Active AI Guardrails (AAIG) screen the incoming user message before AskRAI does any work on it. Every message is evaluated against your configured guardrail rules, and if any rule is violated the request is stopped before the knowledge pipeline runs — so no answer is generated, retrieved, or delivered.
Guardrails inspect the user's question, not the assistant's answer. Each guardrail evaluates the incoming message on its own; there is no output or response guardrail in AskRAI today. A rule phrased as a constraint on the response — "ensure answers include a disclaimer", "keep responses on brand" — cannot be enforced, because no guardrail ever sees the generated text. Write your rules as tests of the request: block the kinds of questions you do not want answered.
How Guardrails Work
Each guardrail contains a natural-language rule that a language model evaluates independently against the user's message. Guardrails run concurrently (up to three at a time) to minimize latency, and each evaluation is recorded in the audit trail. They run before knowledge search — a failing guardrail short-circuits the pipeline entirely, which is also why a blocked request costs you no retrieval or summarisation tokens.
Three key principles govern how guardrails operate:
- Independent evaluation — each guardrail is evaluated by a separate LLM call, so rules never interfere with one another. A safety guardrail cannot affect how a compliance guardrail evaluates the same message.
- Tripwire blocking — if any single guardrail fails, the entire response is blocked. There is no "majority rules" — one violation is enough to prevent delivery.
- Concurrent execution — guardrails run in parallel to keep response times low. The system does not wait for one guardrail to finish before starting the next.
Five Categories
Guardrails are organized into five categories, each targeting a different type of risk:
Safety guardrails stop harmful, abusive, or dangerous requests from entering the system.
Example rules:
- Block requests that contain threats or harassment
- Block requests that ask for medical, legal, or financial advice
- Flag messages that discuss self-harm
Compliance guardrails enforce regulatory and organizational policy requirements.
Example rules:
- Block questions on regulated topics the organization may not advise on
- Block discussion of topics outside the organization's mandate
- Block requests that would breach data handling policy for sensitive information
Quality guardrails keep low-value or off-topic questions out of the pipeline.
Example rules:
- Reject queries that are too vague to produce a useful response
- Reject questions outside the organization's domain
- Block nonsensical or spam-like input
Privacy guardrails protect personally identifiable information (PII) and sensitive data.
Example rules:
- Block messages containing social security numbers, credit card numbers, or addresses
- Prevent the assistant from requesting personal information
- Redact or reject queries that expose sensitive employee data
Custom guardrails address organization-specific requirements that don't fit the other categories.
Example rules:
- Reject questions asked in an unsupported language
- Block requests about competitors or sensitive internal topics
- Block requests that try to elicit internal-only material
Priority and Reporting
Each guardrail has two additional configuration options:
- Priority (1–10) — a label for admin triage. It is the default sort order of the guardrails list and nothing more: it does not affect evaluation order (all guardrails run in parallel), and it does not rank them in analytics, which sort by how often each guardrail fires.
- Reporting level (None / Alerts / All) — intended to control how much of each evaluation reaches the audit trail.
Reporting level has no effect today. Whatever you choose, every guardrail evaluation is written to the audit trail and counted in the analytics — a guardrail set to None is audited exactly like one set to All. The setting is stored and displayed, but no code reads it. Do not rely on it to keep evaluations out of the audit record.
Guardrails and Groups
Guardrails are assigned to groups, not applied globally. This means different user populations can have different safety rules. For example:
- An internal employees group might have relaxed quality guardrails but strict compliance rules
- A public access group might have aggressive safety and privacy guardrails
- An admin group might bypass certain guardrails for testing purposes
A user's effective guardrails are the union of all guardrails from every group they belong to.
False Positive Tracking
Not every guardrail trigger is a real violation. AskRAI supports false positive marking in the conversation logs, allowing you to:
- Identify guardrails that fire too aggressively
- Track false positive rates over time in the guardrail analytics
- Refine guardrail prompts to reduce false triggers without weakening protection
Next Steps
- Guardrails — create, edit, and monitor guardrails in the admin console
- Confidence & Escalation — understand what happens after guardrails pass
- Governance & Audit — learn how guardrail evaluations feed into the audit trail
- Sandbox — test guardrail behavior before deploying to production