AI Agents Fundamentals for Beginners (Free Labs)

AI Fundamentals and LLM Basics [00:00]

• Key concepts like prompt engineering, context windows, tokens, embeddings, RAG, vector DB, agents, Langchain, Langraph, Claude, and Gemini are introduced as topics to be covered.
• The goal is to take an audience with zero prior knowledge to a comprehensive understanding of AI through a single project.
• Large Language Models (LLMs) are a subset of AI that answer user queries.
• Popular LLMs like OpenAI's GPT, Anthropic's Claude, and Google's Gemini are transformer models trained on massive datasets, potentially up to tens of trillions of tokens.

The size of training tokens can go up to tens of trillions of tokens that are used to train these models.

Context Windows and Token Limitations [01:41]

• Passing data to LLMs for specific queries, like internal company documents (e.g., 500 GB of TechCorp data), is necessary.
• Conversation history acts as short-term memory, known as the context window, which is measured in tokens.
• A token is roughly 3/4 of an English word.
• Context window sizes vary significantly, from 2,000-4,000 tokens for smaller models to 1 million tokens for models like Google's Gemini 2.5 Pro.
• Despite large context windows, LLMs have practical limitations in processing information within them, akin to memorizing a long string of digits.

While the context window plays an important role in storing them in memory, there are practical limitations in how LLM treats what's inside the context window.

Embeddings: Transforming Meaning into Numbers [04:56]

• The challenge of LLMs only being able to process a tiny fraction of large datasets (e.g., 500 GB) is highlighted.
• Embeddings are crucial for overcoming this by transforming text meaning into numerical vectors.
• Similar meanings result in mathematically close vector patterns, allowing for semantic similarity searches.
• An embedding model converts text into a vector, typically of 1536 numbers, representing its meaning.
• This enables systems to find relevant documents based on semantic intent, not just exact keywords (e.g., finding "dress code policy" even if "jeans" isn't mentioned).

Embeddings capture that semantic similarity.

Langchain: Simplifying AI Agent Development [06:43]

• Langchain is an abstraction layer designed to simplify the development of AI agents with minimal code.
• It addresses pain points like storing chat messages, maintaining context, connecting to knowledge bases, and provider switching by using pre-built components.
• An AI agent has autonomy, memory, and tools, differentiating it from a static LLM that only answers questions based on training data.
• Langchain offers standardized interfaces for LLM providers, memory management, vector database integration, embedding, and tool integration, reducing the need for extensive custom coding.

Langchain is an abstraction layer that helps you build AI agents with minimal code.

OpenAI API Calls and Response Handling [10:08]

• Labs are designed to take users from zero to making AI API calls.
• Environment verification includes checking Python installation, OpenAI library availability, and API key setup.
• OpenAI's conversational API, "chat completions," involves sending messages (system, user, assistant roles) and receiving responses.
• The response object contains fields like usage statistics and timestamps, with the content field holding the AI's textual response.
• Understanding tokens (prompt, completion, total) and their associated costs is important, as output tokens are more expensive.

The key takeaway is remembering how to navigate the response object with response.content.

Langchain in Practice: Multimodel Support and Prompt Templates [14:36]

• Langchain simplifies working with multiple AI providers (OpenAI, Gemini, Grok) through a single interface, reducing boilerplate code significantly.
• It supports comparing responses from different models using the same prompt structure, beneficial for A/B testing and cost balancing.
• Prompt templates with placeholders allow for dynamic variable filling, eliminating the need to maintain numerous similar prompt files.
• Output parsers transform free-form AI text responses into structured data like lists or JSON objects.
• Chain composition in Langchain allows linking prompts, models, and parsers into a single pipeline for cleaner, more scalable AI development.

With it, you can move from OpenAI to Google's Gemini or XAI's Gro by changing just a single word.

Prompt Engineering Techniques [18:09]

• Prompt engineering is crucial for influencing the quality of AI responses; specific prompts yield better results than vague ones.
• Zero-shot prompting asks AI to perform a task without examples, relying solely on its existing knowledge.
• One-shot and few-shot prompting provide one or multiple examples within the prompt to guide the AI's response format and style.
• Chain-of-thought prompting encourages the AI to show its reasoning step-by-step to solve complex problems, leading to more reliable outputs.
• Labs demonstrate that using structured prompting techniques can make prompts up to 10 times more effective.

The key takeaway is that the right technique can make your prompts 10 times more effective.

Vector Databases and Semantic Search [24:45]

• Vector databases store data by meaning (embeddings) rather than just values, enabling semantic search.
• This contrasts with traditional SQL databases where users must format search terms precisely.
• Popular vector databases include Pinecone and ChromaDB.
• Embeddings convert text into numerical vectors, allowing for similarity searches based on meaning, not exact wording.
• Key concepts in setting up vector databases include dimensionality (typically 1536 dimensions) and retrieval mechanisms like scoring (similarity threshold) and chunk overlap.

So essentially instead of searching by value, we can now search by meaning.

Building a Semantic Search Engine [31:26]

• The process of building a semantic search engine involves installing libraries like Sentence Transformers, Langchain, and ChromaDB.
• Embeddings are the backbone, converting text into numerical vectors where similar meanings are close.
• Document chunking is necessary for embedding large documents, with overlapping chunks preserving context and improving retrieval accuracy.
• Vector stores like ChromaDB efficiently store and search through embeddings, supporting metadata filtering.
• The full pipeline involves embedding user queries, searching the vector store, retrieving relevant chunks, and returning them, achieving a significantly higher success rate (e.g., 95%) compared to keyword search.

By the end, we built a production-ready search engine with 95% success rate.

Retrieval Augmented Generation (RAG) [35:14]

• RAG, or Retrieval Augmented Generation, allows AI assistants to fit large datasets (like 500 GB) into their context window and generate answers.
• The three steps are: Retrieval (converting queries and documents to embeddings and performing semantic search), Augmentation (injecting retrieved data into the prompt at runtime), and Generation (AI generating a response based on the augmented prompt).
• RAG enables AI to rely on up-to-date, private data without needing to fine-tune the LLM.
• The effectiveness of RAG depends on data characteristics, influencing chunking strategies (e.g., paragraph-based for legal docs, sentence-level for transcripts).
• A complete RAG system can answer questions with context, accuracy, and confidence, pointing back to source documents.

This is the same architecture that powers tools like Catchup, Claude, and Gemini.

Langraph for Complex Workflows [42:27]

• Langraph extends Langchain to handle complex, multi-step AI workflows beyond simple Q&A.
• It uses a graph-based approach with nodes (computational units/functions) and edges (defining execution flow).
• State graphs allow information to be shared and updated across the entire workflow.
• Langraph enables conditional branching, loops, and iterative analysis, making AI agents more sophisticated.
• It facilitates tool integration, allowing agents to dynamically choose and use external tools like calculators or web search engines.

With Langgraph, this becomes a graph where each node handles a specific responsibility.

Model Context Protocol (MCP) for Tool Integration [49:01]

• MCP provides a standardized way for AI agents to connect to external systems like databases, support systems, and third-party APIs.
• Unlike traditional APIs, MCP offers self-describing interfaces that AI agents can understand and use autonomously.
• The burden shifts from the developer to the AI agent in integrating tools.
• MCP servers expose tools and schemas, which Langraph integrates with, routing queries intelligently.
• This allows agents to extend their capabilities significantly by leveraging pre-built MCP servers from a community, similar to how USB allows devices to connect to computers.

MCP functions like an API, but with crucial differences that make it perfect for AI agents.

The Complete AI System [55:19]

• By integrating context windows, vector databases, Langchain, Langraph, MCP, and prompt engineering, TechCorp achieved complex document search in under 30 seconds with higher accuracy.
• A chat application UI enhances user satisfaction with conversation history tracking and better intuition, available 24/7.
• This represents a shift from static documents to living, intelligent systems that can proactively solve problems.
• The ultimate goal is to unlock the full value of business knowledge using AI agents.

The shift from static documents to living intelligent system marks a turning point not just for Tech Corp, but for how every other business can unlock a full value of its knowledge using a