Skip to main content

Configure guardrails

Guardrails filter user inputs and AI outputs for a chat experience. This guide walks through setting up a starter configuration and tuning it.

Before you start​

1. Open the guardrails card​

Sidebar → Experiences → open your AI experience → expand the Guardrails card.

You'll see two sections — Input guardrails and Output guardrails — each with its own enable toggle.

2. Configure input guardrails​

Input guardrails run on the user's message before the AI sees it. They prevent off-topic conversations and known-bad terms from reaching the AI.

Step 1 — Turn on input guardrails​

Toggle the Input guardrails enabled switch.

Step 2 — Topic gating​

Topic gating keeps the chat on-domain.

  1. Toggle Domain filter enabled.
  2. Allowed domains — enter a comma-separated list of keywords describing the chat's domain. For a fashion product chat: 
    fashion, clothing, apparel, shoes, accessories, shipping, returns, sizing
  3. Friendly message — what the chat says when it rejects an off-topic message:

    "I can only help with fashion shopping. Can I help you find something else, or have a question about an order?"

  4. Threshold — start at 0.6 (middle ground). Raise if off-topic gets through; lower if legitimate questions get rejected.
  5. Click Generate Domain Terms. Interakt uses AI to expand your keywords into 20–30 semantically related terms. The expanded list is what the topic-gating logic uses — your keywords are the seed.

Verify the expansion looks right (under the Expanded Terms section). If it includes things you don't want covered, narrow your seed keywords and re-generate.

Step 3 — Blocklist (optional)​

Hard exact-match filter for known-bad strings.

  1. Blocked terms — comma-separated list of phrases that should never appear in messages: competitor names, internal codenames, profanity.
  2. Action — pick one:
    • Block — reject the message outright.
    • Warn — let it through but log a warning.
    • Redact — replace the matched text with [REDACTED] and continue.

For most customer-facing chats, Block is right for competitors and profanity.

3. Configure output guardrails​

Output guardrails run on the AI's response before it goes to the user. They're the backstop — even if the AI generates something off-policy, this catches it.

Same structure as input. Common configuration:

  1. Toggle Output guardrails enabled.
  2. Topic gating — usually less aggressive than input. Use the same allowed-domains list, but a lower threshold (the AI is generally on-topic, so you're only catching outliers).
  3. Blocklist — terms the AI should never say. Competitor mentions, deprecated SKUs, anything bad-for-brand.

4. Save​

Click Save on the experience or whatever the save button is at the top of the page.

5. Test in the Chat Playground​

Expand the Chat Playground card on the same experience detail page. Try a battery of test messages:

Test messageExpected outcome
In-domain question ("do you have red sneakers?")Normal response.
Clear off-topic ("help me write a poem")Blocked. Friendly message shown.
Greeting ("hi", "thanks")Should pass through (greeting detection handles it).
Edge case ("do you have any sales going on?")Should pass — sales is a fashion-shopping topic. If it gets blocked, broaden allowed domains or generate more terms.
Blocked term ("is your product better than {competitor}?")Blocked.

The playground shows tool calls and rejections in the sidebar — you can see which rule fired.

6. Iterate​

SymptomFix
Off-topic messages get throughRaise the threshold, or expand the allowed-domains list and re-generate terms.
Legitimate questions get blockedLower the threshold, or expand allowed-domains.
Greeting got blockedShould not happen — greeting detection runs first. If it does, the friendly message is too aggressive; review the topic-gating config.
Blocked term not caughtCheck the blocklist for typos. The match is substring-based but case-insensitive.
Chat feels slowTopic gating adds latency on every message. Skip output guardrails if input is sufficient.

7. (Production) Disable in development​

While iterating on the chat's persona and tools, you can toggle guardrails off so you can test rapidly. Don't forget to re-enable before going live.

Special case — Help Assistant style​

If you're building an internal documentation chatbot (an "ask anything about our system" assistant), turn guardrails off entirely. Topic gating defeats the purpose of an open-ended help bot. The built-in Help Assistant works this way.

Common gotchas​

  • Threshold too strict. Default 0.6 is a starting point — most chats should sit between 0.5 and 0.7. Cranking to 0.9 blocks half of legitimate queries.
  • Generated terms list is stale. If you change keywords without clicking Generate, the chat uses the old expansion. The status badge says "Stale" — fix by clicking Generate.
  • No friendly message set. Defaults to a generic "I can't help with that." Always customise.
  • Forgetting output guardrails. Input clean doesn't mean output clean — the AI can hallucinate things you don't want said. Mirror critical filters on output.
  • Blocklist for things topic-gate should catch. "weather" doesn't need to be blocklisted in a fashion chat — topic-gating already rejects it. Use blocklist for exact phrases that need exact-match precision.

Where to go next​