fix(guardrails): restore the google_text_moderation embeddings fix an earlier main merge dropped

This commit is contained in:
mateo-berri 2026-10-28 23:26:19 -07:00
parent 243f6fd3c3
commit 63b22cb142
4 changed files with 125 additions and 3 deletions

View file

@ -15,7 +15,10 @@ import litellm
from litellm._logging import verbose_proxy_logger
from litellm.integrations.custom_logger import CustomLogger
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.guardrails._content_utils import iter_message_text
from litellm.proxy.guardrails._content_utils import (
is_non_conversational_request,
iter_message_text,
)
from litellm.types.utils import CallTypesLiteral
_THRESHOLD_VALUE: Final = TypeAdapter(object)
@ -94,7 +97,7 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
async def async_moderation_hook(
self,
data: dict,
data: dict[str, object],
user_api_key_dict: UserAPIKeyAuth,
call_type: CallTypesLiteral,
):
@ -102,6 +105,9 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
- Calls Google's Text Moderation API
- Rejects request if it fails safety check
"""
if is_non_conversational_request(call_type, data):
verbose_proxy_logger.debug("Google text moderation: skipping non-conversational call type %s", call_type)
return
# Covers multimodal list content + Responses-API input.
text = "".join(iter_message_text(data))
if text:

View file

@ -62,6 +62,20 @@ def is_non_conversational_call_type(call_type: str) -> bool:
return call_type in NON_CONVERSATIONAL_CALL_TYPES
_PASSTHROUGH_EMBEDDINGS_ENDPOINT: Final = re.compile(r"/?(?:[\w-]+/)*embeddings/?")
def is_non_conversational_request(call_type: str, data: Mapping[str, object]) -> bool:
if is_non_conversational_call_type(call_type):
return True
endpoint: Final = data.get("endpoint")
return (
call_type == "allm_passthrough_route"
and isinstance(endpoint, str)
and _PASSTHROUGH_EMBEDDINGS_ENDPOINT.fullmatch(endpoint) is not None
)
TEXT_PART_TYPES: Final[frozenset[str]] = frozenset(
{"text", "input_text", "output_text", "summary_text", "reasoning_text"}
)

View file

@ -12,6 +12,7 @@ from litellm.proxy.guardrails._content_utils import (
build_inspection_messages,
has_non_string_content,
is_non_conversational_call_type,
is_non_conversational_request,
is_string_batch_input,
iter_message_text,
iter_tool_json_text,
@ -819,6 +820,30 @@ def test_is_non_conversational_call_type_defaults_to_inspecting_unknown_call_typ
assert is_non_conversational_call_type("some_future_call_type") is False
@pytest.mark.parametrize(
("call_type", "endpoint", "expected"),
[
("aembedding", None, True),
("allm_passthrough_route", "embeddings", True),
("allm_passthrough_route", "/v1/embeddings/", True),
("allm_passthrough_route", "chat/completions", False),
("allm_passthrough_route", "model/amazon.titan-embed-text-v2:0/invoke", False),
("allm_passthrough_route", "fine_tuned_embeddings_jobs", False),
("allm_passthrough_route", "chat/completions#embeddings", False),
("allm_passthrough_route", "chat/completions?x=embeddings", False),
("allm_passthrough_route", "chat/completions;embeddings", False),
("allm_passthrough_route", None, False),
("acompletion", "embeddings", False),
],
)
def test_is_non_conversational_request_skips_only_embeddings_requests(
call_type: str, endpoint: str | None, expected: bool
):
data: Final[dict[str, object]] = {"endpoint": endpoint} if endpoint is not None else {}
assert is_non_conversational_request(call_type, data) is expected
@pytest.mark.parametrize(
"tool_input",
(

View file

@ -12,17 +12,19 @@ inspection by the relevant guardrail hook:
not just ``choices[0]``.
"""
from types import SimpleNamespace
from typing import Any, Dict, Final
from unittest.mock import AsyncMock, MagicMock, patch
import pytest
from fastapi import HTTPException
from httpx import Request, Response
import litellm
from litellm import DualCache
from litellm.constants import DEFAULT_OPENAI_MODERATIONS_MODEL
from litellm.proxy._types import UserAPIKeyAuth
from litellm.types.utils import Choices, Message, ModelResponse
from litellm.types.utils import CallTypesLiteral, Choices, Message, ModelResponse
@pytest.fixture
@ -905,3 +907,78 @@ async def test_google_text_moderation_inspects_multimodal_content(user_api_key):
)
assert seen_documents == ["hello world"]
class _FinanceFlaggingLanguageClient:
def __init__(self) -> None:
self.moderated_texts: Final[list[str]] = [] # mutable-ok: records what reached Google
def moderate_text(self, request: SimpleNamespace) -> SimpleNamespace:
self.moderated_texts.append(request.document.content)
return SimpleNamespace(moderation_categories=[SimpleNamespace(name="Finance", confidence=0.885)])
def _google_text_moderation_flagging_finance():
from enterprise.enterprise_hooks.google_text_moderation import (
ENTERPRISE_GoogleTextModeration,
)
guard: Final = ENTERPRISE_GoogleTextModeration.__new__(ENTERPRISE_GoogleTextModeration)
guard.client = _FinanceFlaggingLanguageClient()
guard.language_document = lambda content, type_: SimpleNamespace(content=content)
guard.moderate_text_request = lambda document: SimpleNamespace(document=document)
guard.document_type = "PLAIN_TEXT"
guard.finance_confidence_threshold = 0.8
return guard
@pytest.mark.asyncio
@pytest.mark.parametrize(
("call_type", "data"),
[
("aembedding", {"model": "text-embedding-ada-002", "input": "stocks"}),
("embedding", {"model": "text-embedding-ada-002", "input": ["stocks", "mortgage"]}),
("allm_passthrough_route", {"model": "gigachat-embeddings", "endpoint": "embeddings", "input": ["stocks"]}),
],
)
async def test_google_text_moderation_leaves_embedding_input_unmoderated(
user_api_key: UserAPIKeyAuth, call_type: CallTypesLiteral, data: dict[str, object]
):
guard: Final = _google_text_moderation_flagging_finance()
await guard.async_moderation_hook(data=data, user_api_key_dict=user_api_key, call_type=call_type)
assert guard.client.moderated_texts == []
@pytest.mark.asyncio
@pytest.mark.parametrize(
("call_type", "data"),
[
("acompletion", {"model": "gpt-4o", "messages": [{"role": "user", "content": "stocks"}]}),
("anthropic_messages", {"model": "claude", "messages": [{"role": "user", "content": "stocks"}]}),
("aresponses", {"model": "gpt-4o", "input": "stocks"}),
(
"allm_passthrough_route",
{"model": "gigachat", "endpoint": "chat/completions", "messages": [{"role": "user", "content": "stocks"}]},
),
(
"allm_passthrough_route",
{
"model": "gigachat",
"endpoint": "chat/completions#embeddings",
"messages": [{"role": "user", "content": "stocks"}],
},
),
],
)
async def test_google_text_moderation_still_blocks_conversations(
user_api_key: UserAPIKeyAuth, call_type: CallTypesLiteral, data: dict[str, object]
):
guard: Final = _google_text_moderation_flagging_finance()
with pytest.raises(HTTPException) as exc_info:
await guard.async_moderation_hook(data=data, user_api_key_dict=user_api_key, call_type=call_type)
assert exc_info.value.status_code == 400
assert guard.client.moderated_texts == ["stocks"]