RAG and Agents in 8 Weeks

Theory

• Core LLM terminology: tokens, context window, parameters, attention mechanism
• How LLMs are trained: pre-training, post-training, and inference at runtime
• Embeddings & semantic similarity: turning text into vectors
• Prompting basics: system, user, few-shot

Coding

• Make your first LLM API calls, and compute cosine similarity from scratch

Theory

• Why RAG exists: knowledge cutoff and hallucination problem
• The 5-stage RAG pipeline: ingest → chunk → embed → index → retrieve
• Chunking strategies and when to use each
• Embedding models: choosing the right one for your use case

Coding

• Build an end-to-end RAG pipeline with LangChain + ChromaDB.

Theory

• Why naive RAG hits accuracy ceilings
• Query rewriting techniques: expansion, HyDE, multi-query
• Cross-encoder reranking: LLM-quality scoring after initial retrieval
• Metadata filtering and hybrid search (dense + sparse)

Coding

• Apply advanced retrieval techniques to your Week 2 pipeline and benchmark accuracy improvements with an eval suite

Theory

• Common pitfalls in production RAG systems
• GraphRAG: knowledge graphs as retrieval backends
• KAG (Knowledge-Augmented Generation): structured KB + LLM
• Agentic RAG: LLM decides what to retrieve, when, and how

Coding

• Build a GraphRAG system on a real use-case using Neo4j, and compare it against vanilla RAG

Theory

• What is an agent? LLM + tools + loop
• LLM vs. Agent: how to distinguish pipeline vs single api call
• Tool / function calling: how the LLM triggers Python functions
• The ReAct pattern: Reasoning → Acting → Observing

Coding

• Build a single agent with tools, structured outputs, and basic guardrails.

Theory

• Why multi-agent? Parallelism, specialisation, separation of concerns
• Agentic design patterns: Orchestrator-Worker, Routing
• Problems unique to multi-agent: orchestration, information isolation, planning
• Memory systems for agents: short-term and long-term

Coding

• Build a research + writer + critic multi-agent pipeline in LangGraph with routing.

Theory

• Context engineering vs. prompt engineering
• What goes in the context: instructions, tools, history, retrieved docs
• Quantitative Evals (accuracy, F1, exact match, tool-call accuracy)
• Qualitative evals with LLM-as-a-Judge: rubric design, single-answer grading.

Coding

• Run a full eval suite on your Week 6 multi-agent pipeline with both quantitative and LLM-as-a-Judge evaluations

Theory

• AI engineering best practices: from prototype to production
• Engineering decision framework: choosing RAG type, agent pattern, and stack
• How to scope a capstone: MVP definition, what to cut, what to keep
• Architecture reviews and 1:1 feedback on each project plan

Coding

• Live capstone demos with peer Q&A, eval walkthroughs, and structured feedback