Customizing Titan Text Filters

Amazon Bedrock provides a feature called Guardrails, which is a collection of filtering policies designed to help prevent undesirable and harmful content in generative AI applications. These guardrails can be applied to interactions with foundation models in Bedrock, including Titan models.

The guardrails in Amazon Bedrock offer several configurable policies:

1. Content filters: These can block harmful content such as hate speech, insults, sexual content, violence, and prompt attacks.

2. Denied topics: You can define specific topics to avoid within your application.

3. Word filters: Custom words or phrases can be configured to detect and block, including profanity or competitor names.

4. Sensitive information filters: These can detect and remove personally identifiable information (PII) or custom regex patterns.

5. Contextual grounding check: This helps detect and filter hallucinations or irrelevant responses.

To configure these guardrails, you would use the Amazon Bedrock console or API. In the console, you can create a new guardrail by providing a name and description, configuring messaging for blocked prompts and responses, and setting up the various filters mentioned above.

It's important to note that these guardrails are a feature of Amazon Bedrock and are not specific to Titan models. They can be applied to interactions with various foundation models available through Bedrock.

The customization options you mentioned for Titan models specifically relate to model fine-tuning parameters, such as epochs, batch size, learning rate, and learning rate warmup steps. These are different from content filtering and are used to adjust the model's performance for specific tasks or domains.

If you want to implement custom filtering beyond what Bedrock Guardrails offer, you would need to handle that in your own application logic, either by preprocessing prompts or post-processing responses before presenting them to users.

Sources
Components of a guardrail - Amazon Bedrock
Create a guardrail - Amazon Bedrock
Amazon Titan text model customization hyperparameters - Amazon Bedrock