https://github.com/langchain-samples/framework-agnostic-guardrails
Framework-agnostic guardrails modeled after LangChain pre-built guardrails middleware
https://github.com/langchain-samples/framework-agnostic-guardrails
beginner guardrails middleware
Last synced: 3 months ago
JSON representation
Framework-agnostic guardrails modeled after LangChain pre-built guardrails middleware
- Host: GitHub
- URL: https://github.com/langchain-samples/framework-agnostic-guardrails
- Owner: langchain-samples
- License: mit
- Created: 2025-12-19T00:52:54.000Z (7 months ago)
- Default Branch: main
- Last Pushed: 2025-12-19T03:08:22.000Z (7 months ago)
- Last Synced: 2025-12-21T23:46:55.865Z (7 months ago)
- Topics: beginner, guardrails, middleware
- Language: Python
- Homepage:
- Size: 174 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Framework-Agnostic Guardrails
A simple, framework-agnostic wrapper for adding guardrails to AI agents at the application level.
## Guard() Wrapper
`Guard()` is a wrapper that provides input and output validation for AI agent calls. It operates as application-level middleware:
1. Validates user input before execution (e.g., prompt injection detection, PII screening)
2. Executes your agent using any framework
3. Validates agent output before returning results (e.g., regulatory compliance, hallucination detection)
We believe **application-level guardrails are more effective** than placing guardrails in the observability layer. Security and compliance decisions should be made where they matter most - in your application code, before and after your agent runs.
This project demonstrates how easy it is to integrate guardrails into your agent using open-source tools like [OpenEvals](https://github.com/langchain-ai/openevals).
## Quick Start
### Installation
Using [UV](https://docs.astral.sh/uv/) (recommended):
```bash
# Install dependencies
uv sync
# Set up your OpenAI API key
cp .env.example .env
# Edit .env and add your OPENAI_API_KEY
```
### Basic Usage
```python
from guard import Guard, PIIDetectionGuardrail, JailbreakGuardrail
# Create guard with multiple guardrails
guard = Guard([
PIIDetectionGuardrail(), # Pattern-based
JailbreakGuardrail(), # LLM-based
])
# Works with ANY agent function
def my_agent(user_input: str) -> str:
# Your agent logic (OpenAI, Anthropic, whatever)
return client.chat.completions.create(...)
# Run with guardrails
result = guard.run(
user_input="What's the weather?",
agent_call=my_agent
)
if result.blocked:
print(f"Blocked: {result.findings}")
else:
print(f"Safe: {result.output}")
```
### Flexible Input/Output Validation
Each guardrail can be configured to check inputs, outputs, or both:
```python
guard = Guard([
# Check only user input for jailbreak attempts
JailbreakGuardrail(check_input=True, check_output=False),
# Check only agent output for compliance
RegulatoryComplianceGuardrail(check_input=False, check_output=True),
# Check both input and output for PII
PIIDetectionGuardrail(check_input=True, check_output=True),
])
```
**Smart defaults:** Each guardrail comes with sensible defaults based on its purpose (e.g., jailbreak checks input only, compliance checks output only).
## Available Guardrails
### Pattern-Based (Fast, No API Required)
- **PIIDetectionGuardrail** - Block emails, SSNs, phone numbers
- **ToxicLanguageGuardrail** - Block offensive content
### LLM-Based (Semantic, Uses OpenEvals)
- **JailbreakGuardrail** - Detect manipulation attempts
- **RegulatoryComplianceGuardrail** - Ensure financial/legal compliance
- **BiasDetectionGuardrail** - Catch discriminatory content
- **CorrectnessGuardrail** - Validate factual accuracy
- **HallucinationGuardrail** - Detect made-up information
### Custom
- **CustomGuardrail** - Your own business logic
## Examples
See [`examples.py`](examples.py) for detailed demonstrations of each guardrail, including:
- Input vs output validation
- Pattern-based vs LLM-based checks
- Custom business logic
- Multiple guardrails working together
## Credits
Built using:
- [OpenEvals](https://github.com/langchain-ai/openevals) - LLM-as-judge framework
- Inspired by [LangChain's guardrails middleware](https://docs.langchain.com/oss/python/langchain/guardrails)
## License
MIT