lexigram-ai-guard

AI input/output guard pipeline for the Lexigram Framework — LLM safety and content filtering

Overview

Content safety and guardrails for the Lexigram Framework. Runs an ordered pipeline of input and output guards against every LLM call — blocking prompt injections, redacting PII, enforcing length limits, restricting topics, and optionally using an LLM classifier for advanced jailbreak detection. Zero-config usage starts with sensible defaults.

Install

uv add lexigram-ai-guard

Quick Start

from lexigram import Application
from lexigram.di.module import Module, module

from lexigram.ai.guard import GuardModule
from lexigram.ai.guard.config import GuardConfig

@module(imports=[
    GuardModule.configure(
        GuardConfig(
            injection_detection=True,
            pii_detection=True,
            pii_action="redact",
            max_input_chars=8000,
        )
    )
])
class AppModule(Module):
    pass

app = Application(modules=[AppModule])
if __name__ == "__main__":
    app.run()

Configuration

Zero-config usage: Call GuardModule.configure() with no arguments to use defaults.

Option 1 — YAML file

# application.yaml
ai_guard:
  injection_detection: true
  pii_detection: true
  pii_action: "redact"
  max_input_chars: 8000
  restricted_topics: []

Option 2 — Profiles + Environment Variables (recommended)

export LEX_AI_GUARD__PII_ACTION=redact
# Environment variables for each field

Option 3 — Python

from lexigram.ai.guard.config import GuardConfig
from lexigram.ai.guard import GuardModule

config = GuardConfig(
    injection_detection=True,
    pii_action="redact",
    max_input_chars=8000,
    restricted_topics=["violence", "adult_content"],
)
GuardModule.configure(config)

Config reference

Module Factory Methods

Key Features

• Input guards: PromptInjectionDetector, PIIDetector, TopicRestrictor, InputLengthGuard
• Output guards: PIIRedactor, OutputLengthGuard
• LLM-based guards: Optional LLMInjectionDetector and LLMJailbreakDetector
• Guard results: AggregateGuardResult with passed, blocked, redacted, warned states
• @guarded decorator: Function-level guard attachment
• Lifecycle hooks: GuardInputCheckedHook, GuardOutputCheckedHook, GuardPipelineCompletedHook

Testing

async with Application.boot(modules=[GuardModule.stub(
    GuardConfig(
        injection_detection=True,
        injection_action="block",
    )
)]) as app:
    # your test code
    ...

Key Source Files