LLM Fundamentals

Understand how per-token pricing works, why chat history quietly multiplies your bill, and how to budget AI features.

Understand the context window: the hard token limit every request lives inside, shared between input and output.

Learn why models under-attend to the middle of long contexts, why windows have limits at all, and where to place what matters.

Understand the difference between training a model and using one, and why prompting is your primary lever as a developer.

No new concepts: realistic scenario reps on tokens, cost, context windows, temperature, and inference.

Learn the anatomy of an effective prompt and how each component shapes model output.

Use delimiters to cleanly separate your instructions from user-supplied content, and see why that boundary also matters for security.

Use system prompts to set persona, constraints, and persistent behavior for a whole conversation.

Teach the model exactly what you want by including worked examples directly in your prompt.

Make models reason step by step on complex problems, and know when that helps versus when it just burns tokens.

Understand how AI represents meaning as vectors, and why that unlocks search by meaning instead of keywords.

Learn how cosine similarity turns 'how close in meaning?' into a number you can rank and threshold.

Understand why comparing against every vector breaks at scale, and how approximate nearest neighbor search fixes it.

Tour the vector database landscape, run vector search in Postgres with pgvector, and combine similarity with metadata filters.

Understand the three knowledge problems RAG solves and why it's the most important pattern in AI application development.

Walk the complete RAG machinery: the indexing phase that prepares documents and the retrieval phase that answers queries.

Learn how to split documents into chunks that maximize retrieval accuracy: size, overlap, and semantic boundaries.

Package retrieved chunks into a prompt that grounds the model, supports citations, and respects the context window.

Level up retrieval with hybrid search, cross-encoder re-ranking, and the metrics that prove it's working.

No new concepts: diagnose realistic RAG failures and make the engineering calls, end to end.

Understand what fine-tuning actually changes, and how it compares to prompting on cost, speed, and flexibility.

LoRA, managed fine-tuning services, and the dataset quality bar that decides whether fine-tuning helps or hurts.

A practical decision framework for prompting vs RAG vs fine-tuning, the classic mistakes, and how real systems combine all three.

Understand why LLMs fabricate plausible falsehoods with total confidence, and where that's most dangerous.

The five practical defenses: grounding, uncertainty prompts, verification passes, structured output, and human review.

Understand the attack class of the AI era: direct and indirect prompt injection, and why no perfect defense exists.

The defense-in-depth playbook: structural separation, least privilege, output validation, and human confirmation for agents.

Understand where prompt data goes, how to handle PII, and the regulations that govern AI data flows.

Fairness, transparency, accountability, and content moderation: the practices that make AI systems trustworthy.

No new concepts: realistic incident triage across hallucinations, injection, privacy, and responsible AI.

Walk through a complete AI Knowledge Assistant design, decision by decision, using every module of the course.

Design your own complete AI Knowledge Assistant: architecture, prompts, retrieval, and safety, applying every module of the course.