_list(self):
+        assert _extract_chat_content_parts([]) == ""
+
+    def test_empty_text(self):
+        parts = [{"type": "text", "text": ""}]
+        assert _extract_chat_content_parts(parts) == ""
+
+    def test_pydantic_like_objects(self):
+        """Simulate Pydantic model objects with attribute access (like OpenAI SDK models)."""
+
+        class ContentPart:
+            def __init__(self, **kwargs):
+                for k, v in kwargs.items():
+                    setattr(self, k, v)
+
+        parts = [
+            ContentPart(type="text", text="Return an integer up to 10."),
+            ContentPart(type="image_url", image_url=ContentPart(url="https://example.com/img.png")),
+        ]
+        assert _extract_chat_content_parts(parts) == "Return an integer up to 10.\n[image]"
+
+
 def test_extract_chat_template_with_falsy_values():
     """Test that falsy but valid values (0, False) are preserved in template extraction."""