Beginner Guide to Agentic Workflow using CrewAI, Ollama, GPT-4o, Llama 3, and Gradio

---

---

AI Nuggets

YouTube Video Transcript Extraction Instructions

Social Media Engagement

• Like the video
• Leave comments
• Share with friends
• Subscribe to the channels: YouTube, LinkedIn, TikTok, and Facebook

Agentic AI Workflow Overview

• Concept from a paper on communicative agents in software development
• Multiple AI models with specific roles (CEO, programmer, tester, reviewer)
• Framework to orchestrate collaborative work between AI models

Importance of Agents

• Better performance with agentic frameworks
• Comparison of performance using human eval benchmarks
• Four agentic workflows mentioned by Andrew Ng on deeplearning.ai:
  1. Reflection
  2. Tool use
  3. Planning
  4. Multi-agent collaboration

Crew AI Web Page

• Provides an open-source library for implementing agentic workflows
• Documentation available for learning how to use it
• URL not provided in the transcript

Problem Statement

• Create social media posts for YouTube videos efficiently

Workflow Plan

1. Start with a transcript in text form
2. Extract topics automatically
3. Create a summary for each topic
4. Automate value proposition extraction
5. Generate social media posts

Coding with Crew AI

• Code is about 250 lines, took 4 hours to create
• Run locally using Lama 3 on a laptop
• Serve up Lama server for local use
• Access the application via a local URL (port 7860)
• Interface allows for file upload, model selection, and topic extraction
• Example text from a previous video on prompt engineering
• Use whisper model for transcription

CLI Commands

1. Serve up Lama server:

   lama-server start  

2. Run the code with Gradle:

   gradle run  

3. Access the interface via local URL (port 7860):

   http://localhost:7860  

4. Refresh the interface to load the file and start processing

Creating Agents and Tasks with Crew AI

• Define the role, goal, and backstory for each agent
• Specify tasks with detailed descriptions
• Create crews with multiple agents and tasks
• Use Gradle to define UI elements and handle button clicks

Exporting Results

• Click the Export button to save the generated social media post ideas

Running with Different Models

• Example using GPT-4:

  gradle run -Pport=7861  

• Access the new interface via port 7861:

  http://localhost:7861  

Additional Tips

• Ensure you have enough context length for transcript processing
• Keep API keys in a separate file for security
• Define agents and tasks clearly for effective execution
• Use asynchronous tasks if needed

URLs and Additional Information

• Crew AI webpage and documentation: URL not provided in the transcript
• Local URL for accessing the application interface: http://localhost:7860 and http://localhost:7861

Note

The complete URLs and some specific details are not provided in the transcript and would need to be obtained from the actual webpage or documentation.