Merge pull request #31 from anevjes/openai-pii-stripping

Openai-pii-stripping

aisentry/facade/app.py

@@ -35,12 +35,6 @@
 
 load_dotenv(".env", override=True)
 
-
-# os.environ["AZURE_CLIENT_ID"] = "your_client_id"
-# os.environ["AZURE_TENANT_ID"] = "your_tenant_id"
-# os.environ["AZURE_FEDERATED_TOKEN_FILE"] = "/var/run/secrets/tokens/azure-identity-token"
-
-
 logger.info("Starting Ai-Sentry Facade app")
 app = Quart(__name__)
 
@@ -132,10 +126,10 @@ async def catch_all(path):
             client = endpoint_info["client"]
 
 
-            if openAI_request_headers.get('Api-Key') is not None:
-                logger.info("detected use of api-key header - will use this for authentication")
-                logger.debug(f"Swapping out api-key inside header with {endpoint_info['api-key']} value")
-                openAI_request_headers['Api-Key'] = endpoint_info['api-key']
+            # if openAI_request_headers.get('Api-Key') is not None:
+            #     logger.info("detected use of api-key header - will use this for authentication")
+            #     logger.debug(f"Swapping out api-key inside header with {endpoint_info['api-key']} value")
+            #     openAI_request_headers['Api-Key'] = endpoint_info['api-key']
 
             if endpoint_info['api-key'] is not None:
                 logger.info("No api-key header detected - will use the default api-key for authentication")

aisentry/utils/analyze_pii.py

@@ -58,8 +58,6 @@ async def analyze_pii_async(input_text: List[str]) -> None:
             async for page in pages:
                 document_results.append(page)
 
-        # doc=""
-
         for doc, action_results in zip(chunk, document_results):
             for result in action_results:
 
@@ -71,21 +69,18 @@ async def analyze_pii_async(input_text: List[str]) -> None:
                         logger.debug(f".........Confidence Score: {pii_entity.confidence_score}")
                         if pii_entity.confidence_score >= 0.8 and pii_entity.category != "DateTime":
                             logger.debug(f"Removing PII entity: {pii_entity.text}, category: {pii_entity.category} from the logged payload")
-                            # if pii_entity.text in "\},]":
-                            #     doc = doc.replace(pii_entity.text, "*PII*\"},")
-                            #     logger.info(f"PII-Processing: Replacing PII entity: {pii_entity.text} with extra escaping")
-                            doc = doc.replace(pii_entity.text, "*PII*")
+                            doc = doc.replace(pii_entity.text, "PII_REDACTED")
 
                 elif result.is_error is True:
                     logger.error(f'PII-Processing: An error with code {result.error.code} and message {result.error.message}')
 
 
     #UNTOCHED 
-    if ": *PII*," in doc:
-        doc = doc.replace(": *PII*,", ":\"*PII*\",")
+    if ": PII_REDACTED," in doc:
+        doc = doc.replace(": PII_REDACTED,", ":\"PII_REDACTED\",")
 
-    if "*PII*," in doc:
-        doc = doc.replace("*PII*,", "*PII*\"")
+    if "PII_REDACTED," in doc:
+        doc = doc.replace("PII_REDACTED,", "PII_REDACTED\"")
 
     logger.info(f"PII stripping completed")
     return doc

aisentry/utils/analyze_pii_openai.py

@@ -0,0 +1,91 @@
+import os
+import logging
+import asyncio
+from dotenv import load_dotenv
+from openai import AsyncAzureOpenAI
+
+
+
+# initial setup for logging / env variable loading
+log_level = os.getenv('LOG-LEVEL', 'INFO').upper()
+
+# Set up the logger
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=getattr(logging, log_level),
+                    format='%(asctime)s - %(levelname)s - %(message)s',
+                    datefmt='%d-%m-%Y %H:%M:%S'
+                    )
+
+# Initialize the OpenAI client with your key and endpoint
+
+load_dotenv(".env", override=True)
+
+openai_key = os.environ.get("PII_STRIPPING_OPENAI_API_KEY")
+openai_endpoint = os.environ.get("PII_STRIPPING_OPENAI_ENDPOINT")
+
+
+client = AsyncAzureOpenAI(
+    api_key=openai_key,  
+    api_version="2023-12-01-preview",
+    azure_endpoint=openai_endpoint
+)
+
+pii_stripping_system_prompt = """Objective: Identify and flag any Personally Identifiable Information (PII) within text data to ensure data privacy and compliance with regulations such as GDPR, CCPA, etc.
+
+PII includes but is not limited to:
+Full Names: First and last names
+Addresses: Street address, city, state, zip code
+Phone Numbers: Any format of telephone numbers
+Email Addresses: Any format of email addresses
+Social Security Numbers (SSNs): XXX-XX-XXXX or similar formats
+Credit Card Numbers: Any format of credit/debit card numbers
+Bank Account Numbers: Any format of bank account numbers
+Driver's License Numbers: Any format of driver's license numbers
+Passport Numbers: Any format of passport numbers
+Date of Birth: Full date of birth (MM/DD/YYYY or similar formats)
+IP Addresses: Any format of IPv4 or IPv6 addresses
+API-KEY or Token: Any format of API keys or tokens
+Medical Information: Any health-related information that can identify an individual
+Biometric Data: Fingerprints, facial recognition data, etc.
+
+Instructions for the System:
+Input: Accept text data for analysis.
+Processing:
+Use pattern matching, regular expressions, and machine learning algorithms to identify potential PII.
+Cross-reference detected patterns with known PII formats.
+Output:
+Flag detected PII and categorize it.
+Provide a confidence score for each detected PII item.
+Highlight the specific text containing PII.
+
+Example:
+
+Input Text:
+
+John Doe lives at 123 Maple Street, Springfield, IL 62704. His email is [email protected], and his phone number is (555) 123-4567. He was born on 01/15/1985 and his SSN is 123-45-6789.  
+ 
+Output:
+
+Keep the same text structure but replace the PII with placeholders: [PII-Redacted]
+ 
+Compliance Note: The system must handle all detected PII with strict confidentiality and in accordance with applicable data protection regulations."""
+
+
+async def get_chat_pii_stripped_completion(prompt):
+    # Send the request to Azure OpenAI
+    response = await client.chat.completions.create(
+        model="gpt-4",
+        messages=[
+    {"role": "system", "content": pii_stripping_system_prompt},
+    {"role": "user", "content": f"Rewrite the input and Strip out PII information as per the system message from following input: {prompt}"}
+  ]
+    )
+
+    # Extract the text from the response
+    #completion_text = response.completions[0].data.get("text", "")
+    message_content = response['choices'][0]['message']['content']
+
+    logger.info(f"PII Stripped Completion Text: {message_content}")
+    return message_content
+
+

aisentry/worker/cosmos_logger/cosmos_logger.py

@@ -1,10 +1,10 @@
 from flask import Flask, request, jsonify
 from cloudevents.http import from_http
 from requests.exceptions import HTTPError
-from azure.identity import DefaultAzureCredential
 from azure.core.exceptions import AzureError
 from dapr.clients import DaprClient
 from utils.analyze_pii import analyze_pii_async
+from utils.analyze_pii_openai import get_chat_pii_stripped_completion
 import logging
 import datetime
 import asyncio
@@ -16,6 +16,9 @@
 import uuid
 from typing import List
 
+
+load_dotenv(".env", override=True)
+
 # Get log level from environment variable
 log_level = os.getenv('LOG-LEVEL', 'INFO').upper()
 logger = logging.getLogger(__name__)
@@ -29,10 +32,11 @@
 
 
 
-load_dotenv(".env", override=True)
-
 app_port = os.getenv('APP_PORT', '7000')
 
+# This can be either OPENAI or TEXTANALYTICS
+pii_stripping_service = os.getenv('PII_STRIPPING_SERVICE', 'OPENAI')
+
 # Register Dapr pub/sub subscriptions
 @app.route('/dapr/subscribe', methods=['GET'])
 def subscribe():
@@ -82,9 +86,34 @@ async def oairequests_subscriber():
 
         input_data: List[str] = []
         input_data.append(output_binding_data)
-        output_binding_data = await analyze_pii_async(input_data)
-        logger.debug(f"PII stripped data: {output_binding_data}") 
-
+        if pii_stripping_service == 'TEXTANALYTICS':
+            logger.debug(f"PII stripping service: {pii_stripping_service}")
+            output_binding_data = await analyze_pii_async(input_data)
+
+        #OPENAI BASED PII Stripping
+        else:
+            logger.debug(f"PII stripping service: {pii_stripping_service}")
+
+            # Ensure a new event loop is created if the current one is closed
+            try:
+                loop = asyncio.get_event_loop()
+                if loop.is_closed():
+                    loop = asyncio.new_event_loop()
+                    asyncio.set_event_loop(loop)
+            except RuntimeError:
+                loop = asyncio.new_event_loop()
+                asyncio.set_event_loop(loop)
+
+            if loop.is_running():
+                task = loop.create_task(get_chat_pii_stripped_completion(input_data))
+                result = loop.run_until_complete(task)
+            else:
+                result = loop.run_until_complete(get_chat_pii_stripped_completion(input_data))
+
+            print(result)
+            # output_binding_data = await get_chat_pii_stripped_completion(input_data)
+
+        logger.debug(f"PII stripped data: {output_binding_data}")
 
 
     elif headers['ai-sentry-log-level'] == 'COMPLETE':

aisentry/worker/requirements.txt

@@ -46,6 +46,7 @@ msal==1.28.0
 msal-extensions==1.1.0
 multidict==6.0.5
 openai==1.14.2
+nest_asyncio==1.6.0
 packaging==24.0
 portalocker==2.8.2
 priority==2.0.0

build/build-ai-sentry-containers.ps1

@@ -1,9 +1,9 @@
 param(
     [string]$version = "0.1.1",
-    [string]$containerRegistry = "ariantestacr001.azurecr.io"
+    [string]$containerRegistry = "anevjesacrdevtest.azurecr.io"
 )
 #Uncomment first line for very first time to set to the right acr context
-# az acr login --name ariantestacr001
+# az acr login --name anevjesacrdevtest
 Write-Host "Building AI-Sentry Facade:$version"
 docker build --platform linux/amd64 -t ai-sentry-facadeapp:$version -f Dockerfile.facade ../aisentry/
 docker tag ai-sentry-facadeapp:$version $containerRegistry/ai-sentry-facadeapp:$version

content/documentation/AI-Sentry-config-settings.md

@@ -6,4 +6,6 @@ Each pool is represented as an object within the "pools" array. Each pool object
 Each endpoint object has the following properties:
 
 - url: The URL of the OpenAI instance. You're expected to replace <your-open-ai-instance> with the actual instance name. You need to append
-- api-key: The API key for accessing the OpenAI instance. You're expected to replace your-api-key with the actual API key.
+- api-key: The API key for accessing the OpenAI instance. You're expected to replace your-api-key with the actual API key. 
+
+Please note: We also support JWT auth to backend openAI instances. If you simply set "api-Key": null within the property bags inside the facade layer config; you will leverage aks workload identity to connect to openAi backends - however you will need worklaod identity federated out with managed identity stood up with your AKS cluster - and ofcourse grant the RBAC to the managed identity across all the required openAI instances in the backend.

content/documentation/AKSDeployment.md

@@ -48,7 +48,8 @@ helm repo add bitnami https://charts.bitnami.com/bitnami
 helm install sentry-redis bitnami/redis-cluster
 export REDIS_PASSWORD=$(kubectl get secret --namespace "default" sentry-redis-redis-cluster -o jsonpath="{.data.redis-password}" | base64 --decode)
 
- kubectl create secret generic redis --from-literal=redis-password=$REDIS_PASSWORD -n sentry-ai
+kubectl create namespace ai-sentry
+kubectl create secret generic redis --from-literal=redis-password=$REDIS_PASSWORD -n ai-sentry
 
 ```
 

content/documentation/Workload-identity-config.md

@@ -1,5 +1,10 @@
 # AKS Workload Identity setup
 
+## Enable Workload identity against existing AKS cluster
+```powershell
+az aks update --resource-group "aks-devtest-rg" --name "anevjes-aks-dev" --enable-oidc-issuer --enable-workload-identity
+```
+
 ## MI creation
 ```powershell
 az account set --subscription "subscriptionID"
@@ -31,7 +36,7 @@ export SERVICE_ACCOUNT_NAMESPACE="ai-sentry"
 ## OIDC Issuer url
 
 ```bash
-export AKS_OIDC_ISSUER="$(az aks show --name anevjes-aks --resource-group aks --query "oidcIssuerProfile.issuerUrl" -o tsv)"
+export AKS_OIDC_ISSUER="$(az aks show --name anevjes-aks-dev --resource-group aks-devtest-rg --query "oidcIssuerProfile.issuerUrl" -o tsv)"
 ```
 
 ## Create AKS Service Account

content/documentation/ai-sentry-config.json

@@ -6,11 +6,11 @@
             "endpoints": [
                 {
                     "url": "https://<your-open-ai-instance>.openai.azure.com/openai",
-                    "api-key": "yourapi-key"
+                    "api-key": "yourapi-key" //If you simply set "api-Key": null the facade layer will leverage aks workload identity to connect to openAi backends.
                 },
                 {
                     "url": "https://<your-open-ai-instance>.openai.azure.com/openai",
-                    "api-key": "your-api-key"
+                    "api-key": "your-api-key" //If you simply set "api-Key": null the facade layer will leverage aks workload identity to connect to openAi backends.
                 }
             ]
         },
@@ -20,11 +20,11 @@
             "endpoints": [
                 {
                     "url": "https://<your-open-ai-instance>.openai.azure.com/openai",
-                    "api-key": "your-api-key"
+                    "api-key": "your-api-key" //If you simply set "api-Key": null the facade layer will leverage aks workload identity to connect to openAi backends.
                 },
                 {
                     "url": "https://<your-open-ai-instance>.openai.azure.com/openai",
-                    "api-key": "your-api-key"
+                    "api-key": "your-api-key" //If you simply set "api-Key": null the facade layer will leverage aks workload identity to connect to openAi backends.
                 }
             ]
         }

deploy/aks/ai-sentry-deployment.yaml

@@ -64,6 +64,7 @@ spec:
     metadata:
       labels:
         app: facadeapp
+        azure.workload.identity/use: "true"
       annotations:
         dapr.io/enabled: "true"
         dapr.io/app-id: "facadeapp"
@@ -75,6 +76,8 @@ spec:
         dapr.io/app-health-probe-interval: "3"
         dapr.io/app-health-probe-timeout: "200"
         dapr.io/app-health-threshold: "2"
+        azure.workload.identity/inject-proxy-sidecar: "true"
+        azure.workload.identity/proxy-sidecar-port: "8000"
 
     spec:
       containers:
@@ -94,7 +97,7 @@ spec:
             failureThreshold: 3
           env:
           - name: "AI-SENTRY-ENDPOINT-CONFIG"
-            value: "{\"pools\":[{\"name\":\"pool1\",\"description\":\"pool1 description\",\"endpoints\":[{\"url\":\"https://youropenaiendpoint.openai.azure.com/openai\",\"api-key\":\"yourkey\"}]},{\"name\":\"pool2\",\"description\":\"pool2 description\",\"endpoints\":[{\"url\":\"https://youropenai.openai.azure.com/openai\",\"api-key\":\"yourkey\"},{\"url\":\"https://youropenai.openai.azure.com/openai\",\"api-key\":\"yourkey\"}]}]}"
+            value: "{\"pools\":[{\"name\":\"pool1\",\"description\":\"pool1 description\",\"endpoints\":[{\"url\":\"https://youropenaiendpoint.openai.azure.com/openai\",\"api-key\":\"yourkey or simply null\"}]},{\"name\":\"pool2\",\"description\":\"pool2 description\",\"endpoints\":[{\"url\":\"https://youropenai.openai.azure.com/openai\",\"api-key or simply null\":\"yourkey or simply null\"},{\"url\":\"https://youropenai.openai.azure.com/openai\",\"api-key or simply null\":\"yourkey\"}]}]}"
           - name: "LOG-LEVEL"
             value: "INFO"
 ---
@@ -141,6 +144,12 @@ spec:
               value: "your-key"
             - name: "LOG-LEVEL"
               value: "DEBUG"
+            - name: "PII_STRIPPING_SERVICE"
+              value: "OPENAI"
+            - name: PII_STRIPPING_OPENAI_ENDPOINT
+              value: "https://ptuopendeployment.openai.azure.com/"
+            - name: "PII_STRIPPING_OPENAI_API_KEY"
+              value: "yourapikeytoopenai / apim subscription key"
 ---
 #Summary Logger Logger
 kind: StatefulSet