Reflection Agents

Self-improving AI that critiques and refines its own responses.

Overview

Reflection agents implement a Generate → Critique → Regenerate loop where the model evaluates and improves its own outputs.

Key benefits:

  • Higher quality responses
  • Self-correction of mistakes
  • Minimal human intervention
  • Configurable quality thresholds

How It Works

  1. Generate: Model produces initial response
  2. Critique: Reflection agent scores quality (0.0-1.0) and identifies weaknesses
  3. Regenerate: Model creates improved version based on critique
  4. Repeat: Continues until quality threshold met or max iterations reached

Configuration

{
  reflection: {
    enabled: true,
    max_iterations: 3,
    quality_threshold: 0.7,
    early_stopping: true,
    min_improvement_delta: 0.05
  }
}

Usage

Enable reflection mode in your request:

curl -X POST http://localhost:8080/v1/chat/completions \
  -d '{
    "model": "phi3",
    "mode": "reflection",
    "messages": [{"role": "user", "content": "Write a professional email"}]
  }'

Example Output

The response includes reflection metadata:

{
  "choices": [{
    "message": {
      "content": "Final improved response..."
    }
  }],
  "reflection_metadata": {
    "total_iterations": 2,
    "final_score": 0.85,
    "improvement": 0.25,
    "threshold_met": true
  }
}

Use Cases

  • Content generation: Articles, emails, documentation
  • Code review: Self-critique code before returning
  • Complex reasoning: Multi-step problems with verification
  • Quality assurance: Ensure responses meet standards

Performance

  • Additional latency: 2-3x base latency (due to iterations)
  • Token usage: ~2-4x tokens consumed
  • Quality improvement: Typical 20-30% score increase