Tutorial: Add Prompt Filtering to Your App in 15 Minutes

What You'll Build

By the end of this tutorial you will have a single filtering call in front of your model invocation that returns block, sanitize, or allow — and you will know how to roll it out safely.

Prerequisites

• An app that calls an LLM
• The ability to make an HTTP POST before that call
• An API key for the hosted tier (the offline pattern engine needs no key)

Step 1: Make Your First Call

Send any prompt to the filter endpoint:

curl -X POST https://api.sprapp.com/v1/filter \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_KEY" \
  -d '{"input": "Hello, how do I reset my password?"}'

A benign prompt returns {"action": "allow"}. Now try a hostile one and watch it flip to block.

Step 2: Wrap Your Model Call

async function safeComplete(prompt) {
  const res = await fetch("https://api.sprapp.com/v1/filter", {
    method: "POST",
    headers: { "Content-Type": "application/json",
               "Authorization": "Bearer " + KEY },
    body: JSON.stringify({ input: prompt }),
  });
  const { action, sanitized } = await res.json();

  if (action === "block") throw new Error("Request blocked");
  const finalPrompt = action === "sanitize" ? sanitized : prompt;
  return callYourModel(finalPrompt);
}

Step 3: Start in Log-Only Mode

Do not enforce on day one. Run the filter alongside your normal flow, log every verdict, and forward all prompts as usual. This tells you the false-positive rate on your real traffic before you start blocking anything.

Step 4: Review the Verdicts

After a day or two, look at what got flagged. You will see three buckets: genuine threats (good catches), borderline cases (tune your thresholds), and false positives (legitimate prompts the filter misread). Use this to set per-category sensitivity.

Step 5: Turn On Enforcement

Once the log-only data looks clean, switch block and sanitize to actually take effect. Start with the highest-confidence categories — credential theft, structured PII, obvious injection — and expand from there.

Step 6: Add the Offline Fast-Path (Optional)

For latency-critical paths, run the free pattern engine inline and only escalate uncertain prompts to the hosted cascade. Most traffic resolves locally in sub-millisecond time.

A Note on Expectations

The pattern tier catches roughly 95% of clear-cut threats and the transformer cascade lifts the ambiguous band to 97.1%. It is a strong control, not a guarantee. Keep your other defenses in place.

You're Done

One POST in front of your model, rolled out from log-only to enforcement. Paste prompts into the live demo at https://filter.sprapp.com to see the category breakdown as you tune.