CHAPTER 1: CSE vs ML vs AI — Getting the
Foundations Right
1.1 What is Computer Science (CSE)?
Definition:
Computer Science is the core study of computation, algorithms, hardware-software
systems, and the theory behind programming. It’s the scientific and practical approach to
computation and its applications.
Key Areas in CSE:
● Data Structures & Algorithms: Efficient ways to store, access, and process data.
● Operating Systems: How computers manage hardware and software resources.
● Computer Networks: How data travels between systems over the internet or
intranet.
● Databases: How data is stored, retrieved, and managed efficiently.
● Software Engineering: Designing, developing, testing, and maintaining software.
● Computer Architecture: Understanding processors, memory, and low-level machine
interaction.
Real-World Example:
How does Instagram load so fast when you open it?
Behind the scenes:
● Databases fetch your profile and feed instantly.
● Caching systems (like Redis or Memcached) store recent data for fast access.
● Load balancing and distributed systems handle millions of users concurrently.
● Efficient code and APIs written in scalable programming languages.
Common Programming Languages in CSE:
● Python – Popular for scripting, backend, and ML.
● C++ – Systems programming and performance-focused applications.
● Java – Enterprise applications, Android development.
● JavaScript – Web development (frontend & backend with [Link]).
Top Courses to Learn CSE:
● CS50x by Harvard (Free, beginner-friendly, covers fundamentals to advanced
topics)
● MIT OpenCourseWare – Advanced and in-depth CS courses like Algorithms, OS,
Systems.
1.2 What is Machine Learning (ML)?
Definition:
Machine Learning is a subset of AI where machines learn from data to make decisions or
predictions without being explicitly programmed for each rule.
Core Idea:
Instead of writing rules manually (if X then Y), you provide data (input-output pairs), and the
model learns the mapping automatically.
Types of Machine Learning:
1. Supervised Learning
○ Labeled data is used.
○ Example: Email spam detection (emails labeled “spam” or “not spam”).
○ Algorithms: Linear Regression, Decision Trees, SVM, Neural Networks.
2. Unsupervised Learning
○ No labels, the model finds structure in data.
○ Example: Grouping customers based on behavior (clustering).
○ Algorithms: K-Means, PCA, DBSCAN.
3. Reinforcement Learning
○ Learning through rewards and penalties (trial and error).
○ Example: Teaching a robot to walk or an AI to play games like Chess or Atari.
○ Concepts: Agent, Environment, Reward Function.
Real-World Example:
You show a model 10,000 images of cats, labeled as “cat.”
It learns visual features (ears, shape, patterns) and then predicts whether a new image
contains a cat.
Tools and Libraries in ML:
● Scikit-learn
● TensorFlow
● PyTorch
● Keras
Top Courses to Learn ML:
● Andrew Ng’s ML Course (Coursera, Stanford-based, beginner to intermediate level)
● [Link] ML for Coders (Hands-on, deep learning focused, great for coders)
1.3 What is Artificial Intelligence (AI)?
Definition:
Artificial Intelligence refers to the broader field where machines simulate human-like
intelligence – not only learning patterns (ML), but also thinking, reasoning, planning, and
interacting.
AI is a superset of ML. While ML is mainly data-driven learning, AI combines ML with
logic, memory, decision-making, and even sensory processing (like vision, speech).
Key Subfields of AI:
● Machine Learning – Learning from data (covered above).
● Symbolic AI / Expert Systems – Rule-based systems used before ML boom.
● Natural Language Processing (NLP) – Understanding and generating human
language.
● Computer Vision – Interpreting visual information (images, video).
● Robotics – Intelligent movement and manipulation of physical objects.
● Planning & Reasoning – Decision-making processes, logical inference.
● Knowledge Representation – Storing knowledge in a way machines can reason
about.
♟️
Real-World Example:
A chess-playing AI doesn't just recognize past game patterns (ML), it also:
● Plans several moves ahead.
● Evaluates the game state.
● Remembers previous strategies.
● Adapts to new opponents.
Hence, it uses AI = ML + Planning + Reasoning + Memory + Sensory Input
AI Applications:
● Chatbots (like ChatGPT)
● Self-driving cars
● Fraud detection
● Language translation
● Virtual assistants (like Siri, Alexa)
CHAPTER 2: AI vs Generative AI — Machines That
Think vs Create
2.1 Traditional AI
What is Traditional AI?
Traditional AI refers to systems built to simulate intelligence through logic, rules, and
statistical models. These systems are not creative — they are designed to analyze,
classify, and make decisions based on data patterns.
Key Characteristics:
● Rules-based systems: If-else logic, expert systems.
● Decision trees & regression models: Predictive but not generative.
● Goal: Accuracy in prediction, not content generation.
What Traditional AI Does:
● Detect spam in emails
● Recommend products on Amazon
● Predict if a loan will default
● Flag fraudulent transactions
Real-World Examples:
● Fraud Detection System: Flags suspicious bank transactions based on historical
data.
● Spam Filter: Classifies emails as spam or not spam.
● Recommendation Engine: Suggests YouTube videos or Netflix shows based on
your past behavior.
Top Resources to Learn Traditional AI:
1. Artificial Intelligence by Georgia Tech (Udacity)
(Intermediate Level, focuses on classical AI: search, logic, planning, and
decision-making)
2. Stanford CS221: AI - Principles & Techniques
(University course — advanced, includes logic, search, planning, and probabilistic
reasoning)
3. Edureka’s YouTube Playlist – AI for Beginners
(Free visual walkthrough of traditional AI concepts like search algorithms, decision
trees, expert systems)
2.2 Generative AI (GenAI)
What is Generative AI?
Generative AI refers to AI systems that create new content, such as text, images, audio,
video, and code, by learning from massive datasets. These models go beyond
decision-making and are capable of creative tasks.
How Generative AI Works:
● Trained on large datasets (books, images, code, music).
● Learns patterns and probabilities of what comes next.
● Generates new outputs that resemble the training data, but aren’t direct copies.
Types of Generative AI:
Type Tools/Models
Text GPT-4, Claude, Gemini
Images DALL·E, Midjourney, Stable Diffusion
Audio ElevenLabs, Suno
Video Runway, Pika
Real-World Example:
Prompt: “Create a poem about time in Shakespearean style”
GPT-4 Output: A completely original, stylistically accurate Shakespearean poem about
time.
Top Resources to Learn Generative AI:
1. [Link]: Generative AI Specialization
(Beginner-friendly, explains text/image generation, prompt engineering, and
fine-tuning)
2. OpenAI Cookbook on GitHub
(Practical code examples on using GPT, embeddings, tools, function calling, etc.)
3. YouTube - Fireship: What is Generative AI?
(Full course on generative AI by freecodecamp explains GenAI with clear visuals
and real-world demos)
4. Hugging Face Course
(Build and deploy GenAI models like transformers, language models, and diffusion)
CHAPTER 3: AI Model vs AI Tool — Brains vs
Applications
3.1 What is an AI Model?
Definition:
An AI model is a mathematical or computational algorithm trained on data to make
decisions, predictions, or generate content. It represents the core logic or “brain” behind
intelligent behavior.
Think of it as:
● A trained system that learns from past data.
● Given an input, it produces a predicted or generated output.
Examples:
● GPT-4: A large language model (LLM) trained on massive internet text to generate
human-like responses.
● ResNet: A convolutional neural network for image recognition tasks.
● BERT: A transformer model used for understanding the context of words in text.
● Custom ML Model: A logistic regression model trained to predict customer churn.
Key Features of AI Models:
● Require training on large datasets.
● Can be fine-tuned for specific tasks (e.g., domain-specific chatbots).
● Can be hosted and accessed via APIs (e.g., OpenAI’s API).
Top Resources to Learn About AI Models:
1. Hugging Face – Transformers Course
Covers how transformer-based models like BERT, GPT, and T5 work. Great for text
models.
2. Coursera – Deep Learning Specialization by Andrew Ng
In-depth courses on neural networks, CNNs, RNNs, and model building.
3. YouTube – What is a Machine Learning Model (StatQuest)
Visual, beginner-friendly explanation on what ML models are and how they work.
3.2 What is an AI Tool?
Definition:
An AI Tool is a software product or application that uses one or more AI models under
the hood to provide value to end-users. It often includes:
● A user interface
● Backend APIs
● Storage, analytics, or additional features
Think of it as:
● The "productized" version of AI models.
● A tool that makes complex AI models usable for non-technical users.
Examples:
Tool Built On Purpose
ChatGPT GPT-4 Conversational AI
Midjourney Diffusion Models AI Image Generation
Jasper AI GPT-3.5 / GPT-4 AI Writing Assistant
[Link] LLMs (GPT) Marketing content generation
Notion AI GPT-based + APIs Productivity and writing aid
Components in an AI Tool:
● Frontend/UI: To input prompts and view results.
● Model Layer: Interacts with one or more AI models.
● Additional logic: Workflow handling, file uploads, analytics, etc.
Top Resources to Learn About AI Tools:
1. OpenAI Cookbook: Building with AI APIs
Real code examples to integrate GPT-3/4 into your apps.
2. Build AI-Powered Tools with LangChain (YouTube – Prompt Engineering Guide)
Learn how to build actual tools with memory, reasoning, and chaining APIs.
3. Google Vertex AI Tools
For those interested in building enterprise-grade tools with hosted models.
3.3 Key Differences Between AI Models & AI Tools
Feature AI Model AI Tool
Definition Trained algorithm that makes Product using one or more AI
predictions models
Purpose Core logic or intelligence engine End-user-facing app for
productivity, tasks
User Level Developers, researchers General public, creators, teams
Output Raw predictions (e.g., text, label, Refined interface or end result
score)
Examples GPT-4, BERT, DALL·E ChatGPT, Jasper, [Link], Runway
Training Yes – trained on large datasets No – tools are plug-and-play
Needed?
Access Method APIs, SDKs, inference libraries Web apps, mobile apps, SaaS
platforms
Analogy:
● AI Model = Engine
● AI Tool = Car (with steering, GPS, AC, dashboard using the engine)
Resources to Compare AI Tools vs Models:
1. Understanding AI and Chat GPT
Visual breakdown of tools like ChatGPT vs underlying models like GPT-4.
2. Medium Article: Everything about Ai Models
A beginner-friendly read on the conceptual understanding of models.
3. LangChain Docs
How developers combine multiple models, memory, and logic to build full tools.
CHAPTER 4: Large Language Models
(LLMs)
What is a Large Language Model (LLM)?
A Large Language Model (LLM) is a type of deep learning model, typically based on
transformer architecture, trained on massive datasets containing natural language and
sometimes other modalities (code, images, etc.). These models are capable of generating
human-like responses by learning the statistical patterns in text.
LLMs are pre-trained on large corpora (web data, books, codebases, forums, etc.) and
fine-tuned (or prompted) to perform downstream tasks such as:
● Text generation
● Summarization
● Translation
● Code generation
● Question answering
● Sentiment analysis
● Semantic search
● And more...
Key Examples of LLMs:
Model Developer Notable Features
GPT-4 OpenAI Multimodal (text + image), used in ChatGPT
Claude Anthropic Constitutional AI approach, long context window
Gemini Google DeepMind Formerly Bard; focuses on reasoning + grounding
LLaMA Meta Open-source weights, research & community
driven
Mistral [Link] Small, efficient, open-weight performant models
Falcon TII (UAE) Open-weight model optimized for performance
How Do LLMs Work?
Architecture: Transformer-based Neural Networks
Most LLMs rely on the Transformer architecture, first introduced by Vaswani et al. in the
2017 paper “Attention is All You Need.” The architecture uses self-attention mechanisms to
process input in parallel, enabling much greater scalability compared to RNNs or LSTMs.
Training Objective: Language Modeling
LLMs are trained using one of the following:
● Causal Language Modeling (CLM): Predict the next word/token given previous
ones. Used in GPT-style models.
● Masked Language Modeling (MLM): Predict masked tokens in a sequence. Used in
BERT-style models.
Data Scale:
● GPT-3 was trained on 500B+ tokens
● GPT-4's data is undisclosed but assumed to be trillions of tokens
● Models require massive compute (10K+ GPUs for months), high-quality data, and
careful alignment to reduce bias/toxicity.
What Can LLMs Actually Do?
Task Description
Text Generation Compose essays, blogs, scripts, documentation
Summarization Convert long articles or transcripts into bullet-point summaries
Translation Translate across multiple languages with high fluency
Programming Generate and autocomplete code (e.g., GitHub Copilot, Replit
Ghostwriter)
Semantic Search Return relevant documents based on meaning, not keywords
Reasoning & Logic Solve math, puzzles, or multi-step logic tasks
Conversational AI Used in chatbots, personal assistants, and therapy bots
Tool Use & Integrate with APIs to complete real-world tasks
Planning
Limitations of LLMs
Despite their intelligence, LLMs have several known drawbacks:
Limitation Description
Lack of true LLMs don’t "understand" like humans—they predict based on
understanding patterns
Statelessness Default LLMs don’t retain memory across sessions unless
engineered in
Hallucination May generate convincing but false or misleading information
Outdated Knowledge Limited to data it was trained on; doesn’t know recent or
real-time info
Context Window Limit Can only consider a fixed amount of text (~8K to 200K tokens
depending on model)
Real-World Applications of LLMs
Domain Application Example LLM’s Role
Email Gmail Smart Compose Predicts next phrase based on context
Software Dev GitHub Copilot Suggests code completions and refactors
Legal Ironclad, Spellbook Contract summarization and clause
extraction
Healthcare AI therapy apps like Simulated therapy chats
Woebot, Replika
Customer AI chatbots on websites Handles FAQs, product info, and simple
Support troubleshooting
Building Your Own LLM (Realistically)
You likely won’t train GPT-4 from scratch (costs ~$100M+), but you can build powerful
custom LLMs by fine-tuning or instruct-tuning open-source models.
Options for Building:
Method Description
Open-source models Start with LLaMA 3, Mistral, or Falcon (weights are public)
Hugging Face Most popular open-source NLP framework for model
Transformers training/inference
Fine-tuning Modify model behavior by training on custom data
LoRA / QLoRA Lightweight, low-cost fine-tuning methods using adapters
Creating LLM using Python
Next Step: AI Agents (LLM + Memory + Tools +
Autonomy)
Basic LLMs are reactive — they respond to prompts. AI agents are proactive — they
reason, plan, and act.
What are AI Agents?
An AI agent is a system built on top of an LLM that has access to tools (APIs, search, file
systems), memory, and logic modules to autonomously achieve goals.
Anatomy of an AI Agent:
● LLM core: Language reasoning & generation
● Planner: Determines sequence of actions
● Tools/API access: Executes external operations (e.g., search, booking)
● Memory: Maintains state, context across steps
● Feedback loop: Evaluates outcomes, refines next steps
Example:
Prompt: “Find 5 cheapest flights to Bali next month and book the best one.”
Component Function
LLM Interprets request and breaks it into steps
Tool Access Calls flight APIs, web scrapers
Planning Logic Compares prices, times, airlines
Execution Agent Books the selected flight through the right portal
Memory Remembers preferences and constraints
This is the foundation of Auto-GPT, AgentGPT, OpenAgents, and similar frameworks.
The Next Big Thing: RAG (Retrieval-Augmented
Generation)
Why Do We Need RAG?
LLMs like GPT-4:
● Don’t know your private/company data
● Can’t read PDFs or databases
● Forget context easily
● Hallucinate confidently
RAG solves these problems.
RAG = Retrieval-Augmented Generation
Concept: Feed the LLM with relevant, external knowledge at query time.
How RAG Works:
1. Document Ingestion: PDFs, Notion docs, policies, reports
2. Chunking: Split text into semantically meaningful segments
3. Embedding: Convert chunks into vectors using an embedding model
4. Vector DB Storage: Store vectors in a searchable vector database
5. Retrieval: When prompted, retrieve top-K similar chunks
6. Generation: LLM responds using retrieved context
Real-World Use Case:
Query: “What’s the refund policy in our 200-page HR PDF?”
● GPT-4 (alone): Doesn’t know
● RAG-enabled system: Retrieves “refund policy” chunk → GPT-4 gives a precise
answer
Tools & Stack for RAG
Component Tools / Libraries
Chunking LangChain, LlamaIndex, Haystack
Embedding OpenAI, HuggingFace, Cohere, BAAI
Vector DB Pinecone, Chroma, Weaviate, Qdrant, FAISS
LLM GPT-4, Claude, Mistral, LLaMA
Serving APIs FastAPI, LangServe, Flask
ChatGPT vs Full Custom RAG
Feature ChatGPT Upload Full RAG System
File Upload Yes Yes
Custom Chunking No Fully configurable
Vector Store Access No Full control
Multi-doc Scaling Limited Optimized for 1,000s of files
Persistent Knowledge No Yes (long-term memory)
Integration & Deployment Not customizable Fully deployable as chatbot/API/web app
Verdict: ChatGPT = Smart Assistant
RAG System = Enterprise-Grade Knowledge Engine
What is LangChain?
LangChain is an open-source Python framework that lets you build complex applications
using LLMs + memory + tools.
Modules in LangChain:
Block Type Function
LLM Block Sends prompt to GPT-4, Claude, Mistral
Tool Block Allows LLM to call calculator, search, APIs, SQL, etc.
Memory Block Stores user history and context
Chain Block Executes multi-step workflows
RAG Block Enables doc retrieval + response generation
LangFlow – No-Code LangChain Builder
LangFlow is a visual GUI for LangChain. Think of it as Canva for AI agents.
● Drag-and-drop interface
● Connect LLM → Tools → Memory → Output nodes
● Deploy as API or UI app
What is a Vector Database?
LLMs don’t understand keywords — they understand semantic meaning. A vector
database allows similarity search on semantic embeddings.
Example:
● “How do I get a refund?” = [0.21, 0.88, 0.09…]
● “Cancel and get my money back” → similar vector → matched!
Popular Vector DBs:
Tool Description
Pinecone Scalable, fully managed cloud-native DB
Weaviate Open-source, with modular plugins
ChromaDB Python-native, great for prototyping
Qdrant Fast, open-source with REST & gRPC support
How i build AI Teacher with Vector Database
Multi-Agent Control Panel (MCPs)
As AI agents evolve, you’ll need to coordinate multiple agents for complex workflows.
What is MCP?
MCP (Multi-Agent Control Panel) is a dashboard + orchestration system to manage:
● Specialized agents (planner, researcher, executor)
● Task assignment and monitoring
● Communication between agents
● Dynamic feedback and control
Top Resources to Learn About AI Agents:
1. LangChain + Agents Documentation
LangChain Agents
Build agents that think and act using memory, tools, and LLMs.
2. YouTube – How AutoGPT Works
AutoGPT Video
A short, explanation of AI agents that plan and do.
3. Blog: ReAct Prompting – Reason + Act
Original ReAct Paper
How to guide LLMs with structured thinking and actions.
CHAPTER 5: Retrieval-Augmented
Generation (RAG)
5.1 Why LLMs Alone Aren’t Enough
Limitations of Standalone Large Language Models (LLMs)
While LLMs like GPT-4, Claude, or Mistral are trained on vast datasets, they have inherent
limitations:
● Static Knowledge: LLMs possess knowledge only up to their training cut-off date
and cannot access real-time information.
● Hallucinations: They may generate plausible-sounding but incorrect or fabricated
information.
● Token Limits: There's a maximum limit to the amount of text they can process in a
single prompt (e.g., GPT-4's 128K tokens).
● No Personal Context: LLMs lack awareness of specific user data, such as personal
documents or proprietary company information.
Analogy
Consider asking a knowledgeable individual to analyze your company's latest sales report. If
they've never seen the report and only have outdated business knowledge, their insights
would be limited. Similarly, an LLM without access to current or specific data can't provide
fully informed responses.
5.2 What is Retrieval-Augmented Generation (RAG)?
Definition
Retrieval-Augmented Generation (RAG) is an AI framework that enhances LLMs by
integrating an information retrieval component. This allows the model to fetch relevant data
from external sources, such as documents or databases, and incorporate it into its
responses, thereby grounding its outputs in factual and up-to-date information. (Wikipedia)
Why RAG Works
● Enhanced Accuracy: By accessing external data, RAG reduces the chances of
hallucinations, ensuring responses are based on actual information.
● Real-Time Relevance: It allows AI systems to provide answers that reflect the most
current data available.
● Contextual Responses: RAG enables LLMs to tailor outputs based on specific user
contexts or proprietary information.
RAG Workflow: Step-by-Step Breakdown
1. Embedding: Documents are divided into chunks and transformed into vector
representations using embedding models (e.g., OpenAI's embeddings).
2. Storage: These vectors are stored in a vector database like Pinecone, FAISS, or
Weaviate.
3. Retrieval: Upon receiving a query, the system retrieves the most relevant document
chunks by comparing vector similarities.
4. Generation: The LLM processes the retrieved information alongside the original
query to generate a comprehensive and accurate response.
Analogy
Imagine an AI chef (LLM) preparing a dish (response). Without RAG, the chef relies solely
on memory. With RAG, the chef accesses fresh ingredients (retrieved data) from a
well-stocked pantry (external sources), resulting in a more flavorful and accurate dish.
5.3 Practical Use Case of RAG
Scenario: Summarizing an HR Policy Document
● Without RAG: The LLM might provide a generic overview of HR policies, potentially
missing specific details or including inaccuracies.
● With RAG: The system retrieves the actual HR policy document, extracts pertinent
sections, and generates a precise summary, ensuring accuracy and
relevance.(WIRED)
5.4 Tools and Technologies in the RAG Ecosystem
Core Components
● Embedding Models: Convert text into vector representations. Examples include
OpenAI Embeddings, Cohere, and Hugging Face models.
● Vector Databases: Store and facilitate efficient retrieval of vectorized data. Options
include Pinecone, FAISS, Weaviate, and ChromaDB.
● Retrieval Frameworks: Manage the retrieval process. Notable tools are LangChain,
LlamaIndex, and Haystack.
● LLMs (Generators): Generate human-like text based on inputs. Examples are
GPT-4, Claude, Mixtral, and Gemini.
● Orchestration Tools: Coordinate the various components of the RAG pipeline.
LangChain and RAGFlow are prominent examples.
● Frontend/UI: Provide user interfaces for interaction. Streamlit, Gradio, and custom
applications are commonly used.
Additional Enhancements
● PDF Parsing: Tools like PyMuPDF and LangChain’s PDFLoader facilitate the
ingestion of PDF documents.
● Web Scraping: Libraries such as BeautifulSoup and Scrapy enable real-time data
extraction from websites.
● API Integration: Incorporate live data (e.g., weather, stock prices) using RapidAPI or
SerpAPI.
● Fine-Tuning & Prompt Engineering: Enhance LLM performance through
customized prompts and memory chains.
5.5 ChatGPT and RAG Capabilities
ChatGPT Version RAG Capability
Default ChatGPT No retrieval; relies solely on pre-trained data.
ChatGPT with Can access and retrieve real-time web information.
Browsing
ChatGPT with File Able to retrieve and process information from uploaded
Uploads documents.
Custom GPTs Utilize specific instructions and contextual data provided by the
user.
API with Vector Store Full RAG implementation achievable using tools like LangChain
and Pinecone.
5.6 Resources to Master RAG
● [Link]: Retrieval-Augmented Generation Course: Offers foundational
knowledge and practical labs using OpenAI, Pinecone, and LangChain.
● LangChain RAG Quickstart: Provides a comprehensive tutorial for building
RAG-powered applications.
● LlamaIndex RAG YouTube Walkthrough: An accessible guide for ingesting data
and creating custom LLM interfaces.
● OpenAI Cookbook – RAG Notebooks: Features Jupyter notebooks demonstrating
real-world RAG implementations with GPT.
Got it! Here’s a concise Resources to Learn RAG section with brief descriptions and direct
links:
Resources to Learn Retrieval-Augmented Generation (RAG)
1. [Link] RAG Course
Beginner-friendly course by Andrew Ng explaining RAG fundamentals, vector
search, and building RAG applications.
[Link] RAG Course
2. LangChain Documentation
Official docs and tutorials for building RAG pipelines with vector stores and LLMs
using LangChain framework.
LangChain Docs
3. OpenAI Cookbook – RAG Examples
Code examples and Jupyter notebooks demonstrating RAG workflows with OpenAI
models and vector databases.
OpenAI Cookbook on GitHub
4. Pinecone Vector Database Docs
Documentation on how to store and query vector embeddings for efficient similarity
search in RAG systems.
Pinecone Docs
5. LlamaIndex Tutorials
Tools and tutorials to connect LLMs with external knowledge sources, simplifying
RAG implementation.
LlamaIndex GitHub
6. YouTube Tutorials
● LangChain RAG full tutorial
● LlamaIndex basics Everything you need to Know
● Learning RAG from Scratch
● Master RAG in 5 Hours
