Analyzers Overview

Vow uses a multi-analyzer architecture to detect different types of issues in AI-generated content. Each analyzer specializes in a particular domain and contributes to the overall trust score.

Available Analyzers

How Analyzers Work

1. Content Detection

First, Vow identifies sections that are likely AI-generated using:

• Entropy analysis (detecting unnatural patterns)
• Style consistency checks
• Common AI output markers
• User annotations (<!-- AI-generated -->)

2. Analyzer Pipeline

Each enabled analyzer processes the content independently:

graph LR
    A[Input] --> B[Code Analyzer]
    A --> C[Text Analyzer] 
    A --> D[Security Analyzer]
    B --> E[Rule Engine]
    C --> E
    D --> E
    E --> F[Trust Score]
    F --> G[Output]

3. Issue Detection

Analyzers look for specific issue patterns:

# Example issue types
code_issues:
  - hallucinated_imports    # Non-existent packages
  - invalid_apis           # Fabricated endpoints  
  - syntax_errors          # Malformed code
  - deprecated_usage       # Outdated APIs

text_issues:
  - factual_errors         # Contradicts known facts
  - broken_references      # Invalid URLs/citations
  - inconsistencies        # Self-contradictory content
  - style_anomalies        # Unnatural writing patterns

security_issues:
  - hardcoded_secrets      # API keys, passwords
  - injection_patterns     # SQL injection, XSS
  - dangerous_functions    # eval(), exec(), etc.
  - privilege_escalation   # Unsafe permissions

Analyzer Selection

Automatic Selection

By default, Vow automatically selects analyzers based on file type:

# Python files → Code + Security analyzers
vow check script.py

# Markdown files → Text + Security analyzers  
vow check document.md

# Mixed directory → All analyzers
vow check ./project

Manual Selection

Override automatic selection with --analyzers:

# Use only code analyzer
vow check file.py --analyzers code

# Use multiple specific analyzers
vow check . --analyzers code,security

# Exclude text analyzer
vow check . --exclude-analyzers text

Configuration File

Set default analyzers in .vow.yaml:

analyzers:
  # Enabled analyzers (default: auto)
  enabled:
    - code
    - text
    - security
  
  # Auto-selection rules
  auto_select:
    "*.py": [code, security]
    "*.js": [code, security]
    "*.md": [text]
    "*.rst": [text]
    "*": [code, text, security]

Performance Considerations

Model Loading

• Models are loaded once and cached in memory
• First run may be slower (~2-3 seconds)
• Subsequent runs are faster (~100-500ms per file)

Resource Usage

Optimization Tips

# Skip model loading for syntax-only checks
vow check file.py --no-ml-models

# Use faster, less accurate models
vow check . --model-size small

# Parallel processing for large directories
vow check . --jobs 4

# Cache results for unchanged files
vow check . --cache

Trust Score Calculation

Each analyzer contributes to the overall trust score:

Trust Score = weighted_average(analyzer_scores)

Where:
- Code Analyzer Weight: 40%
- Text Analyzer Weight: 35%
- Security Analyzer Weight: 25%

Score Interpretation

• 0.8-1.0: High confidence (green)
• 0.6-0.8: Medium confidence (yellow)
• 0.3-0.6: Low confidence (orange)
• 0.0-0.3: Very low confidence (red)

Custom Analyzers

Vow supports custom analyzers via WASM plugins:

# Install a custom analyzer
vow analyzers install my-analyzer.wasm

# List installed analyzers
vow analyzers list

# Use custom analyzer
vow check file.py --analyzers my-analyzer

See Writing Custom Analyzers for details.

Debugging Analyzers

Verbose Output

# Show analyzer decisions
vow check file.py --verbose

# Debug specific analyzer
vow check file.py --debug-analyzer code

Analyzer Logs

# Show performance metrics
vow check . --stats

# Export detailed logs
vow check . --log-file vow.log --log-level debug

Validation Mode

# Test analyzers without ML models
vow check file.py --dry-run

# Validate analyzer configuration
vow analyzers validate

Next Steps

• Code Analyzer - Detailed guide to code analysis
• Text Analyzer - Understanding text verification
• Security Analyzer - Security-focused checking
• Configuration - Customize analyzer behavior