Module 3: Large Language Models & AI Tools

This module provides a practical guide to Large Language Model (LLM) technologies and how to use them effectively.

Session 1: LLM Fundamentals & Direct API Interaction This session lays the conceptual and technical groundwork for working with Large Language Models. We will cover the basic mechanics of LLMs (tokens, context windows), introduce essential prompt engineering techniques like Zero-shot and Few-shot prompting, and explain the core API parameters that control the model's output, such as temperature and top-k. The primary goal is to bypass frameworks initially and provide students with hands-on experience by executing direct, low-level API calls to the Gemini endpoint in Python, understanding exactly what is happening under the hood before adding complexity.

Session 2: Grounding, Structured Output, and RAG Concept This class addresses the limitations of bare LLMs, focusing on how to make their output reliable and predictable. We'll first tackle the problem of hallucination by demonstrating the use of the Google Search tool to ground responses in real-time information, observing the citation metadata. Crucially, we will leverage the students' advanced Python knowledge by showing them how to use Pydantic models to define JSON schemas, forcing the model to return structured, strongly-typed data. The session will conclude with a conceptual introduction to Retrieval-Augmented Generation (RAG) as the principle of grounding the model in private data.

Session 3: Frameworks and Retrieval Systems (LangChain & LlamaIndex) Moving from direct API calls, this session justifies and introduces high-level orchestration frameworks, specifically LangChain, for building scalable LLM applications. We will explore the framework's components (Prompts, Models, Parsers) and focus on implementing the Retrieval-Augmented Generation (RAG) pipeline. Students will learn the RAG components in depth: how to load external documents, use Text Splitters to break them into manageable chunks, and use Vector Stores and Retrievers to fetch the most relevant context, demonstrating how to build a basic question-answering system over a private dataset.

Session 4: Agentic Workflows and the ReACT Pattern The final session focuses on building a more sophisticated type of LLM application: an intelligent agent. We will distinguish between simple, fixed chains and dynamic agents that use the LLM as a reasoning engine to determine the next action. The core concept taught is the ReACT (Reasoning and Acting) pattern, where the agent decides whether to use a Tool (an external resource, like a calculator or a code function) or use the general LLM to answer. The ultimate goal is for students to learn how to define custom Python functions and wrap them as tools for the agent, enabling it to perform complex, multi-step tasks.