fix(snowflake): answer 400 for user parts Cortex Claude cannot take, and link an https pdf with a pdf format as a document

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -954,6 +954,11 @@ def create_anthropic_image_param(
             return _anthropic_content_element_factory(
                 convert_to_anthropic_image_obj(openai_image_url=base64_url, format=format)
             )
+        elif format is not None and format.partition(";")[0].strip() == "application/pdf":
+            return AnthropicMessagesDocumentParam(
+                type="document",
+                source=AnthropicContentParamSourceUrl(type="url", url=image_url),
+            )
         else:
             # HTTPS URL - pass directly for regular Anthropic
             return AnthropicMessagesImageParam(

litellm/llms/snowflake/chat/transformation.py

@@ -9,6 +9,7 @@ Ref: https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-rest-api
 """
 
 from collections.abc import Mapping, Sequence
+from itertools import chain
 from typing import TYPE_CHECKING, Any, Final, Protocol, TypedDict
 
 import httpx
@@ -18,6 +19,7 @@ import litellm
 from litellm.exceptions import BadRequestError
 from litellm.litellm_core_utils.prompt_templates.factory import anthropic_messages_pt
 from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media
+from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import message_field, parts_of
 from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.chat.handler import ModelResponseIterator as AnthropicStreamParser
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
@@ -98,6 +100,20 @@ def is_snowflake_claude_model(model: str) -> bool:
     return any(name.startswith(p) for p in _CLAUDE_MODEL_PREFIXES)
 
 
+_CORTEX_CLAUDE_USER_PART_TYPES: Final = frozenset({"text", "image_url", "document", "file"})
+
+
+def _user_parts(message: AllMessageValues) -> tuple[object, ...]:
+    return parts_of(message.get("content")) if message["role"] == "user" else ()
+
+
+def _untranslatable_user_part_types(messages: Sequence[AllMessageValues]) -> frozenset[str]:
+    part_types: Final = frozenset(
+        str(message_field(part, "type")) for part in chain.from_iterable(map(_user_parts, messages))
+    )
+    return part_types - _CORTEX_CLAUDE_USER_PART_TYPES
+
+
 def _clean_input_schema(schema: object) -> object:
     return {key: value for key, value in schema.items() if key != "$schema"} if isinstance(schema, Mapping) else schema
 
@@ -207,6 +223,13 @@ class SnowflakeConfig(SnowflakeBaseConfig, OpenAIGPTConfig):
     ) -> tuple[list[AnthropicSystemMessageContent] | None, list[AllAnthropicPassThroughMessageValues]]:
         """Cortex's /messages endpoint speaks Anthropic's dialect, so the shared translators own the shape."""
         conversation_messages: Final = list(messages)
+        untranslatable: Final = _untranslatable_user_part_types(conversation_messages)
+        if untranslatable and not drop_params:
+            raise SnowflakeException(
+                status_code=400,
+                message=f"Snowflake Cortex Claude cannot take {', '.join(sorted(untranslatable))} parts in a user "
+                "message. Remove them, or pass drop_params=True to drop them",
+            )
         system: Final = [
             part
             for part in AnthropicConfig().translate_system_message(messages=conversation_messages)

tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py

@@ -9436,3 +9436,16 @@ def test_create_anthropic_image_param_sends_a_downloaded_pdf_as_a_document(monke
         "type": "document",
         "source": {"type": "base64", "media_type": "application/pdf", "data": base64.b64encode(pdf_bytes).decode()},
     }
+
+
+def test_create_anthropic_image_param_links_an_https_pdf_as_a_document_only_when_format_says_pdf():
+    url = "https://example.com/report.pdf"
+
+    assert create_anthropic_image_param({"url": url, "format": "application/pdf"}) == {
+        "type": "document",
+        "source": {"type": "url", "url": url},
+    }
+    assert create_anthropic_image_param({"url": url, "format": "image/png"}) == {
+        "type": "image",
+        "source": {"type": "url", "url": url},
+    }

tests/unit/llms/snowflake/chat/test_snowflake_chat_transformation.py

@@ -744,6 +744,32 @@ class TestSnowflakeCortexClaudeFixes:
         )
         assert body["system"] == [{"type": "text", "text": "be terse"}]
 
+    _AUDIO_TURN = (
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "hear"},
+                {"type": "input_audio", "input_audio": {"data": "UklGRg==", "format": "wav"}},
+            ],
+        },
+    )
+
+    def test_user_part_cortex_claude_cannot_take_is_a_400_naming_it(self):
+        with pytest.raises(SnowflakeException) as exc_info:
+            self._transform(list(self._AUDIO_TURN))
+        assert exc_info.value.status_code == 400
+        assert "input_audio" in str(exc_info.value)
+
+    def test_drop_params_drops_a_user_part_cortex_claude_cannot_take(self):
+        body = SnowflakeConfig().transform_request(
+            model="snowflake/claude-sonnet-4-6",
+            messages=list(self._AUDIO_TURN),
+            optional_params={},
+            litellm_params={"drop_params": True},
+            headers={},
+        )
+        assert body["messages"] == [{"role": "user", "content": [{"type": "text", "text": "hear"}]}]
+
     def test_unknown_charset_on_a_text_attachment_is_a_400(self):
         with pytest.raises(SnowflakeException) as exc_info:
             self._transform(