Multi-Step Reasoning (MSR) Framework

A comprehensive agent-based framework that enhances the reasoning capabilities of foundation models through structured multi-step thinking processes and executable tools.

Overview

The MSR Framework is designed to improve problem-solving capabilities of large language models by implementing a structured, agent-based approach to complex tasks. By breaking down problems into explicit reasoning steps with goal-oriented planning, the framework enables more reliable, transparent, and effective solutions across a variety of domains.

Core Components

Agent System

The framework revolves around an agent architecture with the following key components:

• PlannerLLM: Generates structured research plans with explicit goals and dependencies between steps
• StepExecutor: Executes individual steps with tool integration and robust error handling
• Agent: Orchestrates the planning and execution processes with proper state management

Tool Integration

MSR supports multiple tool types to enhance agent capabilities:

• PYTHON_EXECUTION: Execute Python code for data analysis and computation
• TERMINAL_COMMAND: Run shell commands with security constraints
• WEB_SEARCH: Retrieve real-time information from the internet
• FILE_OPERATIONS: Read and write files securely
• DATA_ANALYSIS: Process and analyze structured data
• VISUALIZATION: Generate data visualizations and charts

REST API Server

A Flask-based REST API server provides programmatic access to the MSR framework:

• Create and manage agent sessions
• Generate structured research plans
• Execute steps with or without approval workflows
• Monitor execution progress and retrieve results
• Comprehensive logging and security features

Key Features

• Structured Planning: Generate detailed, multi-step plans with explicit goals for each step
• Dependency Tracking: Manage complex dependencies between steps for proper execution order
• Comprehensive Logging: Detailed logging system for tracking agent activities and debugging
• Tool Integration: Seamlessly integrate with various tools for enhanced capabilities
• REST API: Access all framework capabilities through a well-documented REST API
• Security Controls: Configurable security policies for tool execution and API access
• State Management: Robust state management for resuming interrupted executions

Use Cases

• Complex mathematical reasoning
• Multi-hop question answering
• Logical puzzles and deduction tasks
• Programming and algorithmic problem solving

Project Structure

• src/: Core implementation of the multi-step reasoning architecture
  • models/: Model configuration and adaptation for Qwen QWQ 32B
  • reasoning/: Implementation of reasoning steps and pipeline
  • utils/: Utility functions for prompts and other helpers
• examples/: Sample applications and demonstrations
• eval/: Evaluation frameworks and benchmarks
• tests/: Unit tests for components
• scripts/: Utility scripts for configuration and maintenance

Getting Started

Installation

# Clone the repository
git clone https://github.com/darwin-labs/MSR.git
cd MSR

# Install dependencies
pip install -r requirements.txt

# Install the package in development mode
pip install -e .

# Set up your API keys
cp .env.template .env
# Edit .env with your API keys

Configuration

The MSR framework uses environment variables for configuration, which can be set in a .env file:

# LLM API Configuration
OPENROUTER_API_KEY="your-api-key"
DEFAULT_MODEL="anthropic/claude-3-opus-20240229"

# Server Configuration
MSR_SERVER_HOST="0.0.0.0"
MSR_SERVER_PORT="5000"
MSR_API_KEY="your-secure-api-key"
MSR_API_KEY_REQUIRED="true"

# Logging Configuration
MSR_LOG_LEVEL="INFO"
MSR_LOG_FILE="logs/msr_server.log"

Running the Server

# Start the MSR REST API server
python -m src.server.server

# With custom configuration
python -m src.server.server --host 127.0.0.1 --port 8000 --log-level DEBUG --log-file logs/debug.log

Using the Python Client

from examples.server_client_example import MSRClient

# Create client instance
client = MSRClient("http://localhost:5000", api_key="your-api-key")

# Create an agent
agent_data = client.create_agent(
    task="Research the latest advancements in quantum computing",
    allowed_tools=["WEB_SEARCH", "PYTHON_EXECUTION"],
    model="anthropic/claude-3-opus",
    temperature=0.7
)

# Get agent ID
agent_id = agent_data["agent_id"]

# Generate a plan
plan = client.generate_plan(agent_id, max_steps=5)

# Execute the plan
results = client.execute_plan(agent_id)

# Print results
for step_result in results["results"]["execution_results"]:
    print(f"Step {step_result['step_id']}: {'Success' if step_result['success'] else 'Failed'}")
    print(f"Findings: {step_result['findings'][:100]}...")

Running the Agent Example

# Run the agent example with web search and data analysis
python examples/agent_example.py "Analyze the impact of AI on healthcare in 2023" \
    --enable-web-search --enable-data-analysis --max-steps 5 \
    --log-file logs/agent.log --log-level INFO

Documentation

• API Reference: Comprehensive documentation of the REST API
• Agent System: Detailed explanation of the agent architecture
• Tool Integration: Guide to available tools and their capabilities
• Configuration: Configuration options and environment variables

Contributing

Contributions are welcome! Please check out our contributing guidelines for details on how to get started.

License

This project is licensed under the MIT License - see the LICENSE file for details.