Agents, RAG & Vibe Coding Notes
Subject Overview
Year 3 (Applied AI & Building Products) is where learning becomes building. This semester covers retrieval systems, agent frameworks and AI-native development workflows — the goal being to build sophisticated AI applications using retrieval, agents and AI-assisted development at every step. Vibe coding is treated as a discipline in its own right: directing AI assistants effectively, reviewing generated code critically, and knowing when not to trust the output.
Subject at a glance
8
Units
50
Topics
15
Hours
8
Infographics
Unit-wise Syllabus
8 units — click WhatsApp below to get the full notes for each
Unit 1: Retrieval-Augmented Generation (4 hours)
- Embeddings
- vector databases (Pinecone, Weaviate, Chroma, pgvector)
- chunking strategies
- hybrid search
- 1
Chunk documents
- 2
Embed chunks
Turn text into vectors
- 3
Store in a vector database
- 4
Embed the user's question
- 5
Retrieve the most similar chunks
- 6
LLM answers using those chunks
With citations
Study guide
Retrieval-Augmented Generation gives models access to your own data: embeddings, vector databases (Pinecone, Weaviate, Chroma, pgvector), chunking strategies and hybrid search. Chunking choices often matter more than the model you pick.
Unit 2: Advanced RAG (3 hours)
- Re-ranking
- query rewriting
- multi-hop retrieval
- evaluation (RAGAS)
- handling long context
Hybrid search
Keywords + vectors
Re-ranking
Re-score retrieved chunks
Query rewriting
Clarify vague questions
Multi-hop retrieval
Chain several lookups
Evaluation
RAGAS: faithfulness, relevance
Study guide
Advanced RAG improves quality with re-ranking, query rewriting, multi-hop retrieval, evaluation (RAGAS) and long-context handling. Measure retrieval quality separately from answer quality.
Unit 3: Agent Architectures (4 hours)
- ReAct
- tool use
- multi-agent systems
- planning
- memory
- frameworks — LangChain
- LlamaIndex
- CrewAI
- AutoGen
- 1. Thought: Reason about the next step
- 2. Action: Call a tool
- 3. Observation: Read the tool result
- 4. Repeat or answer: Until the task is done
Study guide
Agents let models plan and act: ReAct, tool use, multi-agent systems, planning and memory, with frameworks such as LangChain, LlamaIndex, CrewAI and AutoGen. Start with a single agent and a few reliable tools.
Unit 4: Anthropic Claude & MCP (3 hours)
- Model Context Protocol
- Claude Code as a coding agent
- building tools that AI assistants can call
- 1AI app (MCP client)
e.g. Claude Code
- 2Discovers tools
From an MCP server
- 3Model decides to call a tool
- 4MCP server runs it
Database, API, files
- 5Result returns to the model
Study guide
You study Anthropic's Claude and the Model Context Protocol: how MCP works, Claude Code as a coding agent and building tools that AI assistants can call. Build a small MCP server to expose your own data or tool.
Unit 5: Vibe Coding as Practice (3 hours)
- Directing Cursor/Claude Code effectively
- spec-driven prompts
- reviewing AI output
- when NOT to trust the AI
- managing context windows in coding agents
- 1
Write a clear spec
Goal, constraints, tests
- 2
Give the agent focused context
- 3
Generate in small steps
- 4
Review every diff
Don't trust blindly
- 5
Run tests and fix
- 6
Commit and repeat
Study guide
Vibe coding as practice means directing AI coding tools effectively: spec-driven prompts, reviewing AI output, knowing when not to trust the AI and managing context windows. Always read and test the code an AI writes.
Unit 6: Multimodal AI (3 hours)
- Vision-language models
- audio (Whisper, TTS)
- generation (Stable Diffusion, FLUX, video models)
Vision-language models
Understand images and text together
Speech to text
Whisper
Text to speech
Natural voices
Image generation
Stable Diffusion, FLUX
Video generation
Text-to-video models
Study guide
Multimodal AI covers vision-language models, audio (Whisper, TTS) and generation (Stable Diffusion, FLUX, video models). Combine two modalities, such as speech in and image out, in a small demo.
Unit 7: Tools introduced
- Cursor
- Windsurf
- Claude Code
- GitHub Copilot Workspace
- v0
- Bolt
- Lovable
- Replit Agent for rapid prototyping
- LangChain
- LlamaIndex
- CrewAI
- Pydantic AI
- vector DBs — Pinecone
- Weaviate
- Chroma
- pgvector
- Qdrant
AI code editors
Cursor, Windsurf, Claude Code
Rapid prototyping
v0, Bolt, Lovable, Replit Agent
Frameworks
LangChain, LlamaIndex, CrewAI, Pydantic AI
Vector databases
Pinecone, Weaviate, Chroma, pgvector, Qdrant
Study guide
The tools include Cursor, Windsurf, Claude Code, Copilot Workspace, v0, Bolt, Lovable, Replit Agent, LangChain, LlamaIndex, CrewAI, Pydantic AI and vector databases such as Pinecone, Weaviate, Chroma, pgvector and Qdrant. Pick two tools and learn them deeply instead of sampling everything.
Unit 8: Mini-projects
- RAG-powered Q&A system over a custom document corpus with evaluation
- multi-agent research assistant that browses
- summarizes and writes reports
- ship a full-stack AI app in 48 hours using v0 + Cursor + Claude as your team
- 1Hour 0–4
Problem, spec and wireframe
- 2Hour 4–16
UI with v0, backend with an AI editor
- 3Hour 16–32
Wire up the LLM, RAG or agent
- 4Hour 32–44
Test, fix, add evals
- 5Hour 44–48
Deploy and demo
Study guide
The mini-projects are a RAG question-answering system with evaluation, a multi-agent research assistant and a full-stack AI app shipped in 48 hours with v0, Cursor and Claude. Shipping fast under constraints is excellent practice for real startup work.
Why Agents, RAG & Vibe Coding Matters
RAG, agents and AI-assisted development define how AI products are built today. The skills here are exactly what AI startups and product teams hire for.
How to Prepare for Agents, RAG & Vibe Coding
- Build a RAG system on documents you know well so you can judge answer quality.
- Add evaluation early; it is what separates demos from products.
- Review every line of AI-generated code and write tests.
Agents, RAG & Vibe Coding: Frequently Asked Questions
What is RAG?
Retrieval-Augmented Generation retrieves relevant documents and gives them to a language model so its answer is grounded in your data.
What is the Model Context Protocol?
MCP is an open standard that lets AI assistants connect to external tools and data sources in a consistent way.