mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-16 09:22:22 +00:00
fix(guardrails): mark Airia MCP evaluations by the gated event, not by request data
This commit is contained in:
parent
a1361e07c1
commit
b6d3eb555e
2 changed files with 54 additions and 25 deletions
|
|
@ -28,7 +28,7 @@ from litellm.proxy._types import UserAPIKeyAuth
|
|||
from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import UnifiedLLMGuardrails
|
||||
from litellm.types.guardrails import GuardrailEventHooks
|
||||
from litellm.types.proxy.guardrails.guardrail_hooks.airia import AiriaGuardrailConfigModel
|
||||
from litellm.types.utils import MCP_GUARDRAIL_CALL_TYPES, CallTypesLiteral, GenericGuardrailAPIInputs
|
||||
from litellm.types.utils import CallTypesLiteral, GenericGuardrailAPIInputs
|
||||
|
||||
GUARDRAIL_PATH: Final = "/v1/guardrails/litellm"
|
||||
|
||||
|
|
@ -46,7 +46,13 @@ SUPPORTED_EVENT_HOOKS: Final = (
|
|||
GuardrailEventHooks.post_mcp_call,
|
||||
)
|
||||
|
||||
MCP_EVENT_HOOKS: Final = frozenset((GuardrailEventHooks.pre_mcp_call, GuardrailEventHooks.post_mcp_call))
|
||||
UNKNOWN_MCP_TOOL_NAME: Final = "unknown"
|
||||
|
||||
_EVALUATING_DURING_CALL: Final[ContextVar[bool]] = ContextVar("airia_evaluating_during_call", default=False)
|
||||
# Request bodies on several routes are merged into the hook data, so MCP keys in it prove nothing; the event the
|
||||
# framework gates on is the only signal a client cannot set
|
||||
_PENDING_EVENT: Final[ContextVar[GuardrailEventHooks | None]] = ContextVar("airia_pending_event", default=None)
|
||||
|
||||
|
||||
def _stream_cannot_carry_a_rewrite(request_data: Mapping[str, object] | None) -> bool:
|
||||
|
|
@ -93,20 +99,13 @@ def _request_route(request_data: Mapping[str, object]) -> str | None:
|
|||
return None
|
||||
|
||||
|
||||
def _mcp_tool_name(request_data: Mapping[str, object], logging_obj: "LiteLLMLoggingObj | None") -> str | None:
|
||||
"""The MCP tool this evaluation is for, or None for model traffic.
|
||||
|
||||
The logger's call type wins when there is one, because a chat client can put MCP keys in its request body.
|
||||
"""
|
||||
logged_call_type: Final[object] = getattr(logging_obj, "call_type", None)
|
||||
if logging_obj is not None and logged_call_type not in MCP_GUARDRAIL_CALL_TYPES:
|
||||
return None
|
||||
def _mcp_tool_name(request_data: Mapping[str, object]) -> str:
|
||||
pre_call_name: Final = request_data.get("mcp_tool_name")
|
||||
if isinstance(pre_call_name, str) and pre_call_name:
|
||||
return pre_call_name
|
||||
call_metadata: Final = _as_object(request_data.get("mcp_tool_call_metadata"))
|
||||
post_call_name: Final = call_metadata.get("name") if call_metadata is not None else None
|
||||
return post_call_name if isinstance(post_call_name, str) and post_call_name else None
|
||||
return post_call_name if isinstance(post_call_name, str) and post_call_name else UNKNOWN_MCP_TOOL_NAME
|
||||
|
||||
|
||||
def _as_object(value: object) -> Mapping[str, object] | None:
|
||||
|
|
@ -158,6 +157,15 @@ class AiriaGuardrail(CustomGuardrail):
|
|||
)
|
||||
verbose_proxy_logger.info("AiriaGuardrail initialized with gateway: %s", self.api_base)
|
||||
|
||||
def should_run_guardrail(
|
||||
self,
|
||||
data: object,
|
||||
event_type: GuardrailEventHooks,
|
||||
) -> bool:
|
||||
runs: Final = super().should_run_guardrail(data, event_type) # pyright: ignore[reportUnknownMemberType] # the parent annotates data as untyped
|
||||
_ = _PENDING_EVENT.set(event_type if runs else None)
|
||||
return runs
|
||||
|
||||
def supports_structured_messages(self) -> bool:
|
||||
return False
|
||||
|
||||
|
|
@ -169,6 +177,8 @@ class AiriaGuardrail(CustomGuardrail):
|
|||
input_type: Literal["request", "response"],
|
||||
logging_obj: "LiteLLMLoggingObj | None" = None,
|
||||
) -> GenericGuardrailAPIInputs:
|
||||
is_mcp_tool_evaluation: Final = _PENDING_EVENT.get() in MCP_EVENT_HOOKS
|
||||
_ = _PENDING_EVENT.set(None)
|
||||
call_id: Final = (
|
||||
logging_obj.litellm_call_id
|
||||
if logging_obj
|
||||
|
|
@ -187,7 +197,7 @@ class AiriaGuardrail(CustomGuardrail):
|
|||
"tool_calls": inputs.get("tool_calls") or (),
|
||||
"model": inputs.get("model"),
|
||||
"litellm_call_id": call_id,
|
||||
"mcp_tool_name": _mcp_tool_name(request_data, logging_obj),
|
||||
"mcp_tool_name": _mcp_tool_name(request_data or {}) if is_mcp_tool_evaluation else None,
|
||||
},
|
||||
headers={"Authorization": f"Bearer {self.api_key}"},
|
||||
)
|
||||
|
|
|
|||
|
|
@ -21,6 +21,7 @@ from litellm.proxy.guardrails.init_guardrails import (
|
|||
init_guardrails_v2, # pyright: ignore[reportUnknownVariableType] # untyped upstream
|
||||
)
|
||||
from litellm.proxy.utils import ProxyLogging
|
||||
from litellm.types.guardrails import GuardrailEventHooks
|
||||
from litellm.types.proxy.guardrails.guardrail_hooks.airia import AiriaGuardrailConfigModel
|
||||
from litellm.types.utils import GenericGuardrailAPIInputs
|
||||
|
||||
|
|
@ -188,28 +189,46 @@ async def test_post_mcp_call_redaction_masks_the_tool_result(monkeypatch: pytest
|
|||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_mcp_keys_in_a_chat_request_body_do_not_mark_it_as_a_tool_call() -> None:
|
||||
guardrail: Final = _make_guardrail_with_response({"action": "NONE"})
|
||||
logging_obj: Final = LiteLLMLoggingObj(
|
||||
model="gpt-4o",
|
||||
messages=[],
|
||||
stream=False,
|
||||
call_type="acompletion",
|
||||
start_time=datetime(2026, 1, 1, tzinfo=timezone.utc),
|
||||
litellm_call_id="call-1",
|
||||
function_id="1",
|
||||
@pytest.mark.parametrize("logging_obj", [None, "acompletion"])
|
||||
async def test_mcp_keys_in_a_request_body_do_not_mark_model_traffic_as_a_tool_call(logging_obj: str | None) -> None:
|
||||
guardrail: Final = _make_guardrail_with_response({"action": "NONE"}, event_hook="pre_call", default_on=True)
|
||||
request_data: Final[dict[str, object]] = {"prompt": "draw a cat", "mcp_tool_name": "send_email"}
|
||||
logger: Final = (
|
||||
LiteLLMLoggingObj(
|
||||
model="gpt-4o",
|
||||
messages=[],
|
||||
stream=False,
|
||||
call_type=logging_obj,
|
||||
start_time=datetime(2026, 1, 1, tzinfo=timezone.utc),
|
||||
litellm_call_id="call-1",
|
||||
function_id="1",
|
||||
)
|
||||
if logging_obj
|
||||
else None
|
||||
)
|
||||
|
||||
assert guardrail.should_run_guardrail(data=request_data, event_type=GuardrailEventHooks.pre_call)
|
||||
await guardrail.apply_guardrail(
|
||||
inputs=_inputs(),
|
||||
request_data={"mcp_tool_name": "send_email", "mcp_tool_call_metadata": {"name": "send_email"}},
|
||||
input_type="request",
|
||||
logging_obj=logging_obj,
|
||||
inputs=_inputs(), request_data=request_data, input_type="request", logging_obj=logger
|
||||
)
|
||||
|
||||
assert _post_mock(guardrail).call_args.kwargs["json"]["mcp_tool_name"] is None
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_an_mcp_evaluation_does_not_mark_the_next_model_evaluation_as_a_tool_call() -> None:
|
||||
guardrail: Final = _make_guardrail_with_response(
|
||||
{"action": "NONE"}, event_hook=["pre_mcp_call", "pre_call"], default_on=True
|
||||
)
|
||||
tool_call: Final[dict[str, object]] = {"mcp_tool_name": "send_email"}
|
||||
assert guardrail.should_run_guardrail(data=tool_call, event_type=GuardrailEventHooks.pre_mcp_call)
|
||||
await guardrail.apply_guardrail(inputs=_inputs(), request_data=tool_call, input_type="request")
|
||||
|
||||
await guardrail.apply_guardrail(inputs=_inputs(), request_data=tool_call, input_type="request")
|
||||
|
||||
assert _post_mock(guardrail).call_args.kwargs["json"]["mcp_tool_name"] is None
|
||||
|
||||
|
||||
def _during_call_request() -> dict[str, object]:
|
||||
return {"model": "gpt-4o", "messages": [{"role": "user", "content": "mail ada@example.com"}]}
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue