Analyze Content for Safety and Harm

Azure Content Safety Python provides SDK patterns for detecting harmful text and images in applications. It covers user-generated and generated content, with analysis results organized by harm category and severity. The workflow supports both general category classification and domain-specific text blocklists.

Text analysis uses ContentSafetyClient with AnalyzeTextOptions and inspects category results for hate, self-harm, sexual content, and violence. Default text severities are 0, 2, 4, and 6, with an optional eight-level mode from 0 through 7 for finer control. The categories distinguish identity-based attacks, sexual material, physical harm, and self-injury content.

Image analysis accepts a local file converted to base64 or a blob URL through ImageData. AnalyzeImageOptions carries that image into the request, and the response exposes category severity results. Severity results use safe, low, medium, and high bands for text and images.

BlocklistClient manages custom text blocklists separately from the analysis client. Create or update a named list, add or update blocked terms, and analyze text against the selected list. Enabling halt_on_blocklist_hit supports immediate rejection on a match, and the response includes matched blocklist item text.

Use this skill when an application needs text or image screening and category-specific severity results before content is shown to users. Custom blocklists are relevant when the domain has particular terms that general category analysis alone does not capture.

Choose thresholds for the actual use case and consider every returned category, because content can be harmful in several ways. Log analysis results for auditing and improvement, and validate the moderation workflow in its environment. Clarify missing inputs, permissions, safety boundaries, or success criteria before execution.

Setup requires an Azure Content Safety endpoint and authentication credentials. Requests supply text, encoded image data, or an image URL; text requests can also name blocklists and select a severity output mode. Outputs include category analyses, severity values, and blocklist matches.

Install the Python SDK:

pip install azure-ai-contentsafety

Use CONTENT_SAFETY_ENDPOINT and CONTENT_SAFETY_KEY for endpoint and key configuration. API-key authentication wraps the key in AzureKeyCredential; Entra ID authentication uses DefaultAzureCredential. Text, image, and blocklist examples share the configured service endpoint.

Text analysis can select AnalyzeTextOutputType.EIGHT_SEVERITY_LEVELS through output_type. In responses, categories_analysis carries each category and severity; blocklists_match exposes matched blocklist_item_text. Hate includes attacks based on race, religion, or gender, while self-harm includes suicide and eating-disorder content.

Sexual content includes relationships and anatomy; violence includes weapons and injury.