Multi stage critical thinking with reasoning models

This repository contains two Python scripts for interacting with OpenRouter's AI models, featuring retry mechanisms and different interaction patterns.

Prerequisites

• OpenRouter API key (set as environment variable OPENROUTER_API_KEY)

dual_model_critique.py

script that implements an iterative critique system using multiple AI models.

Features:

• Multi-turn conversation with alternating models
• Initial response followed by multiple critique iterations
• Maintains conversation history
• Alternates between different critique models
• Saves each iteration to JSON
• Enhanced error handling and retry mechanism

🎥 Watch How It's Built!

Watch the complete build process on Patreon
See exactly how this automation system was created step by step, with detailed explanations and insights into the development process.

❤️Join my AI community & Get 400+ AI Projects

This is one of 400+ fascinating projects in my collection! Support me on Patreon to get:

• 🎯 Access to 400+ AI projects (and growing daily!)
  • Including advanced projects like 2 Agent Real-time voice template with turn taking
• 📥 Full source code & detailed explanations
• 📚 1000x Cursor Course
• 🎓 Live coding sessions & AMAs
• 💬 1-on-1 consultations (higher tiers)
• 🎁 Exclusive discounts on AI tools & platforms (up to $180 value)

Usage:

export OPENROUTER_API_KEY="your-api-key"
python dual_model_critique.py

Configuration:

• MODEL_PRIMARY: Initial response model (default: "deepseek/deepseek-r1")
• MODEL_CRITIC_1: First critique model (default: "deepseek/deepseek-r1")
• MODEL_CRITIC_2: Second critique model (default: "deepseek/deepseek-r1")
• NUM_ITERATIONS: Number of critique iterations (default: 3)
• CRITIQUE_TYPE: Type of critique requested (default: "critical review")
• MAX_RETRIES: Number of retry attempts (default: 5)
• RETRY_DELAYS: Delay in seconds between retries (default: [1, 3, 5, 7, 10])

Process Flow:

1. Initial Response (Iteration 1):
   • Gets response from PRIMARY_MODEL for user prompt
2. First Critique (Iteration 2):
   • MODEL_CRITIC_1 reviews initial response
3. Second Critique (Iteration 3):
   • MODEL_CRITIC_2 reviews both initial response and first critique
4. Additional iterations alternate between critic models

Output Files

openrouter_reasoning.py

• Outputs to: model_responses.json
• Format:

  [
  {
    "timestamp": "ISO-8601-timestamp",
    "reasoning": "model's reasoning",
    "content": "model's response"
  }
  ]

dual_model_critique.py

• Outputs to: iterative_critique_responses.json
• Format:

  [
  {
    "iteration": 1,
    "timestamp": "ISO-8601-timestamp",
    "user_prompt": "original prompt",
    "conversation_history": [
      {
        "role": "user/assistant/critic",
        "content": "message content",
        "reasoning": "reasoning if available",
        "model": "model name",
        "critique_iteration": "iteration number (for critiques)"
      }
    ]
  }
  ]

Notes

• Add ":online" to model names to enable web search capability
• The Gemini model may have rate limits; adjust retry settings accordingly
• All API calls include error handling and retry mechanisms
• Both scripts use colored console output for better visibility
• JSON files are encoded in UTF-8

Error Handling

Both scripts implement robust error handling:

• Retries on rate limits (429 errors)
• Retries on resource exhaustion
• Handles API errors gracefully
• Provides detailed error messages
• Implements exponential backoff for retries

Environment Variables

Required:

• OPENROUTER_API_KEY: Your OpenRouter API key