Selecione suas preferências de cookies

Usamos cookies essenciais e ferramentas semelhantes que são necessárias para fornecer nosso site e serviços. Usamos cookies de desempenho para coletar estatísticas anônimas, para que possamos entender como os clientes usam nosso site e fazer as devidas melhorias. Cookies essenciais não podem ser desativados, mas você pode clicar em “Personalizar” ou “Recusar” para recusar cookies de desempenho.

Se você concordar, a AWS e terceiros aprovados também usarão cookies para fornecer recursos úteis do site, lembrar suas preferências e exibir conteúdo relevante, incluindo publicidade relevante. Para aceitar ou recusar todos os cookies não essenciais, clique em “Aceitar” ou “Recusar”. Para fazer escolhas mais detalhadas, clique em “Personalizar”.

GuardrailContentFilter - Amazon Bedrock
Esta página não foi traduzida para seu idioma. Solicitar tradução

GuardrailContentFilter

Contains filter strengths for harmful content. Guardrails support the following content filters to detect and filter harmful user inputs and FM-generated outputs.

  • Hate – Describes language or a statement that discriminates, criticizes, insults, denounces, or dehumanizes a person or group on the basis of an identity (such as race, ethnicity, gender, religion, sexual orientation, ability, and national origin).

  • Insults – Describes language or a statement that includes demeaning, humiliating, mocking, insulting, or belittling language. This type of language is also labeled as bullying.

  • Sexual – Describes language or a statement that indicates sexual interest, activity, or arousal using direct or indirect references to body parts, physical traits, or sex.

  • Violence – Describes language or a statement that includes glorification of or threats to inflict physical pain, hurt, or injury toward a person, group or thing.

Content filtering depends on the confidence classification of user inputs and FM responses across each of the four harmful categories. All input and output statements are classified into one of four confidence levels (NONE, LOW, MEDIUM, HIGH) for each harmful category. For example, if a statement is classified as Hate with HIGH confidence, the likelihood of the statement representing hateful content is high. A single statement can be classified across multiple categories with varying confidence levels. For example, a single statement can be classified as Hate with HIGH confidence, Insults with LOW confidence, Sexual with NONE confidence, and Violence with MEDIUM confidence.

For more information, see Guardrails content filters.

This data type is used in the following API operations:

Contents

inputStrength

The strength of the content filter to apply to prompts. As you increase the filter strength, the likelihood of filtering harmful content increases and the probability of seeing harmful content in your application reduces.

Type: String

Valid Values: NONE | LOW | MEDIUM | HIGH

Required: Yes

outputStrength

The strength of the content filter to apply to model responses. As you increase the filter strength, the likelihood of filtering harmful content increases and the probability of seeing harmful content in your application reduces.

Type: String

Valid Values: NONE | LOW | MEDIUM | HIGH

Required: Yes

type

The harmful category that the content filter is applied to.

Type: String

Valid Values: SEXUAL | VIOLENCE | HATE | INSULTS | MISCONDUCT | PROMPT_ATTACK

Required: Yes

inputModalities

The input modalities selected for the guardrail content filter.

Type: Array of strings

Array Members: Minimum number of 1 item. Maximum number of 2 items.

Valid Values: TEXT | IMAGE

Required: No

outputModalities

The output modalities selected for the guardrail content filter.

Type: Array of strings

Array Members: Minimum number of 1 item. Maximum number of 2 items.

Valid Values: TEXT | IMAGE

Required: No

See Also

For more information about using this API in one of the language-specific AWS SDKs, see the following:

PrivacidadeTermos do sitePreferências de cookies
© 2025, Amazon Web Services, Inc. ou suas afiliadas. Todos os direitos reservados.