mirror of
https://github.com/hwchase17/langchain.git
synced 2025-05-04 22:58:42 +00:00
Signed-off-by: ChengZi <chen.zhang@zilliz.com> Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com> Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com> Co-authored-by: Dan O'Donovan <dan.odonovan@gmail.com> Co-authored-by: Tom Daniel Grande <tomdgrande@gmail.com> Co-authored-by: Grande <Tom.Daniel.Grande@statsbygg.no> Co-authored-by: Bagatur <baskaryan@gmail.com> Co-authored-by: ccurme <chester.curme@gmail.com> Co-authored-by: Harrison Chase <hw.chase.17@gmail.com> Co-authored-by: Tomaz Bratanic <bratanic.tomaz@gmail.com> Co-authored-by: ZhangShenao <15201440436@163.com> Co-authored-by: Friso H. Kingma <fhkingma@gmail.com> Co-authored-by: ChengZi <chen.zhang@zilliz.com> Co-authored-by: Nuno Campos <nuno@langchain.dev> Co-authored-by: Morgante Pell <morgantep@google.com>
108 lines
3.4 KiB
Python
108 lines
3.4 KiB
Python
import logging
|
|
from typing import Any, Callable, Dict, List, Optional
|
|
|
|
from langchain_core.callbacks import CallbackManagerForLLMRun
|
|
from langchain_core.language_models.llms import LLM
|
|
from pydantic import model_validator
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def default_guardrail_violation_handler(violation: dict) -> str:
|
|
"""Default guardrail violation handler.
|
|
|
|
Args:
|
|
violation (dict): The violation dictionary.
|
|
|
|
Returns:
|
|
str: The canned response.
|
|
"""
|
|
if violation.get("canned_response"):
|
|
return violation["canned_response"]
|
|
guardrail_name = (
|
|
f"Guardrail {violation.get('offending_guardrail')}"
|
|
if violation.get("offending_guardrail")
|
|
else "A guardrail"
|
|
)
|
|
raise ValueError(
|
|
f"{guardrail_name} was violated without a proper guardrail violation handler."
|
|
)
|
|
|
|
|
|
class LayerupSecurity(LLM):
|
|
"""Layerup Security LLM service."""
|
|
|
|
llm: LLM
|
|
layerup_api_key: str
|
|
layerup_api_base_url: str = "https://api.uselayerup.com/v1"
|
|
prompt_guardrails: Optional[List[str]] = []
|
|
response_guardrails: Optional[List[str]] = []
|
|
mask: bool = False
|
|
metadata: Optional[Dict[str, Any]] = {}
|
|
handle_prompt_guardrail_violation: Callable[[dict], str] = (
|
|
default_guardrail_violation_handler
|
|
)
|
|
handle_response_guardrail_violation: Callable[[dict], str] = (
|
|
default_guardrail_violation_handler
|
|
)
|
|
client: Any #: :meta private:
|
|
|
|
@model_validator(mode="before")
|
|
@classmethod
|
|
def validate_layerup_sdk(cls, values: Dict[str, Any]) -> Any:
|
|
try:
|
|
from layerup_security import LayerupSecurity as LayerupSecuritySDK
|
|
|
|
values["client"] = LayerupSecuritySDK(
|
|
api_key=values["layerup_api_key"],
|
|
base_url=values["layerup_api_base_url"],
|
|
)
|
|
except ImportError:
|
|
raise ImportError(
|
|
"Could not import LayerupSecurity SDK. "
|
|
"Please install it with `pip install LayerupSecurity`."
|
|
)
|
|
return values
|
|
|
|
@property
|
|
def _llm_type(self) -> str:
|
|
return "layerup_security"
|
|
|
|
def _call(
|
|
self,
|
|
prompt: str,
|
|
stop: Optional[List[str]] = None,
|
|
run_manager: Optional[CallbackManagerForLLMRun] = None,
|
|
**kwargs: Any,
|
|
) -> str:
|
|
messages = [{"role": "user", "content": prompt}]
|
|
unmask_response = None
|
|
|
|
if self.mask:
|
|
messages, unmask_response = self.client.mask_prompt(messages, self.metadata)
|
|
|
|
if self.prompt_guardrails:
|
|
security_response = self.client.execute_guardrails(
|
|
self.prompt_guardrails, messages, prompt, self.metadata
|
|
)
|
|
if not security_response["all_safe"]:
|
|
return self.handle_prompt_guardrail_violation(security_response)
|
|
|
|
result = self.llm._call(
|
|
messages[0]["content"], run_manager=run_manager, **kwargs
|
|
)
|
|
|
|
if self.mask and unmask_response:
|
|
result = unmask_response(result)
|
|
|
|
messages.append({"role": "assistant", "content": result})
|
|
|
|
if self.response_guardrails:
|
|
security_response = self.client.execute_guardrails(
|
|
self.response_guardrails, messages, result, self.metadata
|
|
)
|
|
if not security_response["all_safe"]:
|
|
return self.handle_response_guardrail_violation(security_response)
|
|
|
|
return result
|