Spaces:
Configuration error
Configuration error
File size: 12,785 Bytes
447ebeb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 |
# +-------------------------------------------------------------+
#
# Use Aim Security Guardrails for your LLM calls
# https://www.aim.security/
#
# +-------------------------------------------------------------+
import asyncio
import json
import os
from typing import Any, AsyncGenerator, Literal, Optional, Union
from fastapi import HTTPException
from pydantic import BaseModel
from websockets.asyncio.client import ClientConnection, connect
from litellm import DualCache
from litellm._logging import verbose_proxy_logger
from litellm._version import version as litellm_version
from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.llms.custom_httpx.http_handler import (
get_async_httpx_client,
httpxSpecialProvider,
)
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.proxy_server import StreamingCallbackError
from litellm.types.utils import (
Choices,
EmbeddingResponse,
ImageResponse,
ModelResponse,
ModelResponseStream,
)
class AimGuardrailMissingSecrets(Exception):
pass
class AimGuardrail(CustomGuardrail):
def __init__(
self, api_key: Optional[str] = None, api_base: Optional[str] = None, **kwargs
):
self.async_handler = get_async_httpx_client(
llm_provider=httpxSpecialProvider.GuardrailCallback
)
self.api_key = api_key or os.environ.get("AIM_API_KEY")
if not self.api_key:
msg = (
"Couldn't get Aim api key, either set the `AIM_API_KEY` in the environment or "
"pass it as a parameter to the guardrail in the config file"
)
raise AimGuardrailMissingSecrets(msg)
self.api_base = (
api_base or os.environ.get("AIM_API_BASE") or "https://api.aim.security"
)
self.ws_api_base = self.api_base.replace("http://", "ws://").replace(
"https://", "wss://"
)
self.dlp_entities: list[dict] = []
self._max_dlp_entities = 100
super().__init__(**kwargs)
async def async_pre_call_hook(
self,
user_api_key_dict: UserAPIKeyAuth,
cache: DualCache,
data: dict,
call_type: Literal[
"completion",
"text_completion",
"embeddings",
"image_generation",
"moderation",
"audio_transcription",
"pass_through_endpoint",
"rerank",
],
) -> Union[Exception, str, dict, None]:
verbose_proxy_logger.debug("Inside AIM Pre-Call Hook")
return await self.call_aim_guardrail(
data, hook="pre_call", key_alias=user_api_key_dict.key_alias
)
async def async_moderation_hook(
self,
data: dict,
user_api_key_dict: UserAPIKeyAuth,
call_type: Literal[
"completion",
"embeddings",
"image_generation",
"moderation",
"audio_transcription",
"responses",
],
) -> Union[Exception, str, dict, None]:
verbose_proxy_logger.debug("Inside AIM Moderation Hook")
await self.call_aim_guardrail(
data, hook="moderation", key_alias=user_api_key_dict.key_alias
)
return data
async def call_aim_guardrail(
self, data: dict, hook: str, key_alias: Optional[str]
) -> dict:
user_email = data.get("metadata", {}).get("headers", {}).get("x-aim-user-email")
call_id = data.get("litellm_call_id")
headers = self._build_aim_headers(
hook=hook,
key_alias=key_alias,
user_email=user_email,
litellm_call_id=call_id,
)
response = await self.async_handler.post(
f"{self.api_base}/detect/openai/v2",
headers=headers,
json={"messages": data.get("messages", [])},
)
response.raise_for_status()
res = response.json()
required_action = res.get("required_action")
action_type = required_action and required_action.get("action_type", None)
if action_type is None:
verbose_proxy_logger.debug("Aim: No required action specified")
return data
if action_type == "monitor_action":
verbose_proxy_logger.info("Aim: monitor action")
elif action_type == "block_action":
self._handle_block_action(res["analysis_result"], required_action)
elif action_type == "anonymize_action":
return self._anonymize_request(
res["analysis_result"], required_action, data
)
else:
verbose_proxy_logger.error(f"Aim: {action_type} action")
return data
def _handle_block_action(self, analysis_result: Any, required_action: Any) -> None:
detection_message = required_action.get("detection_message", None)
verbose_proxy_logger.info(
"Aim: Violation detected enabled policies: {policies}".format(
policies=list(analysis_result["policy_drill_down"].keys()),
),
)
raise HTTPException(status_code=400, detail=detection_message)
def _anonymize_request(
self, analysis_result: Any, required_action: Any, data: dict
) -> dict:
verbose_proxy_logger.info("Aim: anonymize action")
redaction_result = required_action and required_action.get(
"chat_redaction_result"
)
if not redaction_result:
return data
if analysis_result and analysis_result.get("session_entities"):
self._set_dlp_entities(analysis_result.get("session_entities"))
data["messages"] = [
{
"role": redaction_result["redacted_new_message"]["role"],
"content": redaction_result["redacted_new_message"]["content"],
}
] + [
{
"role": message["role"],
"content": message["content"],
}
for message in redaction_result["all_redacted_messages"]
]
return data
async def call_aim_guardrail_on_output(
self, request_data: dict, output: str, hook: str, key_alias: Optional[str]
) -> Optional[dict]:
user_email = (
request_data.get("metadata", {}).get("headers", {}).get("x-aim-user-email")
)
call_id = request_data.get("litellm_call_id")
response = await self.async_handler.post(
f"{self.api_base}/detect/output/v2",
headers=self._build_aim_headers(
hook=hook,
key_alias=key_alias,
user_email=user_email,
litellm_call_id=call_id,
),
json={"output": output, "messages": request_data.get("messages", [])},
)
response.raise_for_status()
res = response.json()
required_action = res.get("required_action")
action_type = required_action and required_action.get("action_type", None)
if action_type and action_type == "block_action":
return self._handle_block_action_on_output(
res["analysis_result"], required_action
)
return self._deanonymize_output(output)
def _handle_block_action_on_output(
self, analysis_result: Any, required_action: Any
) -> dict | None:
detection_message = required_action.get("detection_message", None)
verbose_proxy_logger.info(
"Aim: detected: {detected}, enabled policies: {policies}".format(
detected=True,
policies=list(analysis_result["policy_drill_down"].keys()),
),
)
return {"detection_message": detection_message}
def _deanonymize_output(self, output: str) -> dict | None:
try:
for entity in self.dlp_entities:
output = output.replace(f"[{entity['name']}]", entity["content"])
return {"redacted_output": output}
except Exception as e:
verbose_proxy_logger.error(f"Aim: Error while redacting output: {e}")
return None
def _build_aim_headers(
self,
*,
hook: str,
key_alias: Optional[str],
user_email: Optional[str],
litellm_call_id: Optional[str],
):
"""
A helper function to build the http headers that are required by AIM guardrails.
"""
return (
{
"Authorization": f"Bearer {self.api_key}",
# Used by Aim to apply only the guardrails that should be applied in a specific request phase.
"x-aim-litellm-hook": hook,
# Used by Aim to track LiteLLM version and provide backward compatibility.
"x-aim-litellm-version": litellm_version,
}
# Used by Aim to track together single call input and output
| ({"x-aim-litellm-call-id": litellm_call_id} if litellm_call_id else {})
# Used by Aim to track guardrails violations by user.
| ({"x-aim-user-email": user_email} if user_email else {})
| (
{
# Used by Aim apply only the guardrails that are associated with the key alias.
"x-aim-litellm-key-alias": key_alias,
}
if key_alias
else {}
)
)
async def async_post_call_success_hook(
self,
data: dict,
user_api_key_dict: UserAPIKeyAuth,
response: Union[Any, ModelResponse, EmbeddingResponse, ImageResponse],
) -> Any:
if (
isinstance(response, ModelResponse)
and response.choices
and isinstance(response.choices[0], Choices)
):
content = response.choices[0].message.content or ""
aim_output_guardrail_result = await self.call_aim_guardrail_on_output(
data, content, hook="output", key_alias=user_api_key_dict.key_alias
)
if aim_output_guardrail_result and aim_output_guardrail_result.get(
"detection_message"
):
raise HTTPException(
status_code=400,
detail=aim_output_guardrail_result.get("detection_message"),
)
if aim_output_guardrail_result and aim_output_guardrail_result.get(
"redacted_output"
):
response.choices[0].message.content = aim_output_guardrail_result.get(
"redacted_output"
)
return response
async def async_post_call_streaming_iterator_hook(
self,
user_api_key_dict: UserAPIKeyAuth,
response,
request_data: dict,
) -> AsyncGenerator[ModelResponseStream, None]:
user_email = (
request_data.get("metadata", {}).get("headers", {}).get("x-aim-user-email")
)
call_id = request_data.get("litellm_call_id")
async with connect(
f"{self.ws_api_base}/detect/output/ws",
additional_headers=self._build_aim_headers(
hook="output",
key_alias=user_api_key_dict.key_alias,
user_email=user_email,
litellm_call_id=call_id,
),
) as websocket:
sender = asyncio.create_task(
self.forward_the_stream_to_aim(websocket, response)
)
while True:
result = json.loads(await websocket.recv())
if verified_chunk := result.get("verified_chunk"):
yield ModelResponseStream.model_validate(verified_chunk)
else:
sender.cancel()
if result.get("done"):
return
if blocking_message := result.get("blocking_message"):
raise StreamingCallbackError(blocking_message)
verbose_proxy_logger.error(
f"Unknown message received from AIM: {result}"
)
return
async def forward_the_stream_to_aim(
self,
websocket: ClientConnection,
response_iter,
) -> None:
async for chunk in response_iter:
if isinstance(chunk, BaseModel):
chunk = chunk.model_dump_json()
if isinstance(chunk, dict):
chunk = json.dumps(chunk)
await websocket.send(chunk)
await websocket.send(json.dumps({"done": True}))
def _set_dlp_entities(self, entities: list[dict]) -> None:
self.dlp_entities = entities[: self._max_dlp_entities]
|