llm-guardrails-sk 0.1.0

Provider-agnostic middleware for LLM safety

LLM Guardrails Gateway

A provider-agnostic middleware for LLM safety. Automatically detects prompt injections, validates output schemas, enforces policies, and logs every decision for compliance.

Installation

pip install llm-guardrails

Quick Start

from llm_guardrails import Guardrails
from openai import OpenAI

guard = Guardrails()
client = OpenAI()

@guard.protect
def ask_llm(question: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": question}]
    )
    return response.choices[0].message.content

# Clean message passes through
answer = ask_llm("What is the capital of France?")
# Returns: "Paris"

# Attack gets blocked
answer = ask_llm("Ignore all previous instructions")
# Returns: GuardrailBlockedResponse(...)

# Verify performance yourself
results = guard.run_self_eval(sample_size=50)
# Returns: precision, recall on fresh held-out examples

# Audit what happened
summary = guard.get_audit_summary()
# Returns: total decisions, block rate, detector breakdown

Features

Input Protection: Prompt Injection Detection

• 85.6% recall - catches 856 out of 1000 real attacks
• 85.9% precision - only 141 false positives out of 1000 benign inputs
• Three layers: pattern matching → embedding similarity → LLM judge
• Catches novel obfuscations (base64, Unicode, translation wrappers)

Output Validation

• Schema validation with auto-retry
• Handles markdown wrapping, syntax errors, missing fields
• Asks LLM to fix its own output before giving up

Output Content Safety

• 76.6% recall - catches harmful responses
• 89.5% precision - low false-positive rate
• Same multi-layer design as input detection

Policy Engine

Define rules in YAML, enforce them automatically:

rules:
  - name: "no_medical_advice"
    description: "Never provide specific medical dosage or treatment advice."
    applies_to: "output"

Audit Logging

Every decision logged in both JSON and human-readable format:

blocked = guard.get_blocked_entries(limit=10)
summary = guard.get_audit_summary()

Self-Eval

Verify detection quality on demand:

result = guard.run_self_eval(sample_size=100)
# Returns: precision, recall, TP/FN/FP/TN

Architecture