xiangxinai 2.6.3

Xiangxin AI guardrails Python SDK - An LLM-based context-aware AI guardrail that understands conversation context for security, safety and data leakage detection.

Xiangxin AI Guardrails Python Client

An LLM-based context-aware AI guardrail that understands conversation context for security, safety and data leakage detection.

Features

• 🧠 Context Awareness – Based on LLM conversation understanding rather than simple batch detection
• 🔍 Prompt Injection Detection – Detects malicious prompt injections and jailbreak attacks
• 📋 Content Compliance Detection – Complies with generative AI safety requirements
• 🔐 Sensitive Data Leak Prevention – Detects and prevents personal or corporate data leaks
• 🧩 User-level Ban Policy – Supports user-granular risk recognition and blocking strategies
• 🖼️ Multimodal Detection – Supports image content safety detection
• 🛠️ Easy Integration – OpenAI-compatible API format; plug in with one line of code
• ⚡ OpenAI-style API – Familiar interface design for rapid adoption
• 🚀 Sync/Async Support – Supports both synchronous and asynchronous calls for different scenarios

Installation

pip install xiangxinai

Quick Start

Basic Usage

from xiangxinai import XiangxinAI

# Create a client
client = XiangxinAI(
    api_key="your-api-key",
    base_url="https://api.xiangxinai.cn/v1"  # Cloud API
)

# Check user input
result = client.check_prompt("I want to learn Python programming", user_id="user-123")
print(result.suggest_action)        # Output: pass
print(result.overall_risk_level)    # Output: no_risk
print(result.score)                 # Confidence score, e.g. 0.9993114447238793

# Check model response (context-aware)
result = client.check_response_ctx(
    prompt="Teach me how to cook",
    response="I can teach you some simple home dishes",
    user_id="user-123"  # Optional user-level risk control
)
print(result.suggest_action)      # Output: pass
print(result.overall_risk_level)  # Output: no_risk

Context-Aware Detection (Core Feature)

# Context-based conversation detection - Core feature
messages = [
    {"role": "user", "content": "I want to learn chemistry"},
    {"role": "assistant", "content": "Chemistry is an interesting subject. What part would you like to learn?"},
    {"role": "user", "content": "Teach me reactions for making explosives"}
]

result = client.check_conversation(messages, user_id="user-123")
print(result.overall_risk_level)
print(result.suggest_action)  # Result based on full conversation context
if result.suggest_answer:
    print(f"Suggested answer: {result.suggest_answer}")

Asynchronous API (Recommended)

import asyncio
from xiangxinai import AsyncXiangxinAI

async def main():
    async with AsyncXiangxinAI(api_key="your-api-key") as client:
        # Async prompt check
        result = await client.check_prompt("I want to learn Python programming")
        print(result.suggest_action)  # Output: pass
        
        # Async conversation context check
        messages = [
            {"role": "user", "content": "I want to learn chemistry"},
            {"role": "assistant", "content": "Chemistry is an interesting subject. What part would you like to learn?"},
            {"role": "user", "content": "Teach me reactions for making explosives"}
        ]
        result = await client.check_conversation(messages)
        print(result.overall_risk_level)

asyncio.run(main())

Concurrent Processing

import asyncio
from xiangxinai import AsyncXiangxinAI

async def batch_check():
    async with AsyncXiangxinAI(api_key="your-api-key") as client:
        # Handle multiple requests concurrently
        tasks = [
            client.check_prompt("Content 1"),
            client.check_prompt("Content 2"),
            client.check_prompt("Content 3")
        ]
        results = await asyncio.gather(*tasks)
        
        for i, result in enumerate(results):
            print(f"Content {i+1}: {result.overall_risk_level}")

asyncio.run(batch_check())

Multimodal Image Detection

Supports multimodal detection for image content safety. The system analyzes both text prompt semantics and image semantics for risk.

from xiangxinai import XiangxinAI

client = XiangxinAI(api_key="your-api-key")

# Check a single local image
result = client.check_prompt_image(
    prompt="Is this image safe?",
    image="/path/to/image.jpg"
)
print(result.overall_risk_level)
print(result.suggest_action)

# Check an image from URL
result = client.check_prompt_image(
    prompt="",  # prompt can be empty
    image="https://example.com/image.jpg"
)

# Check multiple images
images = [
    "/path/to/image1.jpg",
    "https://example.com/image2.jpg",
    "/path/to/image3.png"
]
result = client.check_prompt_images(
    prompt="Are all these images safe?",
    images=images
)
print(result.overall_risk_level)

Async version:

import asyncio
from xiangxinai import AsyncXiangxinAI

async def check_images():
    async with AsyncXiangxinAI(api_key="your-api-key") as client:
        # Async check for a single image
        result = await client.check_prompt_image(
            prompt="Is this image safe?",
            image="/path/to/image.jpg"
        )
        print(result.overall_risk_level)

        # Async check for multiple images
        images = ["/path/to/image1.jpg", "/path/to/image2.jpg"]
        result = await client.check_prompt_images(
            prompt="Are these images safe?",
            images=images
        )
        print(result.overall_risk_level)

asyncio.run(check_images())

On-Premise Deployment

# Sync client connecting to local deployment
client = XiangxinAI(
    api_key="your-local-api-key",
    base_url="http://localhost:5000/v1"
)

# Async client connecting to local deployment
async with AsyncXiangxinAI(
    api_key="your-local-api-key",
    base_url="http://localhost:5000/v1"
) as client:
    result = await client.check_prompt("Test content")

API Reference

XiangxinAI Class (Synchronous)

Initialization Parameters

• api_key (str): API key
• base_url (str): Base API URL, defaults to the cloud endpoint
• timeout (int): Request timeout, default 30 seconds
• max_retries (int): Maximum retry count, default 3

Methods

check_prompt(content: str, user_id: Optional[str] = None) -> GuardrailResponse

Checks the safety of a single prompt.

Parameters:

• content: Text content to be checked
• user_id: Optional tenant user ID for per-user risk control and auditing

Returns: GuardrailResponse object

check_conversation(messages: List[Message], model: str = "Xiangxin-Guardrails-Text", user_id: Optional[str] = None) -> GuardrailResponse

Checks conversation context safety (core feature).

Parameters:

• messages: List of messages, each containing role and content
• model: Model name (default: "Xiangxin-Guardrails-Text")
• user_id: Optional tenant user ID

Returns: GuardrailResponse object

AsyncXiangxinAI Class (Asynchronous)

Same initialization parameters as the synchronous version.

Methods

async check_prompt(content: str) -> GuardrailResponse

Asynchronously checks a single prompt.

async check_conversation(messages: List[Message]) -> GuardrailResponse

Asynchronously checks conversation context safety (core feature).

async health_check() -> Dict[str, Any]

Checks API service health.

async get_models() -> Dict[str, Any]

Retrieves available model list.

async close()

Closes async session (automatically handled with async with).

GuardrailResponse Class

Represents detection results.

Attributes

• id: Unique request ID
• result.compliance.risk_level: Compliance risk level
• result.security.risk_level: Security risk level
• result.data.risk_level: Data leak risk level (added in v2.4.0)
• result.data.categories: Detected sensitive data types (added in v2.4.0)
• overall_risk_level: Overall risk level (none / low / medium / high)
• suggest_action: Suggested action (pass / block / substitute)
• suggest_answer: Suggested response (optional, includes redacted content if applicable)
• score: Confidence score of the results

Helper Methods

• is_safe: Whether the content is safe
• is_blocked: Whether the content is blocked
• has_substitute: Whether a substitute answer is provided
• all_categories: Get all detected risk categories

Safety Detection Capabilities

Risk Levels

• High Risk: Sensitive political topics, national image damage, violent crime, prompt attacks
• Medium Risk: General political topics, harm to minors, illegal acts, sexual content
• Low Risk: Hate speech, insults, privacy violations, commercial misconduct
• No Risk: Safe content

Handling Strategies

• High Risk: Recommend blocking
• Medium Risk: Recommend substitution with a safe reply
• Low Risk: Recommend substitution or business-dependent handling
• No Risk: Recommend pass

Error Handling

Synchronous Error Handling

from xiangxinai import XiangxinAI, AuthenticationError, ValidationError, RateLimitError

try:
    result = client.check_prompt("Test content")
except AuthenticationError:
    print("Invalid API key")
except ValidationError as e:
    print(f"Input validation failed: {e}")
except RateLimitError:
    print("Rate limit exceeded")
except Exception as e:
    print(f"Other error: {e}")

Asynchronous Error Handling

import asyncio
from xiangxinai import AsyncXiangxinAI, AuthenticationError, ValidationError, RateLimitError

async def safe_check():
    try:
        async with AsyncXiangxinAI(api_key="your-api-key") as client:
            result = await client.check_prompt("Test content")
            return result
    except AuthenticationError:
        print("Invalid API key")
    except ValidationError as e:
        print(f"Input validation failed: {e}")
    except RateLimitError:
        print("Rate limit exceeded")
    except Exception as e:
        print(f"Other error: {e}")

asyncio.run(safe_check())

Development

# Clone the project
git clone https://github.com/xiangxinai/xiangxin-guardrails
cd xiangxin-guardrails/client

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Code formatting
black xiangxinai
isort xiangxinai

# Type checking
mypy xiangxinai

License

This project is open-sourced under the Apache 2.0 license.

Support

• 📧 Technical Support: [e-mail]
• 🌐 Official Website: https://xiangxinai.cn
• 📖 Documentation: https://docs.xiangxinai.cn
• 🐛 Issue Tracker: https://github.com/xiangxinai/xiangxin-guardrails/issues

---

Made with ❤️ by Xiangxin AI

---