ai_guardrails

Provider-agnostic input/output safety for Dart & Flutter AI apps

ai_guardrails is a pure-Dart safety layer you wrap around any LLM call — local (llama.cpp, gemma, ONNX) or cloud (OpenAI, Anthropic, Gemini, your own gateway). Compose small, deterministic Scanners into an AiGuard and it redacts PII, blocks prompt injection and secret leakage on the way in, and validates model output on the way out. Python has NeMo Guardrails, Guardrails AI, and LLM Guard — this is the first for Dart: pure heuristic, zero runtime dependencies, runs on any platform including mobile and web.

Why on-device #

• EU AI Act / GDPR — redact PII before the text leaves the device.
• Privacy — secrets and PII are stripped locally; they never reach a vendor's logs.
• Offline & cheap — regex/heuristic based, zero dependencies, no per-token cost.

Use cases #

Install #

Dart SDK ^3.5.0 (Flutter 3.24+) · zero runtime dependencies.

dart pub add ai_guardrails      # or: flutter pub add ai_guardrails

import 'package:ai_guardrails/ai_guardrails.dart';

How it flows #

flowchart LR
  U([User input]) --> IS{{Input scanners}}
  IS -- blocked --> BI[GuardOutcome<br/>blocked = true]
  IS -- sanitised --> LLM[[Your LLM call]]
  LLM --> OS{{Output scanners}}
  OS -- blocked --> BO[GuardOutcome<br/>blocked = true]
  OS -- sanitised --> APP([Your app])

Scanners chain — each sees the previous one's transformed text. The pipeline stops at the first scanner that blocks.

Quickstart #

import 'package:ai_guardrails/ai_guardrails.dart';

final guard = AiGuard(
  inputScanners: [
    PiiScanner(action: GuardAction.redact),
    SecretScanner(),
    PromptInjectionScanner(threshold: 0.5),
  ],
  outputScanners: [
    SchemaValidator({
      'type': 'object',
      'required': ['answer'],
      'properties': {'answer': {'type': 'string'}},
    }),
  ],
);

final outcome = await guard.run(
  input: userText,
  llmCall: (sanitized) => myLlm.complete(sanitized),
);

if (outcome.blocked) {
  print('Rejected: ${outcome.blockReason}');
} else {
  print(outcome.output);
}

See example/ai_guardrails_example.dart for a full runnable example. Provider integration: docs/llm-providers.md.

Scanners #

31 built-in scanners. Actions: GuardAction.{ block, redact, hash, warn, transform }. Findings are dotted and predictable — pii.email, secret.aws_access_key, injection.override — so you can route or log by type.

Per-scanner configuration and code examples: docs/scanners.md.

Performance #

~1.25 KB prompt, Apple Silicon. Full methodology: BENCHMARK.md · ReDoS analysis: PERFORMANCE-AUDIT.md.

Privacy & telemetry #

Built-in heuristic scanners (the default set) make zero network calls and collect no telemetry. They run entirely on-device using synchronous string operations — no data leaves the process boundary.

LLM-assisted scanners (HallucinationScanner, FactCheckScanner, TopicSafetyScanner, EmbeddingGroundingScanner) use caller-provided callbacks (LlmCallback / EmbeddingCallback) that will typically call cloud APIs. The package itself makes no network call, but the callback you supply determines where data goes. If you use these scanners, your privacy boundary extends to whatever service backs your callback.

Documentation #

Contributing #

Contributions are issue-first: open a GitHub issue before sending a PR. See CONTRIBUTING.md.

License #

Apache-2.0 © Kryonex Labs Private Limited