Guardrails AI

What it is

Composable validators with repair.

Guardrails AI is a framework for input and output guardrails assembled from small, composable validators: structure, PII, toxicity, competitor mentions, and many more. When a validator fails, the framework can re-ask the model or fix the output rather than simply erroring. A validators hub ships prebuilt checks you can drop in, so most common rules are a line of configuration rather than new code.

The one job

Validate and repair the output.

It sits after the model and enforces your rules on what came back: the shape is right, no PII leaked, nothing toxic, no competitor named. Where it can, it repairs the output to pass rather than dropping it, so the contract is met before the value moves on.

Reach for it when

• You need structural and semantic output checks with automatic repair on failure.
• You want a catalog of ready validators rather than writing each rule yourself.
• You need typed outputs the rest of your code can rely on.
• You want input and output rails from one framework across Python or JS.

Skip it when

• Native structured outputs already cover your structure needs: Zod plus the SDK is enough.
• You specifically need injection detection: Lakera is the focused tool for that.
• You want minimal dependencies and a validator framework is more than the job requires.

A minimal look

Run validators over the output.

import { Guard } from "@guardrails-ai/core";

// declare the rules once; each validator can fix or re-ask on failure
const guard = await Guard()
  .use(NoPII, { onFail: "fix" })
  .use(NoCompetitors, { onFail: "reask" });

const { validatedOutput } = await guard.validate(modelOutput);
// validatedOutput is now safe to return to the caller
return validatedOutput;

The principle it teaches

Where it fits your labs

This spans Week 4 and Week 1. It is the output guardrail in Lab 4, and the same validator idea backs the validation work from Week 1 where a value has to pass a schema before you trust it.