QUICKSTART.md

🚀 Context Offloading - Quick Start

What You'll Learn

This crew demonstrates context offloading: storing information outside the LLM's context window using a persistent scratchpad.

3-Step Setup

1️⃣ Install Dependencies

cd context_offloading
crewai install

2️⃣ Configure API Keys

Create .env file:

GEMINI_API_KEY=your_gemini_key_here
TAVILY_API_KEY=your_tavily_key_here

Get keys:

• Gemini: https://aistudio.google.com/app/apikey
• Tavily: https://tavily.com/ (free tier)

3️⃣ Run the Crew

crewai run

What Happens

1. Planning Agent → Creates research plan → Saves to scratchpad
2. Research Agent → Reads plan → Searches web → Updates scratchpad
3. Synthesis Agent → Reads scratchpad → Creates report

Output

• report.md - Final comprehensive report
• knowledge/scratchpad.json - All intermediate notes

View the Scratchpad

cat knowledge/scratchpad.json | python -m json.tool

Customize

Change research topic: Edit src/context_offloading/main.py

inputs = {
    'topic': 'Your question here',
    'current_year': str(datetime.now().year)
}

Change preferences: Edit knowledge/user_preference.txt

Key Concept

Unlike context pruning (removes info), context offloading moves info externally for later retrieval - like taking notes!

Next Steps

• Read DOCUMENTATION.md for detailed explanation
• Explore knowledge/scratchpad.json to see agent notes
• Compare with context_pruning implementation

---

Questions? See CrewAI Docs