Overview
Retrieval-augmented generation (RAG) looks simple: find some text, paste it into a prompt, ask the model. Most of the work is in everything around that: getting documents in without losing their tables, finding the right piece among thousands, keeping old and private documents out, making the answer cite its source, and proving all of it with numbers. This course teaches each stage through one system you build piece by piece: Hollin Answers, an assistant over the documents of Hollin Home Appliances, a made-up appliance maker, for its support agents, repair partners and customers. Hollin is fictional; every lesson also opens with a real, cited case.
What you will learn
- 95 written parts, yours for good
- 38h 33m of reading, measured not estimated
- Written for the advanced level
- Every future revision included
- UPI, cards and netbanking
Prerequisites
- A Google account, to run code in Google Colab and to keep your hollin-answers repository in Google Drive between sessions. Nothing to install.
- One free API key, from Groq or Google Gemini, to use in Colab. Or no key at all, if you run a model on your own computer with Ollama and work in a code editor.
- Embedding models and rerankers run locally from Hugging Face and need no key.
- For the capstone and a few production lessons (PostgreSQL with pgvector, Docker Compose, CI), your own computer. Those lessons say so.
Curriculum
15 sections · 95 parts · 38h 33m3h 16m free to read · 35h 17m with the course
01Course guide5 parts · 70 min
- Retrieval-Augmented Generation: The Complete CoursePreview5 min
- How This Course WorksPreview6 min
- Setup: Colab, Your Model Provider, and Running LocallyPreview13 min
- The Project: Hollin AnswersPreview11 min
- Shared Hollin DataPreview35 min
02Module 1: RAG Foundations7 parts · 126 min
- Module 1: RAG FoundationsPreview1 min
- Lesson 1: Why a model needs your documents to answerPreview19 min
- Lesson 2: Nine stages, and where answers quietly go wrongPreview25 min
- Lesson 3: Choosing RAG, long context, fine-tuning or toolsPreview22 min
- Lesson 4: The course stack: every tool, one questionPreview19 min
- Lab: Start Hollin Answers with one cited answerPreview32 min
- Incident: Search answers that cited the wrong sourcePreview8 min
03Module 2: Document Ingestion and Parsing8 parts · 144 min
- Module 2: Document Ingestion and Parsing1 min
- Lesson 1: Where Hollin's documents come from, and how to read them23 min
- Lesson 2: Reading PDFs, office files and email without losing answers20 min
- Lesson 3: Tables, charts and footnotes that survive parsing20 min
- Lesson 4: Choosing a parser, then cleaning what it returns21 min
- Lesson 5: Duplicates, versions and the record every chunk inherits24 min
- Lab: Build Hollin's ingestion pipeline end to end26 min
- Incident: Black boxes that hid nothing from a parser9 min
04Module 3: Chunking7 parts · 119 min
- Module 3: Chunking1 min
- Lesson 1: Why chunking matters, and the simple splitters16 min
- Lesson 2: Structure, semantic and proposition chunking21 min
- Lesson 3: Chunk size, overlap and the metadata beside them25 min
- Lesson 4: Keeping the context a chunk loses when cut25 min
- Lab: Build Hollin's chunker and measure it22 min
- Incident: A splitter that silently dropped the rest9 min
05Module 4: Embeddings7 parts · 159 min
- Module 4: Embeddings1 min
- Lesson 1: Inside an embedding: vectors, distances and odd scores22 min
- Lesson 2: Choosing an embedding model: leaderboards, prefixes, limits28 min
- Lesson 3: Embedding variants: smaller vectors, token vectors, sparse terms28 min
- Lesson 4: Embedding operations: batch, cache, version, measure, swap31 min
- Lab: Hollin's embedding decision, measured and switchable35 min
- Incident: A model change the reset never saw14 min
06Module 5: Vector Stores and Indexing7 parts · 141 min
- Module 5: Vector Stores and Indexing1 min
- Lesson 1: How a vector index finds neighbours fast21 min
- Lesson 2: Choosing a vector store and keeping it swappable23 min
- Lesson 3: Payloads, tenants and filters that keep recall22 min
- Lesson 4: Keeping a vector index correct while it changes21 min
- Lab: A filtered, swappable store kept in sync43 min
- Incident: Hidden from people, still served to the model10 min
07Module 6: Search: Dense, Sparse, and Hybrid6 parts · 153 min
- Module 6: Search: Dense, Sparse, and Hybrid1 min
- Lesson 1: Keyword and dense search, measured on Hollin's codes31 min
- Lesson 2: Hybrid search: merging keyword and vector rankings31 min
- Lesson 3: Result sets: how many, whether any, which first33 min
- Lab: Hybrid search on the filtered store47 min
- Incident: Three search paths that skipped the check10 min
08Module 7: Query Understanding and Transformation6 parts · 173 min
- Module 7: Query Understanding and Transformation1 min
- Lesson 1: Rewriting what customers type into what documents say34 min
- Lesson 2: Asking the search several questions instead of one35 min
- Lesson 3: Routing questions, and what each extra step costs33 min
- Lab: Rewrite and route every incoming message63 min
- Incident: A cache that served one user's page to another7 min
09Module 8: Reranking and Post-Retrieval Processing5 parts · 133 min
- Module 8: Reranking and Post-Retrieval Processing1 min
- Lesson 1: Rerankers: reading the question and document together32 min
- Lesson 2: From ranked chunks to the model's context36 min
- Lab: Rerank and assemble the context55 min
- Incident: Summaries that rewrote the news9 min
10Module 9: Generation, Grounding, and Citations7 parts · 164 min
- Module 9: Generation, Grounding, and Citations1 min
- Lesson 1: Prompt design and the answer contract21 min
- Lesson 2: Citations you can check, streamed as they arrive23 min
- Lesson 3: Catching answers the documents do not support21 min
- Lesson 4: Keeping attackers out of the answer25 min
- Lab: The answer step, guarded and checked64 min
- Incident: One email, one image, no click9 min
11Module 10: RAG Architectures and Patterns7 parts · 225 min
- Module 10: RAG Architectures and Patterns1 min
- Lesson 1: Baseline RAG and the patterns that check themselves37 min
- Lesson 2: Agents for retrieval, and when not to use one37 min
- Lesson 3: Summaries, graphs and tables beside the text33 min
- Lesson 4: Conversations, page images, long contexts and live data27 min
- Lab: Four patterns behind one search82 min
- Incident: SQL rules written only in the prompt8 min
12Module 11: Evaluation8 parts · 236 min
- Module 11: Evaluation1 min
- Lesson 1: Measuring retrieval before you measure answers32 min
- Lesson 2: Scoring the answer, not only the search27 min
- Lesson 3: Building an evaluation set people actually ask34 min
- Lesson 4: Model judges and fair experiments34 min
- Lesson 5: Regression gates, user signals and safe online tests32 min
- Lab: The evaluation harness and its gate68 min
- Incident: A chatbot changed after it was tested8 min
13Module 12: Production Engineering8 parts · 222 min
- Module 12: Production Engineering1 min
- Lesson 1: Latency budgets, parallel stages and the slow tail24 min
- Lesson 2: What each answer costs, and how to spend less34 min
- Lesson 3: Caching answers and keeping them fresh27 min
- Lesson 4: Who may see what, and proving it afterwards25 min
- Lesson 5: Seeing every request, and failing honestly28 min
- Lab: Make Hollin Answers ready for production75 min
- Incident: Chat logs open to the internet8 min
14Module 13: Advanced and Emerging Topics5 parts · 144 min
- Module 13: Advanced and Emerging Topics1 min
- Lesson 1: Questions the corpus cannot answer, and what is missing32 min
- Lesson 2: Feedback loops and what the frontier changes35 min
- Lab: Corpus insights and the weekly log review68 min
- Incident: A chatbot that learned from its attackers8 min
15Module 14: Capstone: A Production RAG Service2 parts · 104 min
- Module 14: Capstone: A Production RAG Service2 min
- Lab: The production RAG service and the gauntlet102 min
Part 1 · freeRetrieval-Augmented Generation: The Complete Course · 5 min
How to make a language model answer from your documents, correctly, with citations, and say so when it cannot.
Retrieval-augmented generation (RAG) looks simple: find some text, paste it into a prompt, ask the model. Most of the work is in everything around that: getting documents in without losing their tables, finding the right piece among thousands, keeping old and private documents out, making the answer cite its source, and proving all of it with numbers. This course teaches each stage through one system you build piece by piece: Hollin Answers, an assistant over the documents of Hollin Home Appliances, a made-up appliance maker, for its support agents, repair partners and customers. Hollin is fictional; every lesson also opens with a real, cited case.
Who this course is for
| You are... | Start at |
|---|---|
| New to Python or to calling a language model | The AI Engineering course Module 0, Module 1 and Module 7 first, then Module 1 here |
| An engineer who has built a basic RAG demo | Module 1 as a quick check, then Module 2 |
| Already running RAG in production | Modules 6, 8, 9, 11 and 12, then the capstone |
| Mostly interested in quality | Module 1, Lesson 2, then Module 11 |
| Mostly interested in security and permissions | Module 2, Lesson 1; Module 5, Lesson 3; Module 9, Lesson 4; Module 12, Lesson 4 |
Reviews
No reviews yet. The first comes from a reader who finishes the course.
to review this course once you have finished it.
Next to this one
Transformers: From Attention to Trained, Served and Interpreted Models111 parts · 43h 12m · ₹399Large Language Models: The Complete Course103 parts · 34h 48m · ₹399Forward Deployed Engineer: The Complete Course85 parts · 18h 15m · ₹399AI Engineering: The Complete Course105 parts · 23h 44m · ₹399Agentic AI Engineer: The Complete Course97 parts · 33h 10m · ₹399
Large Language Models80 parts · 27h 0m · Free
Model Context Protocol83 parts · 14h 44m · ₹199
RAG82 parts · 13h 36m · ₹399
Python79 parts · 7h 28m · Free