yugenkairo-sentinel-sdk

Sentinel Python SDK

Python SDK for Sentinel - A self-healing LLM firewall with cryptographic data protection.

Overview

The Sentinel Python SDK provides a secure interface to LLM providers through the Sentinel security pipeline. It acts as a drop-in replacement for popular LLM SDKs while adding enterprise-grade security features including:

• Real-time data redaction and tokenization
• Format-preserving encryption (FF3-1)
• Semantic violation detection
• Constitutional AI reflection
• Prompt rewriting and ranking
• Tool/function call guarding

Installation

pip install yugenkairo-sentinel-sdk

Quick Start

Basic Usage

from sentinel import SentinelClient

# Initialize the client
client = SentinelClient(
    base_url="http://localhost:8080",
    api_key="your-api-key"
)

# Send a chat completion request through Sentinel
response = client.chat_completions.create(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "user", "content": "Hello, world!"}
    ]
)

print(response.choices[0].message.content)

Advanced Usage

from sentinel import SentinelClient

# Initialize with custom configuration
client = SentinelClient(
    base_url="http://localhost:8080",
    api_key="your-api-key",
    timeout=60
)

# Sanitize a prompt before sending to LLM
sanitized = client.sanitize_prompt("Process sensitive data: 123-45-6789")
print(f"Sanitized prompt: {sanitized['sanitizedPrompt']}")

# Process an LLM response for security
response = "Here's the sensitive information: 123-45-6789"
processed = client.process_response(response)
print(f"Processed response: {processed['processedResponse']}")

Features

🔐 Data Protection

• Real-time data redaction/tokenization
• Format-preserving encryption (FF3-1)
• Reversible detokenization with policy gating
• Multi-language PII detection

🛡️ Self-Healing Security

• Semantic violation detection
• Constitutional AI reflection
• Prompt rewriting and ranking
• Tool/function call guarding

🔌 Provider Compatibility

• Drop-in replacement for OpenAI SDK
• Support for all major LLM providers
• Streaming support with mid-stream inspection
• Multi-tenant policy management

⚙️ Advanced Configuration

• Policy engine integration
• Custom security rules
• Audit trails and compliance reporting
• Observability with metrics and tracing

API Reference

SentinelClient

__init__(base_url, api_key, timeout)

Initialize the Sentinel client.

Parameters:

• base_url (str): The base URL for the Sentinel gateway (default: "http://localhost:8080")
• api_key (str, optional): API key for authentication
• timeout (int): Request timeout in seconds (default: 30)

sanitize_prompt(prompt)

Sanitize a prompt before sending to LLM.

Parameters:

• prompt (str): The prompt to sanitize

Returns:

• dict: Sanitized prompt and metadata

process_response(response)

Process an LLM response for security.

Parameters:

• response (str): The LLM response to process

Returns:

• dict: Processed response and metadata

configure_policies(policies)

Configure security policies.

Parameters:

• policies (dict): Policy configuration

Returns:

• dict: Policy update result

ChatCompletions

create(model, messages, temperature, max_tokens, **kwargs)

Create a chat completion through the Sentinel gateway.

Parameters:

• model (str): The model to use
• messages (list): List of message dictionaries
• temperature (float, optional): Sampling temperature
• max_tokens (int, optional): Maximum tokens to generate
• **kwargs: Additional parameters

Returns:

• dict: Chat completion response

Configuration

Environment Variables

• SENTINEL_BASE_URL: Default base URL for the Sentinel gateway
• SENTINEL_API_KEY: Default API key for authentication
• SENTINEL_TIMEOUT: Default request timeout in seconds

Configuration File

You can also configure the client using a configuration file:

import os
from sentinel import SentinelClient

# Load configuration from environment
client = SentinelClient(
    base_url=os.getenv("SENTINEL_BASE_URL", "http://localhost:8080"),
    api_key=os.getenv("SENTINEL_API_KEY"),
    timeout=int(os.getenv("SENTINEL_TIMEOUT", "30"))
)

Error Handling

The SDK raises standard Python exceptions:

from sentinel import SentinelClient
import requests

client = SentinelClient(base_url="http://localhost:8080")

try:
    response = client.chat_completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Hello"}]
    )
except requests.exceptions.RequestException as e:
    print(f"Request failed: {e}")
except Exception as e:
    print(f"An error occurred: {e}")

Examples

Multi-tenant Usage

from sentinel import SentinelClient

# Different clients for different tenants
tenant_a_client = SentinelClient(
    base_url="http://localhost:8080",
    api_key="tenant-a-key"
)

tenant_b_client = SentinelClient(
    base_url="http://localhost:8080",
    api_key="tenant-b-key"
)

Custom Policy Configuration

from sentinel import SentinelClient

client = SentinelClient(base_url="http://localhost:8080")

# Configure custom policies
policies = {
    "pii_detection": {
        "enabled": True,
        "languages": ["en", "es", "fr"],
        "action": "tokenize"
    },
    "prompt_filtering": {
        "enabled": True,
        "threshold": 0.75
    }
}

result = client.configure_policies(policies)
print(f"Policies configured: {result['success']}")

Integration with Popular Frameworks

LangChain Integration

from langchain.llms import Sentinel
from langchain.prompts import PromptTemplate

llm = Sentinel(
    base_url="http://localhost:8080",
    api_key="your-api-key"
)

template = "What is {subject}?"
prompt = PromptTemplate.from_template(template)
chain = prompt | llm

response = chain.invoke({"subject": "artificial intelligence"})
print(response)

LlamaIndex Integration

from llama_index.llms import Sentinel
from llama_index import VectorStoreIndex, SimpleDirectoryReader

llm = Sentinel(
    base_url="http://localhost:8080",
    api_key="your-api-key"
)

documents = SimpleDirectoryReader("data").load_data()
index = VectorStoreIndex.from_documents(documents, llm=llm)
query_engine = index.as_query_engine()

response = query_engine.query("What did the author do growing up?")
print(response)

Development

Installation from Source

git clone https://github.com/swayam8624/Sentinel.git
cd Sentinel/sdk/python
pip install -e .

Running Tests

pip install pytest
pytest tests/

Code Formatting

pip install black flake8
black .
flake8 .

Documentation

For full documentation, visit https://swayam8624.github.io/Sentinel/

Support

For issues, feature requests, or questions, please open an issue on GitHub.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.