azure-ai-contentsafety-py – skill Viết lách & Nội dung cho agent AI

Azure AI Content Safety SDK for Python

Detect harmful user-generated and AI-generated content in applications.

Installation

pip install azure-ai-contentsafety

Environment Variables

CONTENT_SAFETY_ENDPOINT=https://<resource>.cognitiveservices.azure.com
CONTENT_SAFETY_KEY=<your-api-key>

Authentication

API Key

from azure.ai.contentsafety import ContentSafetyClient
from azure.core.credentials import AzureKeyCredential
import os

client = ContentSafetyClient(
    endpoint=os.environ["CONTENT_SAFETY_ENDPOINT"],
    credential=AzureKeyCredential(os.environ["CONTENT_SAFETY_KEY"])
)

Entra ID

from azure.ai.contentsafety import ContentSafetyClient
from azure.identity import DefaultAzureCredential

client = ContentSafetyClient(
    endpoint=os.environ["CONTENT_SAFETY_ENDPOINT"],
    credential=DefaultAzureCredential()
)

Use only an approved endpoint/account, content and cost/retention policy; uploads disclose text/images to a hosted classifier. Bound input bytes and use reviewed SDK models/imports (azure-identity is additionally needed for Entra ID). Classification can miss harm or flag benign context: severity zero is not a universal safety guarantee. Validate thresholds on representative languages/domains, define outage/ambiguous-result handling and human appeal/review. Blocklist mutations can affect real moderation and require explicit scope; redact content in logs and close clients/credentials.

Analyze Text

from azure.ai.contentsafety import ContentSafetyClient
from azure.ai.contentsafety.models import AnalyzeTextOptions, TextCategory
from azure.core.credentials import AzureKeyCredential

client = ContentSafetyClient(endpoint, AzureKeyCredential(key))

request = AnalyzeTextOptions(text="Your text content to analyze")
response = client.analyze_text(request)

# Check each category
for category in [TextCategory.HATE, TextCategory.SELF_HARM, 
                 TextCategory.SEXUAL, TextCategory.VIOLENCE]:
    result = next((r for r in response.categories_analysis 
                   if r.category == category), None)
    if result:
        print(f"{category}: severity {result.severity}")

Analyze Image

from azure.ai.contentsafety import ContentSafetyClient
from azure.ai.contentsafety.models import AnalyzeImageOptions, ImageData
from azure.core.credentials import AzureKeyCredential
import base64

client = ContentSafetyClient(endpoint, AzureKeyCredential(key))

# From file
with open("image.jpg", "rb") as f:
    image_data = f.read()  # SDK serializes bytes to the REST base64 representation

request = AnalyzeImageOptions(
    image=ImageData(content=image_data)
)

response = client.analyze_image(request)

for result in response.categories_analysis:
    print(f"{result.category}: severity {result.severity}")

Image from URL

from azure.ai.contentsafety.models import AnalyzeImageOptions, ImageData

request = AnalyzeImageOptions(
    image=ImageData(blob_url="https://example.com/image.jpg")
)

response = client.analyze_image(request)

Text Blocklist Management

Create Blocklist

from azure.ai.contentsafety import BlocklistClient
from azure.ai.contentsafety.models import TextBlocklist
from azure.core.credentials import AzureKeyCredential

blocklist_client = BlocklistClient(endpoint, AzureKeyCredential(key))

blocklist = TextBlocklist(
    blocklist_name="my-blocklist",
    description="Custom terms to block"
)

result = blocklist_client.create_or_update_text_blocklist(
    blocklist_name="my-blocklist",
    options=blocklist
)

Add Block Items

from azure.ai.contentsafety.models import AddOrUpdateTextBlocklistItemsOptions, TextBlocklistItem

items = AddOrUpdateTextBlocklistItemsOptions(
    blocklist_items=[
        TextBlocklistItem(text="blocked-term-1"),
        TextBlocklistItem(text="blocked-term-2")
    ]
)

result = blocklist_client.add_or_update_blocklist_items(
    blocklist_name="my-blocklist",
    options=items
)

Analyze with Blocklist

from azure.ai.contentsafety.models import AnalyzeTextOptions

request = AnalyzeTextOptions(
    text="Text containing blocked-term-1",
    blocklist_names=["my-blocklist"],
    halt_on_blocklist_hit=True
)

response = client.analyze_text(request)

if response.blocklists_match:
    for match in response.blocklists_match:
        print(f"Blocked: {match.blocklist_item_text}")

Severity Levels

Text analysis returns 4 severity levels (0, 2, 4, 6) by default. For 8 levels (0-7):

from azure.ai.contentsafety.models import AnalyzeTextOptions, AnalyzeTextOutputType

request = AnalyzeTextOptions(
    text="Your text",
    output_type=AnalyzeTextOutputType.EIGHT_SEVERITY_LEVELS
)

Harm Categories

| Category | Description |

|----------|-------------|

| Hate | Attacks based on identity (race, religion, gender, etc.) |

| Sexual | Sexual content, relationships, anatomy |

| Violence | Physical harm, weapons, injury |

| SelfHarm | Self-injury, suicide, eating disorders |

Severity Scale

| Level | Text Range | Image Range | Meaning |

|-------|------------|-------------|---------|

| 0 | Not detected | Not detected | Model did not detect harm in supported categories |

| 2 | Low | Low | Mild references |

| 4 | Medium | Medium | Moderate content |

| 6 | High | High | Severe content |

Client Types

| Client | Purpose |

|--------|---------|

| ContentSafetyClient | Analyze text and images |

| BlocklistClient | Manage custom blocklists |

Best Practices

1. Use blocklists for domain-specific terms

2. Set severity thresholds appropriate for your use case

3. Handle multiple categories — content can be harmful in multiple ways

4. Use halt_on_blocklist_hit for immediate rejection

5. Log analysis results for audit and improvement

6. Consider 8-severity mode for finer-grained control

7. Pre-moderate AI outputs before showing to users

When to Use

This skill is applicable to execute the workflow or actions described in the overview.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.