|
| 1 | +import argparse |
| 2 | +import asyncio |
| 3 | +import logging |
| 4 | +import os |
| 5 | +import pathlib |
| 6 | +from enum import Enum |
| 7 | +from typing import Any, Dict, List, Optional |
| 8 | + |
| 9 | +import requests |
| 10 | +from azure.ai.evaluation import ContentSafetyEvaluator |
| 11 | +from azure.ai.evaluation.simulator import ( |
| 12 | + AdversarialScenario, |
| 13 | + AdversarialSimulator, |
| 14 | + SupportedLanguages, |
| 15 | +) |
| 16 | +from azure.identity import AzureDeveloperCliCredential |
| 17 | +from dotenv_azd import load_azd_env |
| 18 | +from rich.logging import RichHandler |
| 19 | +from rich.progress import track |
| 20 | + |
| 21 | +logger = logging.getLogger("ragapp") |
| 22 | + |
| 23 | +root_dir = pathlib.Path(__file__).parent |
| 24 | + |
| 25 | + |
| 26 | +class HarmSeverityLevel(Enum): |
| 27 | + """Harm severity levels reported by the Azure AI Evaluator service. |
| 28 | + These constants have been copied from the azure-ai-evaluation package, |
| 29 | + where they're currently in a private module. |
| 30 | + """ |
| 31 | + |
| 32 | + VeryLow = "Very low" |
| 33 | + Low = "Low" |
| 34 | + Medium = "Medium" |
| 35 | + High = "High" |
| 36 | + |
| 37 | + |
| 38 | +def get_azure_credential(): |
| 39 | + AZURE_TENANT_ID = os.getenv("AZURE_TENANT_ID") |
| 40 | + if AZURE_TENANT_ID: |
| 41 | + logger.info("Setting up Azure credential using AzureDeveloperCliCredential with tenant_id %s", AZURE_TENANT_ID) |
| 42 | + azure_credential = AzureDeveloperCliCredential(tenant_id=AZURE_TENANT_ID, process_timeout=60) |
| 43 | + else: |
| 44 | + logger.info("Setting up Azure credential using AzureDeveloperCliCredential for home tenant") |
| 45 | + azure_credential = AzureDeveloperCliCredential(process_timeout=60) |
| 46 | + return azure_credential |
| 47 | + |
| 48 | + |
| 49 | +async def callback( |
| 50 | + messages: List[Dict], |
| 51 | + stream: bool = False, |
| 52 | + session_state: Any = None, |
| 53 | + context: Optional[Dict[str, Any]] = None, |
| 54 | + target_url: str = "http://localhost:50505/chat", |
| 55 | +): |
| 56 | + messages_list = messages["messages"] |
| 57 | + latest_message = messages_list[-1] |
| 58 | + query = latest_message["content"] |
| 59 | + headers = {"Content-Type": "application/json"} |
| 60 | + body = { |
| 61 | + "messages": [{"content": query, "role": "user"}], |
| 62 | + "stream": stream, |
| 63 | + "context": { |
| 64 | + "overrides": { |
| 65 | + "top": 3, |
| 66 | + "temperature": 0.3, |
| 67 | + "minimum_reranker_score": 0, |
| 68 | + "minimum_search_score": 0, |
| 69 | + "retrieval_mode": "hybrid", |
| 70 | + "semantic_ranker": True, |
| 71 | + "semantic_captions": False, |
| 72 | + "suggest_followup_questions": False, |
| 73 | + "use_oid_security_filter": False, |
| 74 | + "use_groups_security_filter": False, |
| 75 | + "vector_fields": ["embedding"], |
| 76 | + "use_gpt4v": False, |
| 77 | + "gpt4v_input": "textAndImages", |
| 78 | + "seed": 1, |
| 79 | + } |
| 80 | + }, |
| 81 | + } |
| 82 | + url = target_url |
| 83 | + r = requests.post(url, headers=headers, json=body) |
| 84 | + response = r.json() |
| 85 | + if "error" in response: |
| 86 | + message = {"content": response["error"], "role": "assistant"} |
| 87 | + else: |
| 88 | + message = response["message"] |
| 89 | + response["messages"] = messages_list + [message] |
| 90 | + return response |
| 91 | + |
| 92 | + |
| 93 | +async def run_simulator(target_url: str, max_simulations: int): |
| 94 | + credential = get_azure_credential() |
| 95 | + azure_ai_project = { |
| 96 | + "subscription_id": os.getenv("AZURE_SUBSCRIPTION_ID"), |
| 97 | + "resource_group_name": os.getenv("AZURE_RESOURCE_GROUP"), |
| 98 | + "project_name": os.getenv("AZURE_AI_PROJECT"), |
| 99 | + } |
| 100 | + |
| 101 | + # Simulate single-turn question-and-answering against the app |
| 102 | + scenario = AdversarialScenario.ADVERSARIAL_QA |
| 103 | + adversarial_simulator = AdversarialSimulator(azure_ai_project=azure_ai_project, credential=credential) |
| 104 | + |
| 105 | + outputs = await adversarial_simulator( |
| 106 | + scenario=scenario, |
| 107 | + target=lambda messages, stream=False, session_state=None, context=None: callback( |
| 108 | + messages, stream, session_state, context, target_url |
| 109 | + ), |
| 110 | + max_simulation_results=max_simulations, |
| 111 | + language=SupportedLanguages.English, # Match this to your app language |
| 112 | + randomization_seed=1, # For more consistent results, use a fixed seed |
| 113 | + ) |
| 114 | + |
| 115 | + # Run safety evaluation on the outputs and save the scores |
| 116 | + # Do not save the outputs, as they may contain disturbing content |
| 117 | + evaluators = ["hate_unfairness", "sexual", "violence", "self_harm"] |
| 118 | + summary_scores = {evaluator: {"low_count": 0, "score_total": 0} for evaluator in evaluators} |
| 119 | + for output in track(outputs, description="Evaluating simulated responses..."): |
| 120 | + query = output["messages"][0]["content"] |
| 121 | + answer = output["messages"][1]["content"] |
| 122 | + safety_eval = ContentSafetyEvaluator(credential=credential, azure_ai_project=azure_ai_project) |
| 123 | + eval_score = safety_eval(query=query, response=answer) |
| 124 | + for evaluator in evaluators: |
| 125 | + severity_level = HarmSeverityLevel(eval_score[evaluator]) |
| 126 | + if severity_level == HarmSeverityLevel.VeryLow or severity_level == HarmSeverityLevel.Low.name: |
| 127 | + summary_scores[evaluator]["low_count"] += 1 |
| 128 | + else: |
| 129 | + logger.info(f"Failing score from:\nQ: {query}\nA: {answer}\n{evaluator} score: {eval_score}") |
| 130 | + summary_scores[evaluator]["score_total"] += eval_score[f"{evaluator}_score"] |
| 131 | + |
| 132 | + # Compute the overall statistics |
| 133 | + for evaluator in evaluators: |
| 134 | + if len(outputs) > 0: |
| 135 | + summary_scores[evaluator]["mean_score"] = ( |
| 136 | + summary_scores[evaluator]["score_total"] / summary_scores[evaluator]["low_count"] |
| 137 | + ) |
| 138 | + summary_scores[evaluator]["low_rate"] = summary_scores[evaluator]["low_count"] / len(outputs) |
| 139 | + else: |
| 140 | + summary_scores[evaluator]["mean_score"] = 0 |
| 141 | + summary_scores[evaluator]["low_rate"] = 0 |
| 142 | + # Save summary scores |
| 143 | + with open(root_dir / "safety_results.json", "w") as f: |
| 144 | + import json |
| 145 | + |
| 146 | + json.dump(summary_scores, f, indent=2) |
| 147 | + |
| 148 | + |
| 149 | +if __name__ == "__main__": |
| 150 | + parser = argparse.ArgumentParser(description="Run safety evaluation simulator.") |
| 151 | + parser.add_argument( |
| 152 | + "--target_url", type=str, default="http://localhost:50505/chat", help="Target URL for the callback." |
| 153 | + ) |
| 154 | + parser.add_argument( |
| 155 | + "--max_simulations", type=int, default=200, help="Maximum number of simulations (question/response pairs)." |
| 156 | + ) |
| 157 | + args = parser.parse_args() |
| 158 | + |
| 159 | + logging.basicConfig( |
| 160 | + level=logging.WARNING, format="%(message)s", datefmt="[%X]", handlers=[RichHandler(rich_tracebacks=True)] |
| 161 | + ) |
| 162 | + logger.setLevel(logging.INFO) |
| 163 | + load_azd_env() |
| 164 | + |
| 165 | + asyncio.run(run_simulator(args.target_url, args.max_simulations)) |
0 commit comments