Merge ba4f91b5d1 into e9b1c11098

litellm/litellm_core_utils/token_counter.py

@@ -1812,7 +1812,13 @@ def _count_anthropic_content(
 
 
 _TEXT_FIELD_BY_BLOCK_TYPE: Final = MappingProxyType(
-    {"text": "text", "thinking": "thinking", "redacted_thinking": "thinking", "compaction": "content"}
+    {
+        "text": "text",
+        "thinking": "thinking",
+        "redacted_thinking": "thinking",
+        "compaction": "content",
+        "refusal": "refusal",
+    }
 )
 
 
@@ -1892,7 +1898,7 @@ def _count_content_list(
                     f"Invalid content item type: {content_type}. "
                     f"Expected str or dict with 'type' field "
                     f"(text, image_url, image, document, file, tool_use, tool_result, thinking, redacted_thinking, "
-                    f"compaction, tool_reference, search_result)."
+                    f"compaction, tool_reference, search_result, refusal)."
                 )
         return num_tokens
     except Exception as e:

litellm/types/llms/openai.py

@@ -882,6 +882,11 @@ class ChatCompletionFileObject(TypedDict):
     file: ChatCompletionFileObjectFile
 
 
+class ChatCompletionRefusalObject(TypedDict):
+    type: ReadOnly[Literal["refusal"]]
+    refusal: ReadOnly[str]
+
+
 OpenAIMessageContentListBlock = (
     ChatCompletionTextObject
     | ChatCompletionImageObject
@@ -889,6 +894,7 @@ OpenAIMessageContentListBlock = (
     | ChatCompletionDocumentObject
     | ChatCompletionVideoObject
     | ChatCompletionFileObject
+    | ChatCompletionRefusalObject
 )
 
 OpenAIMessageContent = str | Iterable[OpenAIMessageContentListBlock]

tests/unit/litellm_core_utils/test_token_counter.py

@@ -3625,3 +3625,20 @@ def test_openai_model_check_holds_while_openai_models_are_registered_concurrentl
             registering.result()
     finally:
         litellm.open_ai_chat_completion_models.difference_update(registered)
+
+
+def test_refusal_content_part_counts_as_the_text_it_replaces() -> None:
+    """An assistant refusal part is a documented OpenAI content block, so it must count instead of raising"""
+    with_text = [
+        {"role": "user", "content": "A request"},
+        {"role": "assistant", "content": [{"type": "text", "text": "I cannot help with that request."}]},
+        {"role": "user", "content": "Please suggest a safe alternative."},
+    ]
+    with_refusal = [
+        {"role": "user", "content": "A request"},
+        {"role": "assistant", "content": [{"type": "refusal", "refusal": "I cannot help with that request."}]},
+        {"role": "user", "content": "Please suggest a safe alternative."},
+    ]
+    assert token_counter_old(model="gpt-4o", messages=with_refusal) == token_counter_old(
+        model="gpt-4o", messages=with_text
+    )