LintAI - AI Output Testing & Validation Framework

A production-ready framework for validating AI/LLM outputs against user-defined assertions, confidence scoring, and edge case testing.

📦 Installation

From PyPI (Recommended)

pip install llm-validator

From Source

git clone https://github.com/SoulSniper-V2/lintai.git
cd lintai
pip install -e .

🎯 Features

• ✅ Assertion-Based Validation - Define expected behavior with simple rules
• 📊 Confidence Scoring - Get quantified trust metrics for outputs
• 🧪 Edge Case Testing - Systematically test boundary conditions
• 🤖 Multi-Model Support - Works with OpenAI, Anthropic, Gemini, local LLMs
• 📈 Regression Tracking - Track validation scores over time
• 🔄 CI/CD Integration - Run validations in GitHub Actions pipelines
• 🚀 Auto-Release to PyPI - Tags automatically publish to PyPI

🚀 Quick Start

CLI Usage

# Initialize a validation config
lintai init-config

# Validate with a config file
lintai validate --config validators/my_config.yaml

# Batch validation from JSONL
lintai batch --input test_cases.jsonl --output results.jsonl

from llm_validator import LLMValidator, Assertion, AssertionType

# Initialize validator
validator = LLMValidator(
    model="gpt-4",
    api_key="your-key"
)

# Define assertions
assertions = [
    Assertion(
        name="max_length",
        type=AssertionType.MAX_LENGTH,
        params={"max_tokens": 500},
        weight=0.3
    ),
    Assertion(
        name="no_profanity",
        type=AssertionType.NO_PATTERN,
        params={"pattern": r"(?i)badword|offensive"},
        weight=0.5
    ),
    Assertion(
        name="contains_action_plan",
        type=AssertionType.CONTAINS_TEXT,
        params={"text": "step 1", "count": 1},
        weight=0.2
    )
]

# Validate output
result = validator.validate(
    prompt="Create a plan to increase sales",
    output="Here is a step by step plan...",
    assertions=assertions
)

print(f"Confidence Score: {result.score}/100")
print(f"Passed: {result.passed}")
print(f"Failed: {result.failed_assertions}")

CLI Usage

# Run validation from config
llm-validate --config validators/sales_plan.yaml

# Quick test
llm-validate --prompt "Summarize this" --output "The text says..." --rules "max_tokens:100"

# Batch validation
llm-validate --input test_cases.jsonl --output results.jsonl

Web Dashboard

cd frontend
npm install
npm run dev

Access at http://localhost:5173

📁 Project Structure

llm-validator/
├── llm_validator/
│   ├── __init__.py
│   ├── core.py           # Main validation logic
│   ├── assertions.py     # Assertion types
│   ├── models.py         # Data models
│   └── providers.py      # LLM provider integration
├── frontend/
│   ├── src/
│   │   ├── App.jsx
│   │   └── components/
│   ├── package.json
│   └── vite.config.js
├── tests/
│   ├── test_core.py
│   └── test_assertions.py
├── validators/           # Example validation configs
├── README.md
└── requirements.txt

🛠️ Assertion Types

Type	Description	Example
MAX_LENGTH	Output within token/char limit	max_tokens: 1000
MIN_LENGTH	Output meets minimum length	min_words: 50
CONTAINS_TEXT	Output has required text	text: "step 1"
NO_PATTERN	Output doesn't match pattern	pattern: "error|fail"
REGEX_MATCH	Output matches regex	pattern: r"^\d+\."
SENTIMENT	Output sentiment check	min_positive: 0.6
JSON_VALID	Output is valid JSON	schema: ./schema.json
KEYWORD_COUNT	Keywords present	keywords: ["AI", "ML"]
CUSTOM	Python function validation	function: my_validator.py

📊 Confidence Scoring

The validator calculates a weighted confidence score:

Confidence Score = Σ(passed_weight) / Σ(total_weight) × 100

Individual assertion results:

• ✅ PASS: Assertion met
• ❌ FAIL: Assertion not met
• ⚠️ WARN: Assertion partially met (with penalty)

🎨 Example Validators

Code Review Validator

name: code_review
model: gpt-4
assertions:
  - name: has_tests
    type: CONTAINS_TEXT
    params: { text: "test" }
    weight: 0.3
  
  - name: no_hardcoded_secrets
    type: NO_PATTERN
    params: { pattern: "api_key|password|secret" }
    weight: 0.4
  
  - name: reasonable_length
    type: MAX_LENGTH
    params: { max_tokens: 2000 }
    weight: 0.2
  
  - name: has_error_handling
    type: REGEX_MATCH
    params: { pattern: "except|try|catch" }
    weight: 0.1

Customer Email Validator

name: customer_email
model: claude-3-opus
assertions:
  - name: professional_tone
    type: SENTIMENT
    params: { min_positive: 0.3, max_negative: 0.2 }
    weight: 0.3
  
  - name: has_greeting
    type: CONTAINS_TEXT
    params: { text: "Dear|Hello|Hi" }
    weight: 0.1
  
  - name: has_signature
    type: CONTAINS_TEXT
    params: { text: "Sincerely|Best|Thanks" }
    weight: 0.1
  
  - name: no_pii
    type: NO_PATTERN
    params: { pattern: "\\d{3}-\\d{2}-\\d{4}" }  # SSN pattern
    weight: 0.5

🔧 Configuration

Environment Variables

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=...
GOOGLE_API_KEY=...

Provider Selection

from llm_validator.providers import OpenAIProvider, AnthropicProvider, LocalProvider

# OpenAI
validator = LLMValidator(provider=OpenAIProvider(model="gpt-4"))

# Anthropic
validator = LLMValidator(provider=AnthropicProvider(model="claude-3-opus"))

# Local/Ollama
validator = LLMValidator(provider=LocalProvider(model="llama2"))

🧪 Testing

# Run all tests
pytest tests/

# Run with coverage
pytest --cov=llm_validator tests/

# Run specific test
pytest tests/test_core.py -v

📈 CI/CD Integration

GitHub Actions

name: Validate AI Outputs
on: [push]

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Set up Python
        uses: actions/setup-python@v5
        with: { python-version: '3.11' }
      - name: Install
        run: pip install llm-validator
      - name: Run Validation
        run: |
          llm-validate \
            --config validators/code_review.yaml \
            --output validation_results.json
      - name: Check Score
        run: |
          if [ $(jq '.score' validation_results.json) -lt 80 ]; then
            echo "Score below threshold!"
            exit 1
          fi

🎯 Use Cases

1. Production AI Safety: Validate outputs before showing to users
2. Code Review Automation: Check AI-generated code for quality
3. Content Moderation: Ensure outputs meet guidelines
4. Customer Support: Validate response quality
5. RAG Evaluation: Test retrieval-augmented generation accuracy
6. Model Comparison: Compare output quality across models

🤝 Contributing

1. Fork the repo
2. Create a feature branch
3. Add your assertion type
4. Submit a PR

🔄 Automated Releases

This project uses GitHub Actions for CI/CD:

Workflow	Description
Test	Runs pytest on every push/PR
Build	Builds PyPI package on every push
Publish	Auto-publishes to PyPI when a git tag is pushed

How to Release

# Make changes, commit
git add -A
git commit -m "Description of changes"

# Create a version tag (follows semver)
git tag v0.1.1

# Push tag to trigger PyPI release
git push origin main
git push origin v0.1.1

The CI workflow will:

1. Run tests
2. Build the package
3. Publish to PyPI automatically

Note: Requires PYPI_API_TOKEN secret in GitHub repo settings.

📄 License

MIT License - Build, validate, ship with confidence!

---

Never deploy AI without validation. 🛡️