Merge pull request #40820 from BerriAI/litellm_sanitize_unknown_model_error_message

fix(proxy): keep the raw model string out of the unknown-model spend-log error message

litellm/proxy/hooks/proxy_track_cost_callback.py

@@ -180,7 +180,7 @@ class _ProxyDBLogger(CustomLogger):
         # here because the input above is constructed non-None.
         _error_information = cast(
             StandardLoggingPayloadErrorInformation,
-            _sanitize_error_information_for_spend_logs(_error_information),
+            _sanitize_error_information_for_spend_logs(_error_information, original_exception=original_exception),
         )
         _metadata["error_information"] = _error_information
 

litellm/proxy/route_llm_request.py

@@ -145,12 +145,14 @@ ROUTE_ENDPOINT_MAPPING: Final = {
 }
 
 
+_AVAILABLE_MODELS_HINT: Final = "Call `/v1/models` to view available models for your key."
+
+
 class ProxyModelNotFoundError(HTTPException):
     def __init__(self, route: str, model_name: str, retryable_with_model_read_through: bool = True):
         self.retryable_with_model_read_through: Final = retryable_with_model_read_through
-        detail: Final = {
-            "error": f"{route}: Invalid model name passed in model={model_name}. Call `/v1/models` to view available models for your key."
-        }
+        self.spend_log_error_message: Final = f"{route}: Invalid model name passed in. {_AVAILABLE_MODELS_HINT}"
+        detail: Final = {"error": f"{route}: Invalid model name passed in model={model_name}. {_AVAILABLE_MODELS_HINT}"}
         super().__init__(status_code=status.HTTP_400_BAD_REQUEST, detail=detail)
 
 

litellm/proxy/spend_tracking/spend_tracking_utils.py

@@ -1167,6 +1167,7 @@ def _redact_prompt_fields_in_guardrail_entry(
 
 def _sanitize_error_information_for_spend_logs(
     error_information: StandardLoggingPayloadErrorInformation | None,
+    original_exception: BaseException | None = None,
 ) -> StandardLoggingPayloadErrorInformation | None:
     """
     Sanitize ``error_information`` before it lands in ``LiteLLM_SpendLogs.metadata``.
@@ -1188,7 +1189,12 @@ def _sanitize_error_information_for_spend_logs(
     if error_information is None:
         return None
 
-    sanitized = cast(dict, {**error_information})
+    persisted: Final = (
+        {**error_information, "error_message": original_exception.spend_log_error_message}
+        if isinstance(original_exception, ProxyModelNotFoundError)
+        else error_information
+    )
+    sanitized = cast(dict, {**persisted})
 
     if not should_store_prompts_and_responses_in_spend_logs():
         for field in ("error_message", "traceback"):

tests/store_model_in_db_tests/test_openai_error_handling.py

@@ -210,7 +210,8 @@ async def test_chat_completion_bad_model_with_spend_logs():
         assert "traceback" in error_info
         assert error_info["error_code"] == "400"
         assert error_info["error_class"] in ("ProxyModelNotFoundError", "BadRequestError")
-        assert "non-existent-model" in error_info["error_message"]
+        assert "non-existent-model" not in error_info["error_message"]
+        assert "/chat/completions: Invalid model name passed in" in error_info["error_message"]
 
         # Verify request details
         assert log_entry["cache_hit"] == "False"

tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py

@@ -1,5 +1,7 @@
 import asyncio
+import json
 from datetime import datetime
+from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -7,6 +9,7 @@ import pytest
 from litellm.litellm_core_utils.internal_call_metadata import MODEL_ACCESS_GROUP_METADATA_KEY
 from litellm.proxy._types import SpendLogsPayload, UserAPIKeyAuth
 from litellm.proxy.collector import SpendEventConsumer
+from litellm.proxy.db.db_spend_update_writer import DBSpendUpdateWriter
 from litellm.proxy.db.spend_log_tool_index import response_tool_call_names
 from litellm.proxy.hooks.proxy_track_cost_callback import (
     _get_budget_reservation_from_metadata,
@@ -15,6 +18,7 @@ from litellm.proxy.hooks.proxy_track_cost_callback import (
     _update_database_and_spend_counters,
     run_spend_event,
 )
+from litellm.proxy.route_llm_request import ProxyModelNotFoundError
 from litellm.proxy.spend_tracking.spend_event import SpendEventDecodeError, build_spend_event, decode_spend_event
 from litellm.proxy.spend_tracking.spend_event_producer import SpendEventProducer, UnixAddress
 from litellm.proxy.spend_tracking.spend_tracking_utils import get_logging_payload
@@ -2347,3 +2351,27 @@ async def test_sidecar_ignores_an_undecodable_event():  # test-quality-ok: a dis
         mock_proxy_logging.db_spend_update_writer.update_database = AsyncMock()
         await run_spend_event(b"garbage\n")
     mock_proxy_logging.db_spend_update_writer.update_database.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_async_post_call_failure_hook_persists_no_raw_model_on_an_unknown_model_rejection():
+    raw_model: Final = "opus-4.6 Please summarize my medical records\nPatient has diabetes"
+    writer: Final = MagicMock(spec=DBSpendUpdateWriter)
+    writer.update_database = AsyncMock()
+    logger: Final = _ProxyDBLogger(spend_writer=lambda: writer)
+
+    await logger.async_post_call_failure_hook(
+        request_data={"model": raw_model, "messages": [{"role": "user", "content": "hi"}]},
+        original_exception=ProxyModelNotFoundError(route="/chat/completions", model_name=raw_model),
+        user_api_key_dict=UserAPIKeyAuth(api_key="test_api_key"),
+    )
+
+    error_information: Final = writer.update_database.call_args.kwargs["kwargs"]["litellm_params"]["metadata"][
+        "error_information"
+    ]
+    assert "medical records" not in json.dumps(error_information)
+    assert (
+        error_information["error_message"]
+        == "/chat/completions: Invalid model name passed in. Call `/v1/models` to view available models for your key."
+    )
+    assert error_information["error_class"] == "ProxyModelNotFoundError"

tests/test_litellm/proxy/spend_tracking/test_spend_tracking_utils.py

@@ -20,6 +20,7 @@ from litellm.constants import (
     SESSION_ID_OMITTED_METADATA_KEY,
     UNKNOWN_MODEL_SPEND_LOG_MODEL,
 )
+from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.proxy._types import SpendLogsPayload, UserAPIKeyAuth
 from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
@@ -2755,6 +2756,31 @@ def test_sanitize_error_information_redacts_pydantic_assignment_form(
 # ── _redact_logged_api_key unit tests ──────────────────────────────────────
 
 
+@pytest.mark.parametrize(
+    ("original_exception", "expected_error_message"),
+    [
+        (
+            ProxyModelNotFoundError(route="/chat/completions", model_name=_RAW_MODEL_WITH_PROMPT),
+            "/chat/completions: Invalid model name passed in. Call `/v1/models` to view available models for your key.",
+        ),
+        (ValueError("provider timed out"), "provider timed out"),
+    ],
+)
+def test_sanitize_error_information_persists_no_raw_model_for_an_unknown_model_rejection(
+    original_exception: Exception, expected_error_message: str
+):
+    error_information: Final = StandardLoggingPayloadSetup.get_error_information(original_exception=original_exception)
+
+    sanitized: Final = _sanitize_error_information_for_spend_logs(
+        error_information, original_exception=original_exception
+    )
+
+    assert sanitized is not None
+    assert sanitized["error_message"] == expected_error_message
+    assert "medical records" not in json.dumps(sanitized)
+    assert sanitized["error_class"] == type(original_exception).__name__
+
+
 def test_redact_logged_api_key_none_returns_none():
     assert _redact_logged_api_key(None) is None
 

tests/test_litellm/proxy/test_route_llm_request.py

@@ -3,6 +3,7 @@ import pytest
 
 
 
+from typing import Final
 from unittest.mock import MagicMock
 
 from fastapi import HTTPException
@@ -1297,3 +1298,13 @@ async def test_route_request_a2a_agent_miss_does_not_consume_model_read_through(
 
     assert agents_find_unique.await_count == 2
     assert model_table.find_many_wheres == []
+
+
+def test_proxy_model_not_found_error_keeps_the_raw_model_only_in_the_client_response():
+    raw_model: Final = "opus-4.6 Please summarize my medical records\nPatient has diabetes"
+
+    error: Final = ProxyModelNotFoundError(route="/chat/completions", model_name=raw_model)
+
+    assert raw_model in error.detail["error"]
+    assert raw_model not in error.spend_log_error_message
+    assert error.spend_log_error_message.startswith("/chat/completions: Invalid model name passed in")