Merge 6de68a64bb into 527d882508

litellm/integrations/datadog/datadog_llm_obs.py

@@ -663,6 +663,15 @@ class DataDogLLMObsLogger(CustomBatchLogger):
             redact_prompt_text=redact_payload,
         )
 
+        # Datadog prices llm/embedding spans from top-level meta.model_name /
+        # model_provider. Those values already live in metadata; lift them so
+        # embedding spans are not flagged "Partial cost — Unsupported model provider".
+        # Narrow via isinstance so reportUnknownVariableType stays flat vs
+        # indexing a payload that basedpyright still treats as partially unknown.
+        _meta_model: Final[object] = payload_metadata.get("model_name", "unknown")
+        _meta_provider: Final[object] = payload_metadata.get("model_provider", "unknown")
+        model_name: Final[str] = _meta_model if isinstance(_meta_model, str) else "unknown"
+        model_provider: Final[str] = _meta_provider if isinstance(_meta_provider, str) else "unknown"
         meta: Final[Meta] = {
             "kind": span_kind,
             "input": input_meta,
@@ -670,6 +679,11 @@ class DataDogLLMObsLogger(CustomBatchLogger):
             "metadata": payload_metadata,
             "error": error_info,
             **({"tool_definitions": tool_definitions} if tool_definitions else {}),
+            **(
+                {"model_name": model_name, "model_provider": model_provider}
+                if span_kind in ("llm", "embedding")
+                else {}
+            ),
         }
 
         metrics: Final = self._assemble_metrics(standard_logging_payload, tool_output_tokens)

litellm/types/integrations/datadog_llm_obs.py

@@ -71,6 +71,10 @@ class DDLLMObsError(TypedDict, total=False):
 class Meta(TypedDict, total=False):
     # The span kind: "agent", "workflow", "llm", "tool", "task", "embedding", or "retrieval".
     kind: Literal["llm", "tool", "task", "embedding", "retrieval"]
+    # Top-level model fields required for Datadog cost attribution on llm/embedding spans.
+    # Values are also kept under metadata for backwards compatibility.
+    model_name: ReadOnly[str]
+    model_provider: ReadOnly[str]
     input: InputMeta  # The span's input information.
     output: OutputMeta  # The span's output information.
     metadata: ReadOnly[dict[str, object]]

tests/unit/integrations/datadog/test_datadog_llm_obs.py

@@ -1141,6 +1141,39 @@ def test_reasoning_content_survives_the_mapping(logger: DataDogLLMObsLogger) ->
     assert payload["meta"]["output"]["messages"][0]["reasoning_content"] == "thinking"
 
 
+def test_embedding_span_sets_top_level_model_name(logger: DataDogLLMObsLogger) -> None:
+    """Datadog prices embedding spans from meta.model_name, not meta.metadata (issue #41601)."""
+    kwargs = build_payload(messages=None)
+    kwargs["standard_logging_object"]["call_type"] = "aembedding"
+    kwargs["standard_logging_object"]["model"] = "openai/text-embedding-3-small"
+    kwargs["standard_logging_object"]["custom_llm_provider"] = "openai"
+    kwargs["standard_logging_object"]["response"] = {"data": [{"embedding": [0.1, 0.2]}]}
+    kwargs["litellm_params"]["metadata"]["parent_id"] = "parent-span-1"
+
+    start = datetime(2026, 9, 1, 12, 0, 0)
+    span = json.loads(safe_dumps(logger.create_llm_obs_payload(kwargs, start, start + timedelta(seconds=1))))
+
+    assert span["meta"]["kind"] == "embedding"
+    assert span["meta"]["model_name"] == "openai/text-embedding-3-small"
+    assert span["meta"]["model_provider"] == "openai"
+    # Still present under metadata for backwards compatibility.
+    assert span["meta"]["metadata"]["model_name"] == "openai/text-embedding-3-small"
+    assert span["meta"]["metadata"]["model_provider"] == "openai"
+
+
+def test_llm_span_sets_top_level_model_name(logger: DataDogLLMObsLogger) -> None:
+    """llm spans must keep top-level model_name/model_provider for Datadog cost attribution."""
+    kwargs = build_payload()
+    kwargs["standard_logging_object"]["model"] = "openai/gpt-4.1-mini"
+    kwargs["standard_logging_object"]["custom_llm_provider"] = "openai"
+
+    start = datetime(2026, 9, 1, 12, 0, 0)
+    span = json.loads(safe_dumps(logger.create_llm_obs_payload(kwargs, start, start + timedelta(seconds=1))))
+
+    assert span["meta"]["kind"] == "llm"
+    assert span["meta"]["model_name"] == "openai/gpt-4.1-mini"
+    assert span["meta"]["model_provider"] == "openai"
+
 @pytest.mark.parametrize(
     ("raw_arguments", "shipped"),
     [