Amazon Bedrock’s Guardrails offer customers the ability to implement safeguards tailored to application requirements and enterprise AI policies. These guardrails aim to prevent unwanted content, thwart rapid attacks, and eliminate privacy breaches. By combining various policy types, users can configure these safeguards for different scenarios and extend them to both base models within Amazon Bedrock and custom or third-party models outside the platform. Guardrails can also be integrated with Agents and Knowledge Bases within Amazon Bedrock to provide added security features.
Initially introduced in preview during re:Invent 2023, Guardrails for Amazon Bedrock is now generally available as of April 2024. This feature includes policies for content filtering, denied topics, sensitive information filters, and word filters. One prominent user of Guardrails is the largest insurer in Spain, MAPFRE, which utilizes Guardrails to align their RAG-based chatbot, Mark.IA, with corporate security policies and responsible AI practices.
In addition to the existing features, two new capabilities are introduced: contextual grounding checks to identify hallucinations in model responses, and the ApplyGuardrail API for evaluating input requests and model responses across all base and custom models supported by Amazon Bedrock. Contextual basis checks detect responses not informed by business data, enhancing their credibility. By setting thresholds for grounding and relevance, users can filter out inaccurate or irrelevant responses, ensuring higher response quality.
Users can easily configure contextual grounding with the CreateGuardrail API, as shown in the example provided. ApplyGuardrail now extends beyond Amazon Bedrock to assess input requests and model responses for generative AI applications built on various platforms, ensuring standardized protection measures. With the flexibility offered by the ApplyGuardrail API, users can evaluate and filter incoming user inputs and model responses at different stages of application development.
Through a practical example using the AWS SDK for Python, the process of creating and applying a guardrail for evaluating financial advice is demonstrated. By invoking the apply_guardrail function, users can block messages that violate predefined policies, ensuring compliance with regulations and security standards. The ApplyGuardrail API enhances centralized governance across generative AI applications, regardless of the infrastructure used or the models implemented.
The availability of contextual grounding and ApplyGuardrail APIs in all AWS regions empowers users to enhance the security and reliability of their generative AI applications efficiently. For further information on Guardrails for Amazon Bedrock, visit the product page on Amazon’s website and explore the pricing details for these policies. Engage with the AWS community to gain insights into real-world application scenarios and best practices for leveraging Amazon Bedrock in diverse solutions.
Article Source
https://aws.amazon.com/blogs/aws/guardrails-for-amazon-bedrock-can-now-detect-hallucinations-and-safeguard-apps-built-using-custom-or-third-party-fms/
