How to roll out content safety without breaking your AI app

Turning on a content filter is one click. Turning it on without breaking your product takes a little planning. This guide shows how to roll out Unoblox Content safety on a live, customer-facing app: a support bot, a tutoring assistant or a community feature. What Content safety does.

Step 1: Record for a week before you block

Set When something is found to Only record it, and sensitivity to Balanced.

Everything still goes through, but every finding lands in the Safety log with its category and a confidence score. After a few days of real traffic you'll know:

• which categories actually appear in your product;
• whether findings come mostly from users' prompts or from model replies;
• which API keys, and so which features or customers, they come from.

Step 2: Pick categories for your audience

Product	Sensible starting point
Customer support bot	Harassment, hate speech, sexual content
Education or children's learning	All four, at Strict
Community or user-generated content	All four, at Balanced
Internal tools for staff	Often Record only, so nothing blocks real work

These are starting points, not rules. Your Safety log data from step 1 should decide.

Step 3: Choose the sensitivity level

• Strict stops more and may stop some safe content. Use it where a miss is worse than a false alarm, such as products for young users.
• Balanced is recommended for most products.
• Relaxed only stops clear cases. Use it when false alarms would frustrate expert users, such as medical or security professionals who discuss difficult topics.

Check the confidence scores in your log. If most findings for a category are high-confidence and correct, blocking that category is low risk.

Step 4: Switch to Stop it, and handle refusals well

When you switch to Stop it, a flagged request is refused, and a flagged reply is cut off. Your app should:

• Show a clear, friendly message, not a raw error. For example: "I can't help with that. Try rephrasing your question."
• Handle streams that stop partway. With streaming, part of a reply may already be on screen when it is cut off. Replace or clearly mark the partial text.
• Treat "check unavailable" as retryable. If a safety check can't run, Unoblox refuses the request rather than letting unchecked content through. Retry with backoff.
• Log the request ID. It matches the Safety log, so support can look up exactly what happened.

Step 5: Know what it costs

• The safety checks themselves cost nothing.
• A refused request generates no reply. If a reply is cut off partway, the part already generated is billed like a cancelled request.
• Checking adds a short delay, and streamed replies start slightly later. Measure this on your own traffic before committing to response-time promises.

Step 6: Review the log every week

Filter the Safety log by Blocked and Flagged, and sort by key and model. Look for:

• False alarms that frustrate genuine users: consider a lower sensitivity or dropping a category.
• Repeated abuse from one key or customer: handle it in your product, for example with rate limits or account review.
• Unsafe content coming from model replies: consider a different model, or a tighter system prompt.

A note on limits

Automated content checks reduce risk; they don't eliminate it. Keep your terms of use, reporting tools and human review in place for high-stakes decisions. Content safety covers harmful content. For personal, health and card data, pair it with Unoblox's data-masking guardrails.

Ready to start? Turn on Only record it today and read your first week of data. Open Guardrails or create a Unoblox account.