fix(cost): detect file_citation annotations moved to provider_specific_fields

File search cost detection only looked at message.annotations, but
partition_annotations now moves every non-url_citation annotation into
message.provider_specific_fields["annotations"]. Chat completions that
cited a file were therefore no longer billed for file search.

Fixes #49126

litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py

@@ -618,6 +618,18 @@ class StandardBuiltInToolCostTracking:
                     for annotation in annotations:
                         if annotation.get("type", None) == annotation_type:
                             return True
+                # Non-url_citation annotations (e.g. file_citation) are moved to
+                # provider_specific_fields["annotations"] by partition_annotations.
+                provider_specific_fields: object = getattr(message, "provider_specific_fields", None)
+                if isinstance(provider_specific_fields, dict):
+                    provider_annotations: object = provider_specific_fields.get("annotations")
+                    if isinstance(provider_annotations, list):
+                        for provider_annotation in provider_annotations:
+                            if (
+                                isinstance(provider_annotation, dict)
+                                and provider_annotation.get("type") == annotation_type
+                            ):
+                                return True
         return False
 
     @staticmethod

tests/unit/litellm_core_utils/llm_cost_calc/test_file_search_annotation_detection.py

@@ -0,0 +1,105 @@
+"""
+File search detection must find file_citation annotations after they are
+moved to message.provider_specific_fields["annotations"].
+
+See https://github.com/BerriAI/litellm/issues/49126.
+"""
+
+from typing import Final
+
+import litellm
+from litellm.constants import OPENAI_FILE_SEARCH_COST_PER_1K_CALLS
+from litellm.litellm_core_utils.llm_cost_calc.tool_call_cost_tracking import (
+    StandardBuiltInToolCostTracking,
+)
+from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+    convert_to_model_response_object,
+)
+from litellm.types.utils import ModelResponse
+
+_FILE_CITATION: Final = {
+    "type": "file_citation",
+    "file_id": "file-abc123",
+    "filename": "doc.pdf",
+    "index": 0,
+}
+
+
+def _streamed_response(annotations: tuple[object, ...]) -> ModelResponse:
+    chunks: Final = [
+        litellm.ModelResponseStream(
+            id="chatcmpl-1",
+            model="gpt-4o",
+            choices=[{"index": 0, "delta": {"role": "assistant", "content": "See the doc."}}],
+        ),
+        litellm.ModelResponseStream(
+            id="chatcmpl-1",
+            model="gpt-4o",
+            choices=[
+                {
+                    "index": 0,
+                    "delta": {"annotations": list(annotations)},
+                    "finish_reason": "stop",
+                }
+            ],
+        ),
+    ]
+    response: Final = litellm.stream_chunk_builder(chunks=chunks)
+    assert isinstance(response, ModelResponse)
+    return response
+
+
+def _non_streamed_response(annotations: tuple[object, ...]) -> ModelResponse:
+    response: Final = convert_to_model_response_object(
+        response_object={
+            "id": "chatcmpl-1",
+            "object": "chat.completion",
+            "created": 0,
+            "model": "gpt-4o",
+            "choices": [
+                {
+                    "index": 0,
+                    "finish_reason": "stop",
+                    "message": {
+                        "role": "assistant",
+                        "content": "See the doc.",
+                        "annotations": list(annotations),
+                    },
+                }
+            ],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+        },
+        model_response_object=ModelResponse(),
+    )
+    assert isinstance(response, ModelResponse)
+    return response
+
+
+def test_streamed_file_citation_is_detected_as_file_search() -> None:
+    response: Final = _streamed_response((_FILE_CITATION,))
+    assert StandardBuiltInToolCostTracking.response_object_includes_file_search_call(response_object=response)
+
+
+def test_non_streamed_file_citation_is_detected_as_file_search() -> None:
+    response: Final = _non_streamed_response((_FILE_CITATION,))
+    assert StandardBuiltInToolCostTracking.response_object_includes_file_search_call(response_object=response)
+
+
+def test_url_citation_only_is_not_file_search() -> None:
+    url_citation: Final = {
+        "type": "url_citation",
+        "url_citation": {"start_index": 0, "end_index": 3, "title": "t", "url": "https://example.com"},
+    }
+    response: Final = _streamed_response((url_citation,))
+    assert not StandardBuiltInToolCostTracking.response_object_includes_file_search_call(response_object=response)
+
+
+def test_streamed_file_citation_bills_file_search_cost() -> None:
+    response: Final = _streamed_response((_FILE_CITATION,))
+    cost: Final = StandardBuiltInToolCostTracking.get_cost_for_built_in_tools(
+        model="gpt-4o",
+        response_object=response,
+        custom_llm_provider="openai",
+        standard_built_in_tools_params={"file_search": {"type": "file_search", "vector_store_ids": ["vs_1"]}},
+    )
+    assert cost == OPENAI_FILE_SEARCH_COST_PER_1K_CALLS