Setting Up a Crew AI Team using Lightning AI

Introduction

• Tutorial on optimal setup for Crew AI team
• Techniques from Crew AI founder
• Utilizing Lightning AI – a cloud-based code editor
  • Collaborate on code in the cloud
  • Power open-source models
• Goal:
  • Build a Crew AI team
  • Replace GPT-4 with Mixol or Mistol

Initial Setup

• Sign up for a Lightning account
• Create a new studio: Code Studio
• Entire environment is cloud-based
  • Avoids issues with Python environment management

Structure the Crew AI Code Framework

• Source folder organization
  • Right-click: New Folder -> source
  • Create subfolders for specific AI teams, e.g., financial analyst crew
  • Sub-subfolder: config
• Familiar environment: Similar to VS Code

Define Agents and Tasks

• Use YAML files
  • Create files agents.yaml and tasks.yaml
• Example Tasks:
  1. Research Company Task
     • Description: Use search tools to look up company's stock info
     • Expected Output: Relevant financial information
  2. Analyze Company Task
     • Description: Analyze financial info (profitability, liquidity, solvency ratios)
     • Expected Output: Thorough financial analysis

Define Agents

• Agents describe roles and goals based on tasks
  • Example: company researcher, financial researcher
  • Set delegation and verbosity options

Main File Setup

• Create main.py file
  • Import necessary libraries from Crew AI and Langchain for Grok
  • Define agents and tasks using decorators @agent and @task
  • Create and configure the Crew

Environment Configuration

• Automatic Saves: Lightning AI benefits
• Tools setup using existing libraries e.g., SEC_tools.py from Crew AI examples

Running the Project

• Use Poetry for dependency management
  • Commands: pip install poetry, poetry lock, poetry install
  • Run with: poetry run financial_analyst_crew

Powering with Open-Source Model (Mixol)

• Use Lightning AI GPUs
• Access Studio Templates from Lightning
• Use API Builder to expose endpoint
  • Incorporate Mixol model via Langchain llms.Olama
  • Change model definition in main.py to use Olama Mixol
• Performance Check: Validate by monitoring GPU activity

Conclusion

• Successful setup of Crew AI team
• Switching from GPT-4 to open-source models with Lightning AI GPUs
• Exposing API endpoints and integrating into projects

Final Notes

• Credits: Lightning AI sponsoring the tutorial
• Encouragement to like and subscribe