azure-ai-evaluation 1.8.0__py3-none-any.whl → 1.10.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of azure-ai-evaluation might be problematic. Click here for more details.
- azure/ai/evaluation/__init__.py +51 -6
- azure/ai/evaluation/_aoai/__init__.py +1 -1
- azure/ai/evaluation/_aoai/aoai_grader.py +21 -11
- azure/ai/evaluation/_aoai/label_grader.py +3 -2
- azure/ai/evaluation/_aoai/python_grader.py +84 -0
- azure/ai/evaluation/_aoai/score_model_grader.py +91 -0
- azure/ai/evaluation/_aoai/string_check_grader.py +3 -2
- azure/ai/evaluation/_aoai/text_similarity_grader.py +3 -2
- azure/ai/evaluation/_azure/_envs.py +9 -10
- azure/ai/evaluation/_azure/_token_manager.py +7 -1
- azure/ai/evaluation/_common/constants.py +11 -2
- azure/ai/evaluation/_common/evaluation_onedp_client.py +32 -26
- azure/ai/evaluation/_common/onedp/__init__.py +32 -32
- azure/ai/evaluation/_common/onedp/_client.py +136 -139
- azure/ai/evaluation/_common/onedp/_configuration.py +70 -73
- azure/ai/evaluation/_common/onedp/_patch.py +21 -21
- azure/ai/evaluation/_common/onedp/_utils/__init__.py +6 -0
- azure/ai/evaluation/_common/onedp/_utils/model_base.py +1232 -0
- azure/ai/evaluation/_common/onedp/_utils/serialization.py +2032 -0
- azure/ai/evaluation/_common/onedp/_validation.py +50 -50
- azure/ai/evaluation/_common/onedp/_version.py +9 -9
- azure/ai/evaluation/_common/onedp/aio/__init__.py +29 -29
- azure/ai/evaluation/_common/onedp/aio/_client.py +138 -143
- azure/ai/evaluation/_common/onedp/aio/_configuration.py +70 -75
- azure/ai/evaluation/_common/onedp/aio/_patch.py +21 -21
- azure/ai/evaluation/_common/onedp/aio/operations/__init__.py +37 -39
- azure/ai/evaluation/_common/onedp/aio/operations/_operations.py +4832 -4494
- azure/ai/evaluation/_common/onedp/aio/operations/_patch.py +21 -21
- azure/ai/evaluation/_common/onedp/models/__init__.py +168 -142
- azure/ai/evaluation/_common/onedp/models/_enums.py +230 -162
- azure/ai/evaluation/_common/onedp/models/_models.py +2685 -2228
- azure/ai/evaluation/_common/onedp/models/_patch.py +21 -21
- azure/ai/evaluation/_common/onedp/operations/__init__.py +37 -39
- azure/ai/evaluation/_common/onedp/operations/_operations.py +6106 -5657
- azure/ai/evaluation/_common/onedp/operations/_patch.py +21 -21
- azure/ai/evaluation/_common/rai_service.py +88 -52
- azure/ai/evaluation/_common/raiclient/__init__.py +1 -1
- azure/ai/evaluation/_common/raiclient/operations/_operations.py +14 -1
- azure/ai/evaluation/_common/utils.py +188 -10
- azure/ai/evaluation/_constants.py +2 -1
- azure/ai/evaluation/_converters/__init__.py +1 -1
- azure/ai/evaluation/_converters/_ai_services.py +9 -8
- azure/ai/evaluation/_converters/_models.py +46 -0
- azure/ai/evaluation/_converters/_sk_services.py +495 -0
- azure/ai/evaluation/_eval_mapping.py +2 -2
- azure/ai/evaluation/_evaluate/_batch_run/_run_submitter_client.py +73 -25
- azure/ai/evaluation/_evaluate/_batch_run/eval_run_context.py +2 -2
- azure/ai/evaluation/_evaluate/_evaluate.py +210 -94
- azure/ai/evaluation/_evaluate/_evaluate_aoai.py +132 -89
- azure/ai/evaluation/_evaluate/_telemetry/__init__.py +0 -1
- azure/ai/evaluation/_evaluate/_utils.py +25 -17
- azure/ai/evaluation/_evaluators/_bleu/_bleu.py +4 -4
- azure/ai/evaluation/_evaluators/_code_vulnerability/_code_vulnerability.py +20 -12
- azure/ai/evaluation/_evaluators/_coherence/_coherence.py +6 -6
- azure/ai/evaluation/_evaluators/_common/_base_eval.py +45 -11
- azure/ai/evaluation/_evaluators/_common/_base_prompty_eval.py +24 -9
- azure/ai/evaluation/_evaluators/_common/_base_rai_svc_eval.py +24 -9
- azure/ai/evaluation/_evaluators/_content_safety/_content_safety.py +28 -18
- azure/ai/evaluation/_evaluators/_content_safety/_hate_unfairness.py +11 -8
- azure/ai/evaluation/_evaluators/_content_safety/_self_harm.py +11 -8
- azure/ai/evaluation/_evaluators/_content_safety/_sexual.py +12 -9
- azure/ai/evaluation/_evaluators/_content_safety/_violence.py +10 -7
- azure/ai/evaluation/_evaluators/_document_retrieval/__init__.py +1 -5
- azure/ai/evaluation/_evaluators/_document_retrieval/_document_retrieval.py +37 -64
- azure/ai/evaluation/_evaluators/_eci/_eci.py +6 -3
- azure/ai/evaluation/_evaluators/_f1_score/_f1_score.py +5 -5
- azure/ai/evaluation/_evaluators/_fluency/_fluency.py +3 -3
- azure/ai/evaluation/_evaluators/_gleu/_gleu.py +4 -4
- azure/ai/evaluation/_evaluators/_groundedness/_groundedness.py +12 -8
- azure/ai/evaluation/_evaluators/_intent_resolution/_intent_resolution.py +31 -26
- azure/ai/evaluation/_evaluators/_intent_resolution/intent_resolution.prompty +210 -96
- azure/ai/evaluation/_evaluators/_meteor/_meteor.py +3 -4
- azure/ai/evaluation/_evaluators/_protected_material/_protected_material.py +14 -7
- azure/ai/evaluation/_evaluators/_qa/_qa.py +5 -5
- azure/ai/evaluation/_evaluators/_relevance/_relevance.py +62 -15
- azure/ai/evaluation/_evaluators/_relevance/relevance.prompty +140 -59
- azure/ai/evaluation/_evaluators/_response_completeness/_response_completeness.py +21 -26
- azure/ai/evaluation/_evaluators/_retrieval/_retrieval.py +5 -5
- azure/ai/evaluation/_evaluators/_rouge/_rouge.py +22 -22
- azure/ai/evaluation/_evaluators/_service_groundedness/_service_groundedness.py +7 -6
- azure/ai/evaluation/_evaluators/_similarity/_similarity.py +4 -4
- azure/ai/evaluation/_evaluators/_task_adherence/_task_adherence.py +27 -24
- azure/ai/evaluation/_evaluators/_task_adherence/task_adherence.prompty +354 -66
- azure/ai/evaluation/_evaluators/_tool_call_accuracy/_tool_call_accuracy.py +175 -183
- azure/ai/evaluation/_evaluators/_tool_call_accuracy/tool_call_accuracy.prompty +99 -21
- azure/ai/evaluation/_evaluators/_ungrounded_attributes/_ungrounded_attributes.py +20 -12
- azure/ai/evaluation/_evaluators/_xpia/xpia.py +10 -7
- azure/ai/evaluation/_exceptions.py +10 -0
- azure/ai/evaluation/_http_utils.py +3 -3
- azure/ai/evaluation/_legacy/_batch_engine/_config.py +6 -3
- azure/ai/evaluation/_legacy/_batch_engine/_engine.py +117 -32
- azure/ai/evaluation/_legacy/_batch_engine/_openai_injector.py +5 -2
- azure/ai/evaluation/_legacy/_batch_engine/_result.py +2 -0
- azure/ai/evaluation/_legacy/_batch_engine/_run.py +2 -2
- azure/ai/evaluation/_legacy/_batch_engine/_run_submitter.py +33 -41
- azure/ai/evaluation/_legacy/_batch_engine/_utils.py +1 -4
- azure/ai/evaluation/_legacy/_common/_async_token_provider.py +12 -19
- azure/ai/evaluation/_legacy/_common/_thread_pool_executor_with_context.py +2 -0
- azure/ai/evaluation/_legacy/prompty/_prompty.py +11 -5
- azure/ai/evaluation/_safety_evaluation/__init__.py +1 -1
- azure/ai/evaluation/_safety_evaluation/_safety_evaluation.py +195 -111
- azure/ai/evaluation/_user_agent.py +32 -1
- azure/ai/evaluation/_version.py +1 -1
- azure/ai/evaluation/red_team/__init__.py +3 -1
- azure/ai/evaluation/red_team/_agent/__init__.py +1 -1
- azure/ai/evaluation/red_team/_agent/_agent_functions.py +68 -71
- azure/ai/evaluation/red_team/_agent/_agent_tools.py +103 -145
- azure/ai/evaluation/red_team/_agent/_agent_utils.py +26 -6
- azure/ai/evaluation/red_team/_agent/_semantic_kernel_plugin.py +62 -71
- azure/ai/evaluation/red_team/_attack_objective_generator.py +94 -52
- azure/ai/evaluation/red_team/_attack_strategy.py +2 -1
- azure/ai/evaluation/red_team/_callback_chat_target.py +4 -9
- azure/ai/evaluation/red_team/_default_converter.py +1 -1
- azure/ai/evaluation/red_team/_red_team.py +1947 -1040
- azure/ai/evaluation/red_team/_red_team_result.py +49 -38
- azure/ai/evaluation/red_team/_utils/__init__.py +1 -1
- azure/ai/evaluation/red_team/_utils/_rai_service_eval_chat_target.py +39 -34
- azure/ai/evaluation/red_team/_utils/_rai_service_target.py +163 -138
- azure/ai/evaluation/red_team/_utils/_rai_service_true_false_scorer.py +14 -14
- azure/ai/evaluation/red_team/_utils/constants.py +1 -13
- azure/ai/evaluation/red_team/_utils/formatting_utils.py +41 -44
- azure/ai/evaluation/red_team/_utils/logging_utils.py +17 -17
- azure/ai/evaluation/red_team/_utils/metric_mapping.py +31 -4
- azure/ai/evaluation/red_team/_utils/strategy_utils.py +33 -25
- azure/ai/evaluation/simulator/_adversarial_scenario.py +2 -0
- azure/ai/evaluation/simulator/_adversarial_simulator.py +31 -17
- azure/ai/evaluation/simulator/_conversation/__init__.py +2 -2
- azure/ai/evaluation/simulator/_direct_attack_simulator.py +8 -8
- azure/ai/evaluation/simulator/_indirect_attack_simulator.py +18 -6
- azure/ai/evaluation/simulator/_model_tools/_generated_rai_client.py +54 -24
- azure/ai/evaluation/simulator/_model_tools/_identity_manager.py +7 -1
- azure/ai/evaluation/simulator/_model_tools/_proxy_completion_model.py +30 -10
- azure/ai/evaluation/simulator/_model_tools/_rai_client.py +19 -31
- azure/ai/evaluation/simulator/_model_tools/_template_handler.py +20 -6
- azure/ai/evaluation/simulator/_model_tools/models.py +1 -1
- azure/ai/evaluation/simulator/_simulator.py +21 -8
- {azure_ai_evaluation-1.8.0.dist-info → azure_ai_evaluation-1.10.0.dist-info}/METADATA +46 -3
- {azure_ai_evaluation-1.8.0.dist-info → azure_ai_evaluation-1.10.0.dist-info}/RECORD +141 -136
- azure/ai/evaluation/_common/onedp/aio/_vendor.py +0 -40
- {azure_ai_evaluation-1.8.0.dist-info → azure_ai_evaluation-1.10.0.dist-info}/NOTICE.txt +0 -0
- {azure_ai_evaluation-1.8.0.dist-info → azure_ai_evaluation-1.10.0.dist-info}/WHEEL +0 -0
- {azure_ai_evaluation-1.8.0.dist-info → azure_ai_evaluation-1.10.0.dist-info}/top_level.txt +0 -0
|
@@ -49,8 +49,8 @@ class AdversarialSimulator:
|
|
|
49
49
|
"""
|
|
50
50
|
Initializes the adversarial simulator with a project scope.
|
|
51
51
|
|
|
52
|
-
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
53
|
-
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
52
|
+
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
53
|
+
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
54
54
|
:type azure_ai_project: Union[str, AzureAIProject]
|
|
55
55
|
:param credential: The credential for connecting to Azure AI project.
|
|
56
56
|
:type credential: ~azure.core.credentials.TokenCredential
|
|
@@ -77,7 +77,7 @@ class AdversarialSimulator:
|
|
|
77
77
|
logger=logging.getLogger("AdversarialSimulator"),
|
|
78
78
|
credential=self.credential,
|
|
79
79
|
)
|
|
80
|
-
self.rai_client
|
|
80
|
+
self.rai_client = AIProjectClient(endpoint=azure_ai_project, credential=credential)
|
|
81
81
|
else:
|
|
82
82
|
try:
|
|
83
83
|
self.azure_ai_project = validate_azure_ai_project(azure_ai_project)
|
|
@@ -96,7 +96,7 @@ class AdversarialSimulator:
|
|
|
96
96
|
credential=self.credential,
|
|
97
97
|
)
|
|
98
98
|
self.rai_client = RAIClient(azure_ai_project=self.azure_ai_project, token_manager=self.token_manager)
|
|
99
|
-
|
|
99
|
+
|
|
100
100
|
self.adversarial_template_handler = AdversarialTemplateHandler(
|
|
101
101
|
azure_ai_project=self.azure_ai_project, rai_client=self.rai_client
|
|
102
102
|
)
|
|
@@ -239,8 +239,11 @@ class AdversarialSimulator:
|
|
|
239
239
|
# So randomize a the selection instead of the parameter list directly,
|
|
240
240
|
# or a potentially large deep copy.
|
|
241
241
|
if randomization_seed is not None:
|
|
242
|
-
random
|
|
243
|
-
|
|
242
|
+
# Create a local random instance to avoid polluting global state
|
|
243
|
+
local_random = random.Random(randomization_seed)
|
|
244
|
+
local_random.shuffle(templates)
|
|
245
|
+
else:
|
|
246
|
+
random.shuffle(templates)
|
|
244
247
|
|
|
245
248
|
# Prepare task parameters based on scenario - but use a single append call for all scenarios
|
|
246
249
|
tasks = []
|
|
@@ -264,7 +267,9 @@ class AdversarialSimulator:
|
|
|
264
267
|
|
|
265
268
|
# Limit to max_simulation_results if needed
|
|
266
269
|
if len(template_parameter_pairs) > max_simulation_results:
|
|
267
|
-
template_parameter_pairs = template_parameter_pairs[
|
|
270
|
+
template_parameter_pairs = template_parameter_pairs[
|
|
271
|
+
:max_simulation_results
|
|
272
|
+
] # Create a seeded random instance for jailbreak selection if randomization_seed is provided
|
|
268
273
|
jailbreak_random = None
|
|
269
274
|
if _jailbreak_type == "upia" and randomization_seed is not None:
|
|
270
275
|
jailbreak_random = random.Random(randomization_seed)
|
|
@@ -364,10 +369,21 @@ class AdversarialSimulator:
|
|
|
364
369
|
target=target, role=ConversationRole.ASSISTANT, template=template, parameters=parameters, scenario=scenario
|
|
365
370
|
)
|
|
366
371
|
bots = [user_bot, system_bot]
|
|
367
|
-
|
|
372
|
+
|
|
373
|
+
async def run_simulation(session_obj):
|
|
374
|
+
async with semaphore:
|
|
375
|
+
_, conversation_history = await simulate_conversation(
|
|
376
|
+
bots=bots,
|
|
377
|
+
session=session_obj,
|
|
378
|
+
turn_limit=max_conversation_turns,
|
|
379
|
+
api_call_delay_sec=api_call_delay_sec,
|
|
380
|
+
language=language,
|
|
381
|
+
)
|
|
382
|
+
return conversation_history
|
|
383
|
+
|
|
368
384
|
if isinstance(self.rai_client, AIProjectClient):
|
|
369
385
|
session = self.rai_client
|
|
370
|
-
else:
|
|
386
|
+
else:
|
|
371
387
|
session = get_async_http_client().with_policies(
|
|
372
388
|
retry_policy=AsyncRetryPolicy(
|
|
373
389
|
retry_total=api_call_retry_limit,
|
|
@@ -375,13 +391,7 @@ class AdversarialSimulator:
|
|
|
375
391
|
retry_mode=RetryMode.Fixed,
|
|
376
392
|
)
|
|
377
393
|
)
|
|
378
|
-
|
|
379
|
-
bots=bots,
|
|
380
|
-
session=session,
|
|
381
|
-
turn_limit=max_conversation_turns,
|
|
382
|
-
api_call_delay_sec=api_call_delay_sec,
|
|
383
|
-
language=language,
|
|
384
|
-
)
|
|
394
|
+
conversation_history = await run_simulation(session)
|
|
385
395
|
|
|
386
396
|
return self._to_chat_protocol(
|
|
387
397
|
conversation_history=conversation_history,
|
|
@@ -391,7 +401,11 @@ class AdversarialSimulator:
|
|
|
391
401
|
def _get_user_proxy_completion_model(
|
|
392
402
|
self, template_key: str, template_parameters: TemplateParameters, simulation_id: str = ""
|
|
393
403
|
) -> ProxyChatCompletionsModel:
|
|
394
|
-
endpoint_url =
|
|
404
|
+
endpoint_url = (
|
|
405
|
+
self.rai_client._config.endpoint + "/redTeams/simulation/chat/completions/submit"
|
|
406
|
+
if isinstance(self.rai_client, AIProjectClient)
|
|
407
|
+
else self.rai_client.simulation_submit_endpoint
|
|
408
|
+
)
|
|
395
409
|
return ProxyChatCompletionsModel(
|
|
396
410
|
name="raisvc_proxy_model",
|
|
397
411
|
template_key=template_key,
|
|
@@ -419,13 +419,13 @@ class MultiModalConversationBot(ConversationBot):
|
|
|
419
419
|
contents = []
|
|
420
420
|
for msg in messages:
|
|
421
421
|
if msg.startswith("image_understanding/"):
|
|
422
|
-
if
|
|
422
|
+
if isinstance(self.rai_client, RAIClient):
|
|
423
423
|
encoded_image = await self.rai_client.get_image_data(msg)
|
|
424
424
|
else:
|
|
425
425
|
response = self.rai_client.red_teams.get_template_parameters_image(path=msg, stream="true")
|
|
426
426
|
image_data = b"".join(response)
|
|
427
427
|
encoded_image = base64.b64encode(image_data).decode("utf-8")
|
|
428
|
-
|
|
428
|
+
|
|
429
429
|
contents.append(
|
|
430
430
|
{"type": "image_url", "image_url": {"url": f"data:image/png;base64,{encoded_image}"}},
|
|
431
431
|
)
|
|
@@ -28,8 +28,8 @@ class DirectAttackSimulator:
|
|
|
28
28
|
Initialize a UPIA (user prompt injected attack) jailbreak adversarial simulator with a project scope.
|
|
29
29
|
This simulator converses with your AI system using prompts designed to interrupt normal functionality.
|
|
30
30
|
|
|
31
|
-
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
32
|
-
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
31
|
+
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
32
|
+
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
33
33
|
:type azure_ai_project: Union[str, AzureAIProject]
|
|
34
34
|
:param credential: The credential for connecting to Azure AI project.
|
|
35
35
|
:type credential: ~azure.core.credentials.TokenCredential
|
|
@@ -46,16 +46,16 @@ class DirectAttackSimulator:
|
|
|
46
46
|
|
|
47
47
|
def __init__(self, *, azure_ai_project: Union[str, AzureAIProject], credential: TokenCredential):
|
|
48
48
|
"""Constructor."""
|
|
49
|
-
|
|
49
|
+
|
|
50
50
|
if is_onedp_project(azure_ai_project):
|
|
51
51
|
self.azure_ai_project = azure_ai_project
|
|
52
|
-
self.credential=cast(TokenCredential, credential)
|
|
52
|
+
self.credential = cast(TokenCredential, credential)
|
|
53
53
|
self.token_manager = ManagedIdentityAPITokenManager(
|
|
54
54
|
token_scope=TokenScope.COGNITIVE_SERVICES_MANAGEMENT,
|
|
55
55
|
logger=logging.getLogger("AdversarialSimulator"),
|
|
56
|
-
credential=self.credential
|
|
56
|
+
credential=self.credential,
|
|
57
57
|
)
|
|
58
|
-
self.rai_client
|
|
58
|
+
self.rai_client = AIProjectClient(endpoint=azure_ai_project, credential=credential)
|
|
59
59
|
else:
|
|
60
60
|
try:
|
|
61
61
|
self.azure_ai_project = validate_azure_ai_project(azure_ai_project)
|
|
@@ -67,14 +67,14 @@ class DirectAttackSimulator:
|
|
|
67
67
|
category=e.category,
|
|
68
68
|
blame=e.blame,
|
|
69
69
|
) from e
|
|
70
|
-
self.credential = cast(TokenCredential, credential)
|
|
70
|
+
self.credential = cast(TokenCredential, credential)
|
|
71
71
|
self.token_manager = ManagedIdentityAPITokenManager(
|
|
72
72
|
token_scope=TokenScope.DEFAULT_AZURE_MANAGEMENT,
|
|
73
73
|
logger=logging.getLogger("AdversarialSimulator"),
|
|
74
74
|
credential=self.credential,
|
|
75
75
|
)
|
|
76
76
|
self.rai_client = RAIClient(azure_ai_project=self.azure_ai_project, token_manager=self.token_manager)
|
|
77
|
-
|
|
77
|
+
|
|
78
78
|
self.adversarial_template_handler = AdversarialTemplateHandler(
|
|
79
79
|
azure_ai_project=self.azure_ai_project, rai_client=self.rai_client
|
|
80
80
|
)
|
|
@@ -5,7 +5,8 @@
|
|
|
5
5
|
# noqa: E501
|
|
6
6
|
import asyncio
|
|
7
7
|
import logging
|
|
8
|
-
|
|
8
|
+
import random
|
|
9
|
+
from typing import Callable, cast, Union, Optional
|
|
9
10
|
|
|
10
11
|
from tqdm import tqdm
|
|
11
12
|
|
|
@@ -30,8 +31,8 @@ class IndirectAttackSimulator(AdversarialSimulator):
|
|
|
30
31
|
"""
|
|
31
32
|
Initializes the XPIA (cross domain prompt injected attack) jailbreak adversarial simulator with a project scope.
|
|
32
33
|
|
|
33
|
-
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
34
|
-
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
34
|
+
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
35
|
+
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
35
36
|
:type azure_ai_project: Union[str, AzureAIProject]
|
|
36
37
|
:param credential: The credential for connecting to Azure AI project.
|
|
37
38
|
:type credential: ~azure.core.credentials.TokenCredential
|
|
@@ -51,13 +52,13 @@ class IndirectAttackSimulator(AdversarialSimulator):
|
|
|
51
52
|
|
|
52
53
|
if is_onedp_project(azure_ai_project):
|
|
53
54
|
self.azure_ai_project = azure_ai_project
|
|
54
|
-
self.credential=cast(TokenCredential, credential)
|
|
55
|
+
self.credential = cast(TokenCredential, credential)
|
|
55
56
|
self.token_manager = ManagedIdentityAPITokenManager(
|
|
56
57
|
token_scope=TokenScope.COGNITIVE_SERVICES_MANAGEMENT,
|
|
57
58
|
logger=logging.getLogger("AdversarialSimulator"),
|
|
58
|
-
credential=self.credential
|
|
59
|
+
credential=self.credential,
|
|
59
60
|
)
|
|
60
|
-
self.rai_client
|
|
61
|
+
self.rai_client = AIProjectClient(endpoint=azure_ai_project, credential=credential)
|
|
61
62
|
self.adversarial_template_handler = AdversarialTemplateHandler(
|
|
62
63
|
azure_ai_project=self.azure_ai_project, rai_client=self.rai_client
|
|
63
64
|
)
|
|
@@ -105,6 +106,7 @@ class IndirectAttackSimulator(AdversarialSimulator):
|
|
|
105
106
|
api_call_retry_sleep_sec: int = 1,
|
|
106
107
|
api_call_delay_sec: int = 0,
|
|
107
108
|
concurrent_async_task: int = 3,
|
|
109
|
+
randomization_seed: Optional[int] = None,
|
|
108
110
|
**kwargs,
|
|
109
111
|
):
|
|
110
112
|
"""
|
|
@@ -130,6 +132,9 @@ class IndirectAttackSimulator(AdversarialSimulator):
|
|
|
130
132
|
:keyword concurrent_async_task: The number of asynchronous tasks to run concurrently during the simulation.
|
|
131
133
|
Defaults to 3.
|
|
132
134
|
:paramtype concurrent_async_task: int
|
|
135
|
+
:keyword randomization_seed: The seed used to randomize prompt selection. If unset, the system's
|
|
136
|
+
default seed is used. Defaults to None.
|
|
137
|
+
:paramtype randomization_seed: Optional[int]
|
|
133
138
|
:return: A list of dictionaries, each representing a simulated conversation. Each dictionary contains:
|
|
134
139
|
|
|
135
140
|
- 'template_parameters': A dictionary with parameters used in the conversation template,
|
|
@@ -190,6 +195,13 @@ class IndirectAttackSimulator(AdversarialSimulator):
|
|
|
190
195
|
ncols=100,
|
|
191
196
|
unit="simulations",
|
|
192
197
|
)
|
|
198
|
+
|
|
199
|
+
# Apply randomization to templates if seed is provided
|
|
200
|
+
if randomization_seed is not None:
|
|
201
|
+
# Create a local random instance to avoid polluting global state
|
|
202
|
+
local_random = random.Random(randomization_seed)
|
|
203
|
+
local_random.shuffle(templates)
|
|
204
|
+
|
|
193
205
|
for template in templates:
|
|
194
206
|
for parameter in template.template_parameters:
|
|
195
207
|
tasks.append(
|
|
@@ -6,6 +6,7 @@ import os
|
|
|
6
6
|
from typing import Dict, List, Optional, Union
|
|
7
7
|
|
|
8
8
|
from azure.core.credentials import TokenCredential
|
|
9
|
+
from azure.core.pipeline.policies import UserAgentPolicy
|
|
9
10
|
from azure.ai.evaluation._model_configurations import AzureAIProject
|
|
10
11
|
from azure.ai.evaluation.simulator._model_tools import ManagedIdentityAPITokenManager
|
|
11
12
|
from azure.ai.evaluation._common.raiclient import MachineLearningServicesClient
|
|
@@ -13,30 +14,35 @@ from azure.ai.evaluation._constants import TokenScope
|
|
|
13
14
|
from azure.ai.evaluation._common.utils import is_onedp_project
|
|
14
15
|
from azure.ai.evaluation._common.onedp import AIProjectClient
|
|
15
16
|
from azure.ai.evaluation._common import EvaluationServiceOneDPClient
|
|
17
|
+
from azure.ai.evaluation._user_agent import UserAgentSingleton
|
|
16
18
|
import jwt
|
|
17
19
|
import time
|
|
18
20
|
import ast
|
|
19
21
|
|
|
22
|
+
|
|
20
23
|
class GeneratedRAIClient:
|
|
21
24
|
"""Client for the Responsible AI Service using the auto-generated MachineLearningServicesClient.
|
|
22
|
-
|
|
23
|
-
:param azure_ai_project: The
|
|
24
|
-
|
|
25
|
+
|
|
26
|
+
:param azure_ai_project: The Azure AI project, which can either be a string representing the project endpoint
|
|
27
|
+
or an instance of AzureAIProject. It contains subscription id, resource group, and project name.
|
|
28
|
+
:type azure_ai_project: Union[str, ~azure.ai.evaluation.AzureAIProject]
|
|
25
29
|
:param token_manager: The token manager
|
|
26
30
|
:type token_manager: ~azure.ai.evaluation.simulator._model_tools._identity_manager.APITokenManager
|
|
27
31
|
"""
|
|
28
|
-
|
|
32
|
+
|
|
29
33
|
def __init__(self, azure_ai_project: Union[AzureAIProject, str], token_manager: ManagedIdentityAPITokenManager):
|
|
30
34
|
self.azure_ai_project = azure_ai_project
|
|
31
35
|
self.token_manager = token_manager
|
|
32
|
-
|
|
36
|
+
|
|
37
|
+
user_agent_policy = UserAgentPolicy(base_user_agent=UserAgentSingleton().value)
|
|
38
|
+
|
|
33
39
|
if not is_onedp_project(azure_ai_project):
|
|
34
40
|
# Service URL construction
|
|
35
41
|
if "RAI_SVC_URL" in os.environ:
|
|
36
42
|
endpoint = os.environ["RAI_SVC_URL"].rstrip("/")
|
|
37
43
|
else:
|
|
38
44
|
endpoint = self._get_service_discovery_url()
|
|
39
|
-
|
|
45
|
+
|
|
40
46
|
# Create the autogenerated client
|
|
41
47
|
self._client = MachineLearningServicesClient(
|
|
42
48
|
endpoint=endpoint,
|
|
@@ -46,20 +52,24 @@ class GeneratedRAIClient:
|
|
|
46
52
|
credential=self.token_manager,
|
|
47
53
|
).rai_svc
|
|
48
54
|
else:
|
|
49
|
-
self._client = AIProjectClient(
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
55
|
+
self._client = AIProjectClient(
|
|
56
|
+
endpoint=azure_ai_project, credential=token_manager, user_agent_policy=user_agent_policy
|
|
57
|
+
).red_teams
|
|
58
|
+
self._evaluation_onedp_client = EvaluationServiceOneDPClient(
|
|
59
|
+
endpoint=azure_ai_project, credential=token_manager, user_agent_policy=user_agent_policy
|
|
60
|
+
)
|
|
61
|
+
|
|
53
62
|
def _get_service_discovery_url(self):
|
|
54
63
|
"""Get the service discovery URL.
|
|
55
|
-
|
|
64
|
+
|
|
56
65
|
:return: The service discovery URL
|
|
57
66
|
:rtype: str
|
|
58
67
|
"""
|
|
59
68
|
import requests
|
|
69
|
+
|
|
60
70
|
bearer_token = self._fetch_or_reuse_token(self.token_manager)
|
|
61
71
|
headers = {"Authorization": f"Bearer {bearer_token}", "Content-Type": "application/json"}
|
|
62
|
-
|
|
72
|
+
|
|
63
73
|
response = requests.get(
|
|
64
74
|
f"https://management.azure.com/subscriptions/{self.azure_ai_project['subscription_id']}/"
|
|
65
75
|
f"resourceGroups/{self.azure_ai_project['resource_group_name']}/"
|
|
@@ -68,7 +78,7 @@ class GeneratedRAIClient:
|
|
|
68
78
|
headers=headers,
|
|
69
79
|
timeout=5,
|
|
70
80
|
)
|
|
71
|
-
|
|
81
|
+
|
|
72
82
|
if response.status_code != 200:
|
|
73
83
|
msg = (
|
|
74
84
|
f"Failed to connect to your Azure AI project. Please check if the project scope is configured "
|
|
@@ -79,51 +89,71 @@ class GeneratedRAIClient:
|
|
|
79
89
|
|
|
80
90
|
# Parse the discovery URL
|
|
81
91
|
from urllib.parse import urlparse
|
|
92
|
+
|
|
82
93
|
base_url = urlparse(response.json()["properties"]["discoveryUrl"])
|
|
83
94
|
return f"{base_url.scheme}://{base_url.netloc}"
|
|
84
|
-
|
|
85
|
-
async def get_attack_objectives(
|
|
95
|
+
|
|
96
|
+
async def get_attack_objectives(
|
|
97
|
+
self,
|
|
98
|
+
*,
|
|
99
|
+
risk_type: Optional[str] = None,
|
|
100
|
+
risk_category: Optional[str] = None,
|
|
101
|
+
application_scenario: str = None,
|
|
102
|
+
strategy: Optional[str] = None,
|
|
103
|
+
scan_session_id: Optional[str] = None,
|
|
104
|
+
) -> Dict:
|
|
86
105
|
"""Get attack objectives using the auto-generated operations.
|
|
87
|
-
|
|
106
|
+
|
|
107
|
+
:param risk_type: Optional risk type to filter the attack objectives
|
|
108
|
+
:type risk_type: Optional[str]
|
|
88
109
|
:param risk_category: Optional risk category to filter the attack objectives
|
|
89
110
|
:type risk_category: Optional[str]
|
|
90
111
|
:param application_scenario: Optional description of the application scenario for context
|
|
91
112
|
:type application_scenario: str
|
|
92
113
|
:param strategy: Optional strategy to filter the attack objectives
|
|
93
114
|
:type strategy: Optional[str]
|
|
115
|
+
:param scan_session_id: Optional unique session ID for the scan
|
|
116
|
+
:type scan_session_id: Optional[str]
|
|
94
117
|
:return: The attack objectives
|
|
95
118
|
:rtype: Dict
|
|
96
|
-
"""
|
|
119
|
+
"""
|
|
97
120
|
try:
|
|
98
121
|
# Send the request using the autogenerated client
|
|
99
122
|
response = self._client.get_attack_objectives(
|
|
100
|
-
risk_types=[
|
|
123
|
+
risk_types=[risk_type],
|
|
124
|
+
risk_category=risk_category,
|
|
101
125
|
lang="en",
|
|
102
126
|
strategy=strategy,
|
|
127
|
+
headers={"x-ms-client-request-id": scan_session_id},
|
|
103
128
|
)
|
|
104
129
|
return response
|
|
105
|
-
|
|
130
|
+
|
|
106
131
|
except Exception as e:
|
|
107
132
|
# Log the exception for debugging purposes
|
|
108
133
|
import logging
|
|
134
|
+
|
|
109
135
|
logging.error(f"Error in get_attack_objectives: {str(e)}")
|
|
110
136
|
raise
|
|
111
|
-
|
|
112
|
-
async def get_jailbreak_prefixes(self) -> List[str]:
|
|
137
|
+
|
|
138
|
+
async def get_jailbreak_prefixes(self, scan_session_id: Optional[str] = None) -> List[str]:
|
|
113
139
|
"""Get jailbreak prefixes using the auto-generated operations.
|
|
114
|
-
|
|
140
|
+
|
|
141
|
+
:param scan_session_id: Optional unique session ID for the scan
|
|
142
|
+
:type scan_session_id: Optional[str]
|
|
115
143
|
:return: The jailbreak prefixes
|
|
116
144
|
:rtype: List[str]
|
|
117
145
|
"""
|
|
118
146
|
try:
|
|
119
147
|
# Send the request using the autogenerated client
|
|
120
|
-
response = self._client.get_jail_break_dataset_with_type(
|
|
148
|
+
response = self._client.get_jail_break_dataset_with_type(
|
|
149
|
+
type="upia", headers={"x-ms-client-request-id": scan_session_id}
|
|
150
|
+
)
|
|
121
151
|
if isinstance(response, list):
|
|
122
152
|
return response
|
|
123
153
|
else:
|
|
124
154
|
self.logger.error("Unexpected response format from get_jail_break_dataset_with_type")
|
|
125
155
|
raise ValueError("Unexpected response format from get_jail_break_dataset_with_type")
|
|
126
|
-
|
|
156
|
+
|
|
127
157
|
except Exception as e:
|
|
128
158
|
return [""]
|
|
129
159
|
|
|
@@ -78,7 +78,13 @@ class APITokenManager(ABC):
|
|
|
78
78
|
|
|
79
79
|
@abstractmethod
|
|
80
80
|
def get_token(
|
|
81
|
-
|
|
81
|
+
self,
|
|
82
|
+
scopes: Union[str, None] = None,
|
|
83
|
+
claims: Union[str, None] = None,
|
|
84
|
+
tenant_id: Union[str, None] = None,
|
|
85
|
+
enable_cae: bool = False,
|
|
86
|
+
**kwargs: Any,
|
|
87
|
+
) -> AccessToken:
|
|
82
88
|
"""Async method to get the API token. Subclasses should implement this method.
|
|
83
89
|
|
|
84
90
|
:return: API token
|
|
@@ -9,8 +9,8 @@ import uuid
|
|
|
9
9
|
from typing import Any, Dict, List, Optional, cast, Union
|
|
10
10
|
|
|
11
11
|
from azure.ai.evaluation._http_utils import AsyncHttpPipeline, get_async_http_client
|
|
12
|
-
from azure.ai.evaluation._user_agent import
|
|
13
|
-
from azure.core.exceptions import HttpResponseError
|
|
12
|
+
from azure.ai.evaluation._user_agent import UserAgentSingleton
|
|
13
|
+
from azure.core.exceptions import HttpResponseError, ServiceResponseError
|
|
14
14
|
from azure.core.pipeline.policies import AsyncRetryPolicy, RetryMode
|
|
15
15
|
from azure.ai.evaluation._common.onedp._client import AIProjectClient
|
|
16
16
|
from azure.ai.evaluation._common.onedp.models import SimulationDTO
|
|
@@ -166,7 +166,7 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
166
166
|
proxy_headers = {
|
|
167
167
|
"Authorization": f"Bearer {token}",
|
|
168
168
|
"Content-Type": "application/json",
|
|
169
|
-
"User-Agent":
|
|
169
|
+
"User-Agent": UserAgentSingleton().value,
|
|
170
170
|
}
|
|
171
171
|
|
|
172
172
|
headers = {
|
|
@@ -193,7 +193,7 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
193
193
|
time_start = time.time()
|
|
194
194
|
full_response = None
|
|
195
195
|
|
|
196
|
-
if
|
|
196
|
+
if isinstance(session, AIProjectClient):
|
|
197
197
|
sim_request_dto = SimulationDTO(
|
|
198
198
|
headers=headers,
|
|
199
199
|
params=params,
|
|
@@ -203,7 +203,7 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
203
203
|
)
|
|
204
204
|
response_data = session.red_teams.submit_simulation(sim_request_dto, headers=headers, params=params)
|
|
205
205
|
operation_id = response_data["location"].split("/")[-1]
|
|
206
|
-
|
|
206
|
+
|
|
207
207
|
request_count = 0
|
|
208
208
|
flag = True
|
|
209
209
|
while flag:
|
|
@@ -211,6 +211,7 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
211
211
|
response = session.evaluations.operation_results(operation_id, headers=headers)
|
|
212
212
|
except Exception as e:
|
|
213
213
|
from types import SimpleNamespace # pylint: disable=forgotten-debug-statement
|
|
214
|
+
|
|
214
215
|
response = SimpleNamespace(status_code=202, text=str(e), json=lambda: {"error": str(e)})
|
|
215
216
|
if isinstance(response, dict):
|
|
216
217
|
response_data = response
|
|
@@ -224,14 +225,33 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
224
225
|
sleep_time = RAIService.SLEEP_TIME**request_count
|
|
225
226
|
await asyncio.sleep(sleep_time)
|
|
226
227
|
else:
|
|
227
|
-
|
|
228
|
+
# Retry policy for POST request to RAI service
|
|
229
|
+
service_call_retry_policy = AsyncRetryPolicy(
|
|
230
|
+
retry_on_exceptions=[ServiceResponseError],
|
|
231
|
+
retry_total=7,
|
|
232
|
+
retry_backoff_factor=10.0,
|
|
233
|
+
retry_backoff_max=180,
|
|
234
|
+
retry_mode=RetryMode.Exponential,
|
|
235
|
+
)
|
|
236
|
+
|
|
237
|
+
response = None
|
|
238
|
+
async with get_async_http_client().with_policies(retry_policy=service_call_retry_policy) as retry_client:
|
|
239
|
+
try:
|
|
240
|
+
response = await retry_client.post(
|
|
241
|
+
url=self.endpoint_url, headers=proxy_headers, json=sim_request_dto.to_dict()
|
|
242
|
+
)
|
|
243
|
+
except ServiceResponseError as e:
|
|
244
|
+
self.logger.error("ServiceResponseError during POST request to rai svc after retries: %s", str(e))
|
|
245
|
+
raise
|
|
246
|
+
|
|
228
247
|
# response.raise_for_status()
|
|
229
248
|
if response.status_code != 202:
|
|
230
249
|
raise HttpResponseError(
|
|
231
|
-
message=f"Received unexpected HTTP status: {response.status_code} {response.text()}",
|
|
250
|
+
message=f"Received unexpected HTTP status: {response.status_code} {response.text()}",
|
|
251
|
+
response=response,
|
|
232
252
|
)
|
|
233
253
|
response_data = response.json()
|
|
234
|
-
|
|
254
|
+
|
|
235
255
|
self.result_url = cast(str, response_data["location"])
|
|
236
256
|
retry_policy = AsyncRetryPolicy( # set up retry configuration
|
|
237
257
|
retry_on_status_codes=[202], # on which statuses to retry
|
|
@@ -252,14 +272,14 @@ class ProxyChatCompletionsModel(OpenAIChatCompletionsModel):
|
|
|
252
272
|
proxy_headers = {
|
|
253
273
|
"Authorization": f"Bearer {token}",
|
|
254
274
|
"Content-Type": "application/json",
|
|
255
|
-
"User-Agent":
|
|
275
|
+
"User-Agent": UserAgentSingleton().value,
|
|
256
276
|
}
|
|
257
277
|
response = await exp_retry_client.get( # pylint: disable=too-many-function-args,unexpected-keyword-arg
|
|
258
278
|
self.result_url, headers=proxy_headers
|
|
259
279
|
)
|
|
260
280
|
response.raise_for_status()
|
|
261
281
|
response_data = response.json()
|
|
262
|
-
|
|
282
|
+
|
|
263
283
|
self.logger.info("Response: %s", response_data)
|
|
264
284
|
|
|
265
285
|
# Copy the full response and return it to be saved in jsonl.
|
|
@@ -10,7 +10,7 @@ import json
|
|
|
10
10
|
from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
|
|
11
11
|
from azure.ai.evaluation._http_utils import AsyncHttpPipeline, get_async_http_client, get_http_client
|
|
12
12
|
from azure.ai.evaluation._model_configurations import AzureAIProject
|
|
13
|
-
from azure.ai.evaluation._user_agent import
|
|
13
|
+
from azure.ai.evaluation._user_agent import UserAgentSingleton
|
|
14
14
|
from azure.core.pipeline.policies import AsyncRetryPolicy, RetryMode
|
|
15
15
|
|
|
16
16
|
from ._identity_manager import APITokenManager
|
|
@@ -148,7 +148,7 @@ class RAIClient: # pylint: disable=client-accepts-api-version-keyword
|
|
|
148
148
|
headers = {
|
|
149
149
|
"Authorization": f"Bearer {token}",
|
|
150
150
|
"Content-Type": "application/json",
|
|
151
|
-
"User-Agent":
|
|
151
|
+
"User-Agent": UserAgentSingleton().value,
|
|
152
152
|
}
|
|
153
153
|
|
|
154
154
|
session = self._create_async_client()
|
|
@@ -184,7 +184,7 @@ class RAIClient: # pylint: disable=client-accepts-api-version-keyword
|
|
|
184
184
|
headers = {
|
|
185
185
|
"Authorization": f"Bearer {token}",
|
|
186
186
|
"Content-Type": "application/json",
|
|
187
|
-
"User-Agent":
|
|
187
|
+
"User-Agent": UserAgentSingleton().value,
|
|
188
188
|
}
|
|
189
189
|
|
|
190
190
|
session = self._create_async_client()
|
|
@@ -209,9 +209,11 @@ class RAIClient: # pylint: disable=client-accepts-api-version-keyword
|
|
|
209
209
|
blame=ErrorBlame.USER_ERROR,
|
|
210
210
|
)
|
|
211
211
|
|
|
212
|
-
async def get_attack_objectives(
|
|
212
|
+
async def get_attack_objectives(
|
|
213
|
+
self, risk_categories: List[str], application_scenario: str = None, strategy: str = None
|
|
214
|
+
) -> Any:
|
|
213
215
|
"""Get the attack objectives based on risk categories and application scenario
|
|
214
|
-
|
|
216
|
+
|
|
215
217
|
:param risk_categories: List of risk categories to generate attack objectives for
|
|
216
218
|
:type risk_categories: List[str]
|
|
217
219
|
:param application_scenario: Optional description of the application scenario for context
|
|
@@ -224,17 +226,17 @@ class RAIClient: # pylint: disable=client-accepts-api-version-keyword
|
|
|
224
226
|
# Create query parameters for the request
|
|
225
227
|
if application_scenario:
|
|
226
228
|
raise NotImplementedError("Application scenario is not supported yet")
|
|
227
|
-
|
|
229
|
+
|
|
228
230
|
params = {
|
|
229
231
|
"api-version": "2022-11-01-preview",
|
|
230
232
|
"riskTypes": ",".join(risk_categories),
|
|
231
|
-
"lang": "en" # Default to English
|
|
233
|
+
"lang": "en", # Default to English
|
|
232
234
|
}
|
|
233
|
-
|
|
235
|
+
|
|
234
236
|
# Add strategy parameter if provided
|
|
235
237
|
if strategy:
|
|
236
238
|
params["strategy"] = strategy
|
|
237
|
-
|
|
239
|
+
|
|
238
240
|
try:
|
|
239
241
|
# Make the request using the existing get method
|
|
240
242
|
result = await self.get(self.attack_objectives_endpoint)
|
|
@@ -254,25 +256,11 @@ class RAIClient: # pylint: disable=client-accepts-api-version-keyword
|
|
|
254
256
|
except Exception:
|
|
255
257
|
# If the API fails or isn't implemented yet, return a mock response
|
|
256
258
|
# This is temporary until the API endpoint is fully implemented
|
|
257
|
-
return [
|
|
258
|
-
|
|
259
|
-
"lang": "en",
|
|
260
|
-
"
|
|
261
|
-
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
|
|
265
|
-
]
|
|
266
|
-
},
|
|
267
|
-
"messages": [
|
|
268
|
-
{
|
|
269
|
-
"role": "user",
|
|
270
|
-
"content": "Risky content"
|
|
271
|
-
}
|
|
272
|
-
],
|
|
273
|
-
"modality": "text",
|
|
274
|
-
"source": [
|
|
275
|
-
"source"
|
|
276
|
-
]
|
|
277
|
-
}]
|
|
278
|
-
|
|
259
|
+
return [
|
|
260
|
+
{
|
|
261
|
+
"metadata": {"lang": "en", "target_harms": [{"risk-type": "violence", "risk-subtype": ""}]},
|
|
262
|
+
"messages": [{"role": "user", "content": "Risky content"}],
|
|
263
|
+
"modality": "text",
|
|
264
|
+
"source": ["source"],
|
|
265
|
+
}
|
|
266
|
+
]
|