Aisert 🚀

Assert-style validation library for AI outputs - ensure your LLMs behave exactly as expected.

Installation

# Basic installation
pip install aisert

# With semantic validation (sentence-transformers)
pip install aisert[sentence-transformers]

# With HuggingFace support
pip install aisert[huggingface]

# All optional features
pip install aisert[all]

Quick Start

from aisert import Aisert, AisertConfig

# Simple validation (no dependencies required)
result = (
    Aisert("Paris is the capital of France.")
    .assert_contains(["Paris", "France"])
    .assert_not_contains(["spam", "inappropriate"])
    .collect()
)

print(f"Validation passed: {result.status}")

# Advanced validation with token counting and semantic similarity
config = AisertConfig(
    token_provider="openai",
    token_model="gpt-4",
    semantic_provider="sentence_transformers",
    semantic_model="all-MiniLM-L6-v2"
)

result = (
    Aisert("AI is transforming technology.", config)
    .assert_tokens(max_tokens=50)
    .assert_semantic_matches("artificial intelligence technology", threshold=0.7)
    .collect()
)

Features

• 🔗 Fluent Interface: Chain multiple validations with readable API
• 📝 Multiple Validators: Schema, content, token count, semantic similarity
• ⚡ Optional Dependencies: Install only what you need
• 🎯 Flexible Modes: Strict (exceptions) or non-strict (collect errors)
• 🌐 Multi-Provider: OpenAI, Anthropic, HuggingFace, Google
• 🔧 Extensible: Custom validators via base classes
• 🚀 Production Ready: Thread-safe with model caching

Use Cases

• 🛡️ Content Moderation: Filter inappropriate content in real-time
• ✅ API Response Validation: Ensure LLM outputs meet quality standards
• 🧪 Testing AI Systems: Automated testing for AI applications
• 📊 Quality Monitoring: Track AI model performance in production
• 🔄 CI/CD Integration: Validate AI-generated content in pipelines
• 📈 A/B Testing: Compare different AI model outputs

Validation Types

# Content validation (no dependencies)
Aisert(content).assert_contains(["required", "keywords"])
Aisert(content).assert_not_contains(["spam", "inappropriate"])

# Schema validation (Pydantic models)
Aisert(json_content).assert_schema(UserModel)

# Token counting (requires API keys)
Aisert(content, config).assert_tokens(max_tokens=100)

# Semantic similarity (requires sentence-transformers)
Aisert(content).assert_semantic_matches("expected meaning", threshold=0.8)

Configuration

# Simple configuration
config = AisertConfig(
    token_provider="openai",
    token_model="gpt-4",
    semantic_provider="sentence_transformers",
    semantic_model="all-MiniLM-L6-v2"
)

# Set API keys (for token counting)
export OPENAI_API_KEY="your-key"
export ANTHROPIC_API_KEY="your-key"

Error Handling

# Strict mode (default) - raises exceptions
try:
    Aisert(content).assert_contains(["required"])
except AisertError as e:
    print(f"Validation failed: {e}")

# Non-strict mode - collects all errors
result = (
    Aisert(content)
    .assert_contains(["term1"], strict=False)
    .assert_tokens(100, strict=False)
    .collect()
)

if not result.status:
    print("Some validations failed:", result.rules)

Documentation

• 📚 Examples - Configuration, usage patterns, production use cases
• 📖 API Reference - Complete API documentation
• 🔧 Custom Validators - Extend with your own validators

Requirements

• Python: >= 3.9
• Dependencies: Optional based on features used
• API Keys: Only for token counting (OpenAI, Anthropic, etc.)
• Memory: 100-500MB for semantic models (optional)

License

MIT License

Links

• GitHub - Source code and issues
• Documentation - Complete documentation
• PyPI - Package repository