Choosing a Guardrail Provider

AISIX enforces content policy through guardrails: checks that run on the request before it reaches the upstream model, on the response before it reaches the caller, or both. Each guardrail is one kind with its own provider, capabilities, and configuration. This page compares the supported kinds so you can choose the right one before reading its dedicated guide.

Provider Comparison

Kind	Provider	Deployment	Detects	Actions	Hooks
keyword	Built-in	In-process	Literal / regex patterns	block	input, output
pii	Built-in	In-process	Rule-based PII (email, phone, ID, card, keys, …) + custom regex	mask, block	input, output
presidio	Microsoft Presidio	Self-hosted containers	Rule and NER/ML PII entities (PERSON, LOCATION, …)	mask (replace / mask / hash / redact operators), block	input, output
lakera	Lakera Guard	Managed cloud API	Prompt injection, jailbreak, content, PII	block; PII-only detections mask	input, output
azure_content_safety	Azure AI Content Safety	Managed cloud API	Prompt injection / jailbreak (Prompt Shield)	block	input, output
azure_content_safety_text_moderation	Azure AI Content Safety	Managed cloud API	Category severity (hate, sexual, self-harm, violence) + blocklists	block	input, output
openai_moderation	OpenAI Moderation	Managed cloud API (free)	Content categories (harassment, hate, violence, sexual, self-harm, …)	block	input, output
bedrock	AWS Bedrock Guardrails	Managed cloud service	Content filters, denied topics, sensitive information, word filters	block; PII ANONYMIZE masks	input, output
aliyun_text_moderation	Alibaba Cloud AI Guardrails	Managed cloud API	Risk-level content moderation	block	input, output

All kinds share the same platform controls — hook points, enforcement modes, mandatory fail-closed policy, and streaming hold-back — described in Guardrail Behavior.

Picking by Goal

Detect or redact PII. Start with the built-in pii kind: it runs in-process with no external calls and covers rule-based detectors plus custom regex. Move to presidio when you need entities a regex cannot express (person names, locations, nationalities), a self-hosted compliance posture, or hash/placeholder replacement instead of a fixed redaction token. AWS Bedrock sensitive-information filters and Lakera PII detection are alternatives when you already use those providers.

Detect prompt injection and jailbreaks. Lakera Guard and Azure Prompt Shield both screen adversarial input, including indirect injection carried in retrieved documents or tool results that appear in the conversation. They are independent providers — pick the one your security review names, or attach both.

Moderate content categories. OpenAI Moderation is free and covers the standard category set. Azure Text Moderation adds per-category severity thresholds and server-side blocklists; Alibaba Cloud is the natural choice inside that ecosystem.

Enforce organization-specific rules. The built-in keyword kind blocks literal or regex patterns with no external dependency.

Detection Is Probabilistic

Every detection-based guardrail — PII recognizers, injection classifiers, category moderation — carries false-positive and false-negative risk. Before enforcing a new guardrail on production traffic:

1. Create it with enforcement_mode: monitor. The guardrail evaluates every request and records what it would have blocked or masked, without affecting traffic.
2. Review the observations and tune the configuration — detector selection, thresholds, entities.
3. Switch the guardrail to enforcement_mode: block once the hit rate looks right.

See Guardrail Behavior for the mechanics.

What the Gateway Records

Every request records which guardrails governed it, whether one blocked it, and (for fail-open bypasses) why a guardrail was skipped. Matched content never appears in gateway logs, usage records, or error responses — only detector, entity, or category names. See the per-kind guides for the exact redaction-token and block-envelope shapes.

Next Steps

Read Guardrail Behavior for the platform controls shared by every kind, then follow the dedicated guide for the provider you picked.