Welcome to the official wiki for Murmur, a sophisticated multi-agent system that orchestrates different specialized AI agents using local LLM models. This wiki serves as the central knowledge base for understanding, using, and contributing to the project.

• Overview
• System Architecture
• Getting Started
• Agent Details
• Configuration Guide
• Development Guidelines
• Troubleshooting
• FAQ

Murmur is an intelligent orchestration system that integrates multiple Large Language Models (LLMs) built on the Ollama architecture. The system implements a sophisticated pipeline of specialized agents to process and respond to user queries with high accuracy and reliability.

• 🤖 Multi-agent architecture with specialized roles
• 🔄 Asynchronous processing pipeline
• 💾 Local LLM integration via REST API
• 🧠 Conversation memory management
• 📊 Confidence-based response evaluation
• ⚠️ Comprehensive error handling

• Version: 1.1.0
• Stability: Beta
• Python Support: 3.7+
• License: MIT

User Input → Interpreter → Reasoner → Generator → Critic → Final Response

Each agent in the pipeline serves a specific purpose:

1. Interpreter Agent

   • Model: mistral-nemo:latest
   • Purpose: Analyzes user intent and context
   • Key functions:
     • Intent recognition
     • Context analysis
     • Requirement identification

2. Reasoner Agent

   • Model: llama3.2-vision:11b
   • Purpose: Develops logical approach
   • Key functions:
     • Problem decomposition
     • Solution strategy development
     • Consideration analysis

3. Generator Agent

   • Model: gemma2:9b
   • Purpose: Creates initial responses
   • Key functions:
     • Content generation
     • Response structuring
     • Context incorporation

4. Critic Agent

   • Model: llama3.2-vision:11b
   • Purpose: Reviews and refines content
   • Key functions:
     • Accuracy verification
     • Clarity assessment
     • Content refinement

• LocalLLMConnector: Manages communication with local LLM server
• AgentOrchestrator: Coordinates agent pipeline execution
• Message System: Handles inter-agent communication
• Confidence Scoring: Evaluates response quality

• Python 3.7+
• Local LLM server (Ollama)
• Required packages: aiohttp, asyncio

1. Clone the repository:

git clone https://github.com/KazKozDev/murmur.git
cd murmur

2. Install dependencies:

pip install -r requirements.txt

3. Configure local LLM server:

# Default server URL: http://localhost:11434
# Modify in config if needed

# Start the application
python src/main.py

# Enter queries when prompted
Enter your message: What is the capital of France?

# Review response and confidence score
Response: The capital of France is Paris.
Confidence: 0.95

Models can be configured in the AgentOrchestrator initialization:

self.agents = {
    AgentRole.INTERPRETER: BaseAgent("Interpreter", "mistral-nemo:latest", AgentRole.INTERPRETER),
    AgentRole.REASONER: BaseAgent("Reasoner", "llama3.2-vision:11b", AgentRole.REASONER),
    AgentRole.GENERATOR: BaseAgent("Generator", "gemma2:9b", AgentRole.GENERATOR),
    AgentRole.CRITIC: BaseAgent("Critic", "llama3.2-vision:11b", AgentRole.CRITIC)
}

Default server settings in LocalLLMConnector:

base_url = "http://localhost:11434"
timeout = 30.0
max_retries = 3

• Follow PEP 8 guidelines
• Use Black formatter
• Sort imports with isort
• Include type hints
• Write comprehensive docstrings

• Write unit tests for new features
• Ensure test coverage > 80%
• Run tests before submitting PR:

pytest tests/

1. Fork the repository
2. Create feature branch
3. Implement changes
4. Add tests
5. Submit PR

1. Connection Errors

   • Verify LLM server is running
   • Check server URL configuration
   • Ensure network connectivity

2. High Resource Usage

   • Reduce conversation memory size
   • Adjust model configurations
   • Monitor system resources

3. Low Confidence Scores

   • Check input quality
   • Verify model availability
   • Review agent configurations

Enable debug logging:

logging.basicConfig(level=logging.DEBUG)

Q: Can I use different LLM models? A: Yes, any model compatible with the Ollama API can be configured.

Q: How is conversation context maintained? A: Through a deque-based memory system with configurable size.

Q: What's the recommended hardware? A: Minimum 16GB RAM, modern multi-core CPU, and SSD storage.

For more information, visit our GitHub repository or contact the maintainers.