Capstone: Forging a Custom AI Assistant
System Admin
AI Architecture // Syllabus Corp
LLMs alone are just text predictors. By giving them "Tools" (APIs) and "Memory" (Context), we transform them from passive chatbots into autonomous Agents capable of executing real-world tasks.
1. The Core: System Prompting
Every AI assistant begins with a System Prompt. Unlike user messages, the system prompt defines the unshakeable rules of engagement. If you are building an assistant for a hospital, the system prompt strictly forbids giving legal advice.
2. The Bridge: Function Calling
LLMs cannot natively browse the web or check databases. Function Calling bridges this gap. You provide the LLM with a JSON description of your APIs (e.g., check_inventory(sku)). When the user asks "Do we have shoes in stock?", the LLM realizes it needs external data and outputs a JSON command telling your code to run that function.
User: "What's the weather in Tokyo?"
LLM: { "tool_call": "get_weather", "args": { "location": "Tokyo" } }
App: *Fetches API and returns 22°C*
LLM: "It is currently 22°C in Tokyo."
3. The Engine: ReAct Agent Loop
For complex tasks, a single tool call isn't enough. The ReAct (Reasoning and Acting) framework allows the model to loop: it creates a Thought, takes an Action, views the Observation, and repeats until the goal is met.
❓ Architecture FAQs
What is the difference between RAG and Function Calling?
RAG (Retrieval-Augmented Generation): You automatically fetch relevant documents (via Vector DB) *before* asking the LLM the question, injecting it as context. Ideal for knowledge bases.
Function Calling: The LLM decides *dynamically* if it needs to trigger a function (like a math calculator or fetching a user profile) based on the conversation flow. Ideal for actions and live data.
How do AI Agents retain memory?
LLMs are stateless. To simulate memory, developers use a Conversation Buffer—an array of the entire conversation history (User, Assistant, User, Assistant) that is re-sent to the API with every single new prompt.