Here are the key components:
The system message sets the behavior, tone, and role of the AI. It's like giving the model instructions before the conversation begins (e.g., "You are a helpful coding assistant"). This shapes how the model interprets and responds to user messages.
Example:
System: "You are a Python expert who explains code concepts using simple analogies. Always provide code examples and be encouraging to beginners."
User: "How do I use list comprehensions?"
Assistant: "Think of list comprehensions like a recipe shortcut! Instead of writing a full loop, you can create a new list in one line..."
Tool calls allow LLMs to interact with external systems and APIs. When a model determines it needs information it doesn't have (like current weather or database queries), it can call predefined functions and use their outputs to generate better responses.
Example:
User: "What's the weather in San Francisco?"
Model: *calls get_weather(location="San Francisco")*
→ Returns: {temp: 62°F, conditions: "Partly cloudy"}
Assistant: "The weather in San Francisco is currently 62°F and partly cloudy."
Reasoning refers to the model's ability to think through problems step-by-step before providing an answer. Advanced models can break down complex questions, consider multiple approaches, and show their "thought process" to arrive at more accurate conclusions.
Example:
User: "If a train leaves at 2 PM going 60 mph and another leaves at 3 PM going 80 mph, when will they meet?"
Model thinking: "Let me work through this step by step:
1. First train has a 1-hour head start = 60 miles ahead
2. Speed difference = 80 - 60 = 20 mph closing speed
3. Time to close gap = 60 miles ÷ 20 mph = 3 hours
4. They meet at 3 PM + 3 hours = 6 PM"
Assistant: "They will meet at 6 PM."
RAG combines the model's general knowledge with specific information retrieved from external sources. When you ask a question, the system first searches relevant documents or databases, then uses that context to generate a more informed, accurate response.
Example:
User: "What's our company's PTO policy?"
System: *searches company handbook*
→ Retrieves: "Employees receive 15 days PTO annually, accrued monthly..."
Assistant: "According to the company handbook, employees receive 15 days of PTO annually, which accrues monthly. You can find the full policy in the Employee Handbook section 4.2."