Guardrails

Every bot has platform safety checks you cannot switch off. They run before the model is called, in English, Hindi (Devanagari and romanised), Bengali, Assamese, Spanish and Arabic, and they see through tricks such as spaced-out letters, look-alike characters and encoded text.

Each refusal except self-harm counts as a strike. Three strikes within 10 minutes pause the visitor for 5 minutes; six block them for 24 hours. Every hit appears on the Flagged page so your team can review it.

Tip: Answers are checked too: if a reply would repeat the bot's hidden instructions, it is stopped and flagged.

Topics the bot should never discuss, one per line (up to 50, each up to 80 characters). A message that mentions one gets a fixed, polite decline in the visitor's language, with no model call, so it costs nothing. The hit is flagged as low severity and does not count as a strike.

Matching is by whole word, ignoring case, and also catches a simple plural (politics, election, elections). It does not translate: add each language and script your visitors use.

• Use single, distinctive words or short phrases. Very short terms (under 3 letters) are ignored.
• List the variants people actually type: election, elections, chunav, चुनाव.
• Do not block words your customers need. Blocking "loan" on a bank's bot would decline real questions.

Blocked topics are a fixed list. Off-topic questions are everything unrelated to your business: weather, homework, trivia, writing a poem. The model decides whether a question relates to your business, based on your knowledge.

• Off (default): the bot says in one sentence that it cannot help with that, and what it can help with.
• On: the bot gives a brief general answer and then offers help with your business.

Tip: Keep it off for most support bots. Turn it on only when friendly small talk helps, for example a coaching institute whose students ask general study questions.

List competitors (up to 50). The bot is told to stay neutral about them, and every answer is checked sentence by sentence: a sentence that names a competitor together with a negative word (worse, scam, overpriced, bekaar, ghatiya and similar) is blocked. The visitor then gets a safe reply asking them to contact you directly, and the turn is flagged.

Neutral comparisons are fine: "Both offer home delivery" passes. The check is about disparagement, not about mentioning the name.

Before an answer reaches the visitor, sentences about refunds, discounts, coupons, guarantees, prices and delivery dates are checked. A sentence that says it will not happen ("we do not offer refunds") is never blocked.

A blocked answer is replaced with: "Sorry, I can't give you a reliable answer to that. For anything about orders, refunds, prices or commitments, please contact your business directly." It is flagged as high severity so you can see which question needs better content.

Tip: The fix for a blocked but correct answer is almost always content: add the policy with its exact numbers (for example "Refunds within 7 days of delivery") to a knowledge source, and the same answer passes.

Offer a person instead of refusing: when on, refusals for jailbreaks, requests for other people's data and similar ask whether the visitor wants to talk to your team. Turn on live support in the Handoff tab first. It does not change blocked topics or abuse replies.

Moderate with my provider: also sends each visitor message to your OpenAI key's moderation endpoint (free on OpenAI). It only works with an OpenAI key and never uses a platform key. If the check is slow or fails, the message goes through.

Turnstile stops automated scripts from opening chats and running up your model bill. It is checked once when a chat starts, not on every message.

1. 01

   Create a widget in Cloudflare

   In your Cloudflare account, open Turnstile and add a widget. Add every hostname that embeds the chat, for example shop.example.com.

2. 02

   Paste both keys

   Copy the site key and the secret key into this card, turn on Require Turnstile for new chats and save. The secret is encrypted and never shown again; only its last four characters.

3. 03

   Allow it in your site's CSP

   If your site sends a Content-Security-Policy header, allow https://challenges.cloudflare.com in script-src and frame-src.

4. 04

   Test in a private window

   Open your site, start a chat and complete the check. A visitor who fails sees: Please complete the human check to start chatting.

Watch out: If Cloudflare cannot be reached, new chats are refused rather than let through. Turn it on when bots or abuse are a real problem, not by default.

Open Flagged in the sidebar. Each card shows the reason, severity, a short excerpt, the visitor and how many times it repeated within 10 minutes. Test chats in the Playground are never flagged.

• Open conversation to see the full context.
• Mark reviewed when no action is needed.
• Block visitor for 24 hours, 7 days or 30 days if someone keeps trying; unblock at any time.
• Many "Commitment blocked" flags on real questions mean your knowledge is missing a policy. Add it rather than loosening the setting.