Agent From Scratch

A lightweight LLM agent implementation using Google's Gemini model, built without intermediate libraries. This project demonstrates how to create an AI agent that can use tools and respond to queries directly.

How It Works

1. User input analysis
   • Agent receives user input
   • Analyzes query using Gemini model
   • Determines if tools are needed

2. Decision making and response output
   • If query matches tool capability:
     • Selects appropriate tool
     • Formats tool inputs
     • Executes tool function
   • If no tool needed:
     • Generates direct response using LLM

Configuring environment

# Clone repository
git clone https://github.com/whanyu1212/agent-from-scratch.git
cd agent-from-scratch

# Syncing the dependencies listed
uv sync

# Activate the environment
source venv/bin/activate  # Linux/Mac

# or
.\venv\Scripts\activate  # Windows

# Set up environment variables
cp .env.example .env
# Add your GEMINI_API_KEY to .env

Usage

Please refer to example notebook for a simple demonstration.