Agentic Data Plane

rpk ai guardrail create

Create a Guardrail.

Choose the config by setting flags from one group; the group you use selects it: bedrock-config, local-config. Setting flags from two groups is an error.

This command was introduced in ai version 0.2.94.

Usage

rpk ai guardrail create NAME [flags]

Examples

This section provides examples of how to use rpk ai guardrail create.

rpk ai guardrail create my-guardrail \
  --blocked-input-message Your request was blocked by a content policy. \
  --blocked-output-message The response was blocked by a content policy. \
  --bedrock-config.backend.region us-east-1

Flags

Value Type Description

--display-name

string

Name shown in the UI, up to 253 characters. Defaults to the identifier.

--description

string

Notes shown with the guardrail in the UI, up to 1,024 characters.

--blocked-input-message

string

Returned when an input policy blocks a prompt. 1 to 500 characters. Shared by every input-blocking policy. (required).

--blocked-output-message

string

Returned when an output policy blocks a response. 1 to 500 characters. Shared by every output-blocking policy. (required).

--enabled

bool

Turns the whole guardrail on or off. While off, no policy runs and configuration is kept. Changes take effect within about 30 seconds. (true when set; pass --enabled=false to disable).

--bedrock-config.backend.region

string

AWS region where the Bedrock guardrail is created, for example us-east-1. Content filters, contextual grounding, and automated reasoning aren’t available in every region. (required in its group).

--bedrock-config.backend.llm-provider-ref

string

Supplies AWS credentials and region. Selecting a provider doesn’t attach the guardrail to it.

--bedrock-config.backend.standalone.default-chain

bool

Default chain. (selects this credentials mode; takes no value).

--bedrock-config.backend.standalone.static-credentials.access-key-id-ref

string

Secret store key that holds the AWS access key ID, in UPPER_SNAKE_CASE (for example, AWS_ACCESS_KEY_ID). The credentials need permission to manage Bedrock guardrails in the configured region. (required in its group).

--bedrock-config.backend.standalone.static-credentials.secret-access-key-ref

string

Secret store key that holds the AWS secret access key, in UPPER_SNAKE_CASE (for example, AWS_SECRET_ACCESS_KEY). (required in its group).

--bedrock-config.backend.standalone.assume-role.role-arn

string

ARN of the IAM role to assume through STS (arn:aws:iam::<account-id>:role/<role-name>). The trust policy must allow Redpanda to assume the role, and the role must be able to manage Bedrock guardrails in the configured region. Creating the guardrail calls AWS, so before you create it, the trust policy must require the External ID <environment ID>/guardrail/<guardrail name>. A role in another AWS account, or one whose trust policy names a whole account, must be under the /redpanda-adp/ path. (required in its group).

--bedrock-config.backend.standalone.assume-role.session-name

string

Session name recorded in CloudTrail for calls this guardrail makes. Defaults to a Redpanda-supplied name.

--bedrock-config.content-filter-policy.enabled

bool

Turns all content categories on or off. Configuration is kept while off. Categories run only while the guardrail is enabled. (true when set; pass --bedrock-config.content-filter-policy.enabled=false to disable).

--bedrock-config.content-filter-policy.hate.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.hate.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.hate.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.hate.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.hate.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.hate.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.insults.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.insults.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.insults.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.insults.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.insults.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.insults.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.sexual.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.sexual.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.sexual.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.sexual.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.sexual.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.sexual.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.violence.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.violence.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.violence.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.violence.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.violence.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.violence.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.misconduct.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.misconduct.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.misconduct.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.misconduct.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.misconduct.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.misconduct.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.prompt-attack.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.prompt-attack.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.prompt-attack.input.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.content-filter-policy.prompt-attack.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--bedrock-config.content-filter-policy.prompt-attack.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--bedrock-config.content-filter-policy.prompt-attack.output.modalities

strings

Content types to inspect. Text only when empty. Image requires a region and model with multimodal guardrail support. (one of: text, image; repeatable; replaces the full list on update).

--bedrock-config.word-filter-policy.enabled

bool

Turns word filtering on or off. Configuration is kept while off. Matching is exact and case-insensitive, so misspellings and paraphrases don’t match. (true when set; pass --bedrock-config.word-filter-policy.enabled=false to disable).

--bedrock-config.word-filter-policy.custom-words

stringArray

Custom word groups. Every word in a group shares the group’s input and output action. Up to 10,000 words across all groups. (repeatable; protojson object; replaces the full list on update).

--bedrock-config.word-filter-policy.managed-lists

stringArray

Platform-curated word lists, such as the AWS-managed profanity list, with an action per direction. (repeatable; protojson object; replaces the full list on update).

--bedrock-config.denied-topics-policy.enabled

bool

Turns denied topics on or off. Configuration is kept while off. Matching is semantic, so paraphrases and misspellings match. (true when set; pass --bedrock-config.denied-topics-policy.enabled=false to disable).

--bedrock-config.denied-topics-policy.topics

stringArray

Topics denied by meaning, each with a name, definition, and optional examples. 1 to 30 topics. (repeatable; name (bare values comma-split) or protojson object; replaces the full list on update).

--bedrock-config.pii-filter-policy.enabled

bool

Turns sensitive-information detection on or off. Configuration is kept while off. (true when set; pass --bedrock-config.pii-filter-policy.enabled=false to disable).

--bedrock-config.pii-filter-policy.entities

stringArray

Rules for the 31 built-in entity types, such as email address, phone number, credit card number, or US Social Security number. One rule per type. (repeatable; protojson object; replaces the full list on update).

--bedrock-config.pii-filter-policy.regexes

stringArray

RE2 patterns for identifiers the built-in types don’t cover, such as internal account IDs. Bedrock doesn’t support lookahead or lookbehind. (repeatable; name (bare values comma-split) or protojson object; replaces the full list on update).

--bedrock-config.grounding-policy.enabled

bool

Turns contextual grounding on or off. Configuration is kept while off. Checks run on output only and require a grounding source and query in the request. (true when set; pass --bedrock-config.grounding-policy.enabled=false to disable).

--bedrock-config.grounding-policy.grounding.enabled

bool

Turns this check on or off independently. Configuration is kept while off. (true when set; pass --bedrock-config.grounding-policy.grounding.enabled=false to disable).

--bedrock-config.grounding-policy.grounding.threshold

float

Minimum score from 0.0 to 0.99. The action fires below it. Higher values flag more responses.

--bedrock-config.grounding-policy.grounding.action

string

Action on a low score. None records the result and passes the response through. Block returns the blocked output message and stops the request. (one of: none, block).

--bedrock-config.grounding-policy.relevance.enabled

bool

Turns this check on or off independently. Configuration is kept while off. (true when set; pass --bedrock-config.grounding-policy.relevance.enabled=false to disable).

--bedrock-config.grounding-policy.relevance.threshold

float

Minimum score from 0.0 to 0.99. The action fires below it. Higher values flag more responses.

--bedrock-config.grounding-policy.relevance.action

string

Action on a low score. None records the result and passes the response through. Block returns the blocked output message and stops the request. (one of: none, block).

--bedrock-config.automated-reasoning-policy.enabled

bool

Turns automated reasoning on or off. Configuration is kept while off. Findings are added to the response trace. Requests are never blocked. (true when set; pass --bedrock-config.automated-reasoning-policy.enabled=false to disable).

--bedrock-config.automated-reasoning-policy.policy-arns

strings

ARNs of one or two published Bedrock Automated Reasoning policies. Each ARN ends in a numeric version such as :1. DRAFT versions are rejected. (repeatable; replaces the full list on update).

--bedrock-config.automated-reasoning-policy.confidence-threshold

float

Confidence required for a definitive finding, from 0.0 to 1.0. Lower-confidence results are reported as ambiguous.

--local-config.bedrock-backend.region

string

AWS region of the companion Bedrock guardrail for delegated families, for example us-east-1. Feature availability varies by region. (required in its group).

--local-config.bedrock-backend.llm-provider-ref

string

Identifier of a Bedrock LLM provider whose AWS credentials the delegated families use.

--local-config.bedrock-backend.standalone.default-chain

bool

Default chain. (selects this credentials mode; takes no value).

--local-config.bedrock-backend.standalone.static-credentials.access-key-id-ref

string

Secret store key that holds the AWS access key ID, in UPPER_SNAKE_CASE (for example, AWS_ACCESS_KEY_ID). The credentials need permission to manage Bedrock guardrails in the configured region. (required in its group).

--local-config.bedrock-backend.standalone.static-credentials.secret-access-key-ref

string

Secret store key that holds the AWS secret access key, in UPPER_SNAKE_CASE (for example, AWS_SECRET_ACCESS_KEY). (required in its group).

--local-config.bedrock-backend.standalone.assume-role.role-arn

string

ARN of the IAM role to assume through STS (arn:aws:iam::<account-id>:role/<role-name>). The trust policy must allow Redpanda to assume the role, and the role must be able to manage Bedrock guardrails in the configured region. Creating the guardrail calls AWS, so before you create it, the trust policy must require the External ID <environment ID>/guardrail/<guardrail name>. A role in another AWS account, or one whose trust policy names a whole account, must be under the /redpanda-adp/ path. (required in its group).

--local-config.bedrock-backend.standalone.assume-role.session-name

string

Session name recorded in CloudTrail for calls this guardrail makes. Defaults to a Redpanda-supplied name.

--local-config.word-filter-policy.engine

string

Engine for word filters. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.word-filter-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.word-filter-policy.shadow-local=false to disable).

--local-config.word-filter-policy.enabled

bool

Turns word filtering on or off. Configuration is kept while off. Matching is exact and case-insensitive, so misspellings and paraphrases don’t match. (true when set; pass --local-config.word-filter-policy.enabled=false to disable).

--local-config.word-filter-policy.custom-words

stringArray

Custom word groups. Every word in a group shares the group’s input and output action. Up to 10,000 words across all groups. (repeatable; protojson object; replaces the full list on update).

--local-config.word-filter-policy.managed-lists

stringArray

Platform-curated word lists, such as the profanity list, with an action per direction. (repeatable; protojson object; replaces the full list on update).

--local-config.pattern-filter-policy.engine

string

Engine for custom patterns. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.pattern-filter-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.pattern-filter-policy.shadow-local=false to disable).

--local-config.pattern-filter-policy.enabled

bool

Turns custom-pattern detection on or off. Configuration is kept while off. (true when set; pass --local-config.pattern-filter-policy.enabled=false to disable).

--local-config.pattern-filter-policy.patterns

stringArray

RE2 patterns for identifiers the built-in types don’t cover, such as internal account IDs. Lookahead and lookbehind aren’t supported. (repeatable; name (bare values comma-split) or protojson object; replaces the full list on update).

--local-config.pii-filter-policy.engine

string

Engine for built-in PII entities. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.pii-filter-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.pii-filter-policy.shadow-local=false to disable).

--local-config.pii-filter-policy.enabled

bool

Turns sensitive-information detection on or off. Configuration is kept while off. (true when set; pass --local-config.pii-filter-policy.enabled=false to disable).

--local-config.pii-filter-policy.entities

stringArray

Rules for the 31 built-in entity types, such as email address, phone number, credit card number, or US Social Security number. (repeatable; protojson object; replaces the full list on update).

--local-config.prompt-attack-policy.engine

string

Engine for prompt attacks. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.prompt-attack-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.prompt-attack-policy.shadow-local=false to disable).

--local-config.prompt-attack-policy.enabled

bool

Turns prompt-attack detection on or off. (true when set; pass --local-config.prompt-attack-policy.enabled=false to disable).

--local-config.prompt-attack-policy.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.prompt-attack-policy.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input message and stops the request. (one of: none, block).

--local-config.prompt-attack-policy.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.engine

string

Engine for harm categories. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.content-filter-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.content-filter-policy.shadow-local=false to disable).

--local-config.content-filter-policy.enabled

bool

Turns all content categories on or off. Configuration is kept while off. Categories run only while the guardrail is enabled. (true when set; pass --local-config.content-filter-policy.enabled=false to disable).

--local-config.content-filter-policy.hate.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.hate.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.hate.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.hate.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.hate.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.hate.output.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.insults.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.insults.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.insults.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.insults.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.insults.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.insults.output.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.sexual.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.sexual.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.sexual.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.sexual.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.sexual.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.sexual.output.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.violence.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.violence.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.violence.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.violence.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.violence.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.violence.output.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.misconduct.input.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.misconduct.input.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.misconduct.input.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.content-filter-policy.misconduct.output.strength

string

Detection sensitivity. Low flags only high-confidence matches. High flags any non-negligible match. None records a score without acting. (one of: none, low, medium, high).

--local-config.content-filter-policy.misconduct.output.action

string

Action on a match. None records the match and passes content through. Block returns the blocked input or output message and stops the request. (one of: none, block).

--local-config.content-filter-policy.misconduct.output.modalities

strings

Content types to inspect. Text only when empty. Image requires engine support. (one of: text, image; repeatable; replaces the full list on update).

--local-config.denied-topics-policy.engine

string

Engine for denied topics. Local runs in the gateway and keeps text in the cluster. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.denied-topics-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.denied-topics-policy.shadow-local=false to disable).

--local-config.denied-topics-policy.enabled

bool

Turns denied topics on or off. Configuration is kept while off. Matching is semantic, so paraphrases and misspellings match. (true when set; pass --local-config.denied-topics-policy.enabled=false to disable).

--local-config.denied-topics-policy.topics

stringArray

Topics denied by meaning, each with a name, definition, and optional examples. 1 to 30 topics. (repeatable; name (bare values comma-split) or protojson object; replaces the full list on update).

--local-config.grounding-policy.engine

string

Engine for grounding. Local runs in the gateway. Bedrock runs in the companion guardrail. Unset uses the platform default. (one of: local, bedrock).

--local-config.grounding-policy.shadow-local

bool

Runs the local engine alongside Bedrock and reports differences without enforcing them. Not yet available. (true when set; pass --local-config.grounding-policy.shadow-local=false to disable).

--local-config.grounding-policy.enabled

bool

Turns contextual grounding on or off. Checks run on output only and require a grounding source and query in the request. (true when set; pass --local-config.grounding-policy.enabled=false to disable).

--local-config.grounding-policy.grounding.enabled

bool

Turns this check on or off independently. (true when set; pass --local-config.grounding-policy.grounding.enabled=false to disable).

--local-config.grounding-policy.grounding.threshold

float

Minimum score from 0.0 to 0.99. The action fires below it.

--local-config.grounding-policy.grounding.action

string

Action on a low score. None records the result. Block returns the blocked output message. (one of: none, block).

--local-config.grounding-policy.relevance.enabled

bool

Turns this check on or off independently. (true when set; pass --local-config.grounding-policy.relevance.enabled=false to disable).

--local-config.grounding-policy.relevance.threshold

float

Minimum score from 0.0 to 0.99. The action fires below it.

--local-config.grounding-policy.relevance.action

string

Action on a low score. None records the result. Block returns the blocked output message. (one of: none, block).

--local-config.automated-reasoning-policy.engine

string

Automated reasoning is available only on Amazon Bedrock, so its engine can’t be set to local. (one of: local, bedrock).

--local-config.automated-reasoning-policy.shadow-local

bool

Shadow evaluation is not available for automated reasoning. (true when set; pass --local-config.automated-reasoning-policy.shadow-local=false to disable).

--local-config.automated-reasoning-policy.enabled

bool

Turns automated reasoning on or off. Findings are added to the response trace. Requests are never blocked. (true when set; pass --local-config.automated-reasoning-policy.enabled=false to disable).

--local-config.automated-reasoning-policy.policy-arns

strings

ARNs of one or two published Bedrock Automated Reasoning policies. Each ARN ends in a numeric version such as :1. DRAFT versions are rejected. (repeatable; replaces the full list on update).

--local-config.automated-reasoning-policy.confidence-threshold

float

Confidence required for a definitive finding, from 0.0 to 1.0. Lower-confidence results are reported as ambiguous.

--tags

stringArray

Tags. (repeatable key=value; replaces the full map on update).

-f, --filename

string

manifest to create from (YAML or JSON; "-" for stdin); flags override file values.

--dry-run

bool

print the request that would be sent (with the computed update mask) and exit without calling the server help for create.

Global flags

Value Type Description

--config

string

Redpanda or rpk config file; default search paths are ~/.config/rpk/rpk.yaml, $PWD/redpanda.yaml, and /etc/redpanda/redpanda.yaml.

-X, --config-opt

stringArray

Override rpk configuration settings; -X help for detail or -X list for terser detail.

--ignore-profile

bool

Ignore rpk.yaml and redpanda.yaml; use default settings.

--profile

string

rpk profile to use.

-v, --verbose

bool

Enable verbose logging.