100% Local Agents - DeepSeek R1 Agents with SmolAgents!

---

---

AI Summary

Summary of Video Transcript

Introduction

• The video by One Little Coder teaches how to build agents using DeepSpeed R1, a model from a Chinese company DeepSpeed.
• The tutorial focuses on creating two types of agents:
  1. Text to SQL agent
  2. A simple agent for basic understanding
• The agents will be built using the Small Agent framework from Hugging Face and powered by DeepSpeed R1’s distilled version of the Quen model.
• The tutorial is hands-on and requires attention to detail.

Prerequisites

• LM Studio must be installed and the DeepSpeed model should be up and running.
• The model should be loaded in LM Studio, indicated by an eject button.
• The developer tab in LM Studio should show the model as running and available at a specific endpoint.

Steps to Build Agents

1. Ensure LM Studio and the DeepSpeed R1 Quen model are set up.
2. Install Small Agents using pip install small_agents.
3. Import code_agent and openai_server_model from Small Agents.
4. Define the model with the model ID, API base, and API key from LM Studio.
5. Define the agent using code_agent without any tools initially.
6. Run the agent with a task, such as finding a number in the Fibonacci sequence.

Building a Text to SQL Agent

• Create a dummy SQL table with columns and insert values.
• Define a tool in Small Agents with detailed schema information.
• Use the same model setup as before but add the SQL engine tool.
• Ask the agent SQL-related questions and observe it generating and executing SQL queries.

Conclusion

• The tutorial demonstrates initiating the DeepSpeed R1 distilled model through LM Studio and serving it locally.
• It uses the Small Agent framework to create basic agents and a text to SQL engine.
• The process is entirely local, with no internet required after the initial setup.
• The video concludes with an invitation for suggestions for new use cases to develop and create tutorials.

Detailed Instructions and URLs

• No specific CLI commands, website URLs, or detailed instructions were provided in the transcript.