RAG Learning Plan: A 4-Week Guide for Beginners
This plan is for learners who have never written code and want to understand RAG first. It follows 4 weeks, 5 sessions per week, and about 30 minutes per session, roughly 10 hours in total. A “day” means a learning session, so you can work at your own pace.
By the end, you should be able to explain the RAG workflow, manually simulate knowledge-base questions and answers, and identify where to look when an answer goes wrong. Building an automated RAG program can be a goal for the later coding stage.
This plan now includes 20 lessons. Click a study day below to open its concepts, examples, exercises, and expandable answers. Keep the practice pack for A–H sources, ten test questions, and record templates; begin with Lesson 1.
1. Understand What RAG Does
RAG (Retrieval-Augmented Generation) retrieves relevant information and then gives that information and the question to a language model to generate an answer. This usually does not require retraining the model. Think of it as an open-book exam: retrieval finds the relevant pages, and the model reads them and organizes an answer.Microsoft's explanation of the fundamentals
It helps to divide the workflow into two parts:
- Prepare the information: collect documents → organize the content → split it into suitable chunks → build an index that supports searching.
- Answer a question: receive the question → retrieve relevant chunks → put the question and chunks into a prompt → generate an answer with sources.
Vector search is a common approach, and keyword search can also support RAG. RAG helps ground answers in information, but it does not guarantee that every answer is correct.
2. A Routine for Each Session
- 10 minutes to understand: read about one concept and find an everyday analogy.
- 15 minutes to practice: complete a task using your notes or the sample material below.
- 5 minutes to reflect: write down “what I learned, an example, and what I still do not understand.”
Use the other two days each week for rest or review. If you cannot explain a weekly checkpoint yet, repeat the relevant exercises before moving on.
| Week | Focus | Outcome |
|---|---|---|
| Week 1 | Language models, prompts, context, and the basic RAG workflow | A RAG workflow diagram you can explain |
| Week 2 | Document chunks, embeddings, retrieval, and sources | A small knowledge base with chunk identifiers and sources |
| Week 3 | Manually simulate retrieval and generation | A reusable question-answering prompt and records of 5 experiments |
| Week 4 | Evaluate answers, locate problems, and improve the workflow | A 10-question test set and a one-page review |
3. Week 1: Build a Basic Understanding
| Session | What to understand | Exercise |
|---|---|---|
| Day 1 | What language models can do and why a fluent answer can still be wrong | List 3 common uses and identify which answers need verification |
| Day 2 | A prompt gives instructions; context is the information available for this response | Use the same short public text and ask for “a summary” and “an explanation in three points,” then compare |
| Day 3 | A model may not know a private document or a recent policy | List 3 questions that require external information |
| Day 4 | The roles of retrieval, adding context, and generation | Draw the workflow for a question about library opening hours |
| Day 5 | How RAG differs from retraining a model and how an agent can use RAG | Explain RAG to someone in 3 minutes without looking at your notes |
Checkpoint: explain “why retrieve information first,” “where that information goes,” and “why an answer can still be wrong after retrieval.” Use your own words.
4. Week 2: Understand How Information Is Found
| Session | What to understand | Exercise |
|---|---|---|
| Day 6 | Document quality affects retrieval; titles, dates, and sources help assess evidence | Select 3–5 short public documents or your own non-sensitive notes and record a title and date for each |
| Day 7 | Chunking splits documents; tiny chunks may lose context, while large chunks may include irrelevant content | Split a short article by paragraph and by topic, then compare which works better for questions |
| Day 8 | Embeddings represent text features numerically to help compare semantic similarity | Write 3 pairs of differently worded expressions with similar meanings, such as “when do you close” and “closing time” |
| Day 9 | The difference between keyword search and semantic search | Ask one question using the document's wording and another using synonyms, then manually find passages and compare |
| Day 10 | An index supports retrieval; Top-k selects the first k candidates; source information helps verify answers | Give chunks identifiers, record their sources, and select the 3 most relevant chunks for one question |
Checkpoint: produce at least 8 chunks, each with an identifier and source. Explain why “a highly ranked chunk may not answer the question” and why similarity alone is insufficient.
5. Week 3: Manually Simulate RAG
This week, you retrieve information manually and then ask an existing AI conversation tool to answer from the selected chunks. The exercise makes each step visible; later, a program can replace manual retrieval.
| Session | What to understand | Exercise |
|---|---|---|
| Day 11 | A clear question helps you select relevant information | Write 5 questions covering direct lookup, different wording, combined information, and missing information |
| Day 12 | Retrieval quality and answer quality should be examined separately | Before generating answers, manually select chunks for each question and record why you chose them |
| Day 13 | A prompt organizes the question, evidence, and response requirements | Use the template below to produce one answer with sources |
| Day 14 | Missing evidence calls for acknowledging uncertainty; conflicting documents need an explanation | Provide irrelevant chunks and then conflicting documents, checking whether the model invents facts or chooses without justification |
| Day 15 | Keeping experiment records makes comparisons possible | Complete 5 questions and record the question, selected chunks, model answer, and your assessment |
Sample Documents for Practice
The following material is entirely fictional and intended for learning. Each row can be treated as a chunk with a source identifier.
| Identifier | Source | Content |
|---|---|---|
| A | Library opening-hours notice, October 2026 edition | Monday–Friday: 9:00–21:00; Saturday–Sunday: 9:00–17:00 |
| B | Library borrowing rules, October 2026 edition | Each reader may borrow up to 3 books at a time; the loan period is 14 days per book |
| C | Renewal notice, October 2026 edition | Books not reserved by another reader may be renewed once before the due date for another 7 days |
| D | Temporary closure notice, October 12, 2026 | The library will be closed throughout October 18, 2026, for equipment maintenance |
| E | Library borrowing rules, archived 2025 edition | Each reader may borrow up to 2 books at a time; this edition has been replaced by the 2026 rules |
Try asking: “Can I visit at 18:00 on Sunday?” “How many books can I borrow at once?” “Is the library open on October 18, 2026?” “What is the daily overdue fine?”
A Prompt You Can Use
Copy the instructions below, then add your question and selected information:
Answer the question using the information below, and identify the supporting chunks after each key conclusion.
If the information is insufficient, explain what is missing rather than guessing.
If the documents conflict, explain the conflict. Choose an applicable version only when the documents explicitly state their effective dates or replacement relationship.
Question: enter your question.
Information: paste the selected chunks, keeping their identifiers, sources, and dates.
Checkpoint: observe an evidence-supported answer, a response to missing information, and an explanation of conflicting sources. Check that citations actually support the conclusions. The sample has no overdue-fine information, so it cannot establish an amount.
6. Week 4: Evaluate and Improve
| Session | What to understand | Exercise |
|---|---|---|
| Day 16 | Test questions should include easy, difficult, and unanswerable cases | Prepare 10 questions: 4 direct lookups, 2 paraphrases, 2 combined or conflicting cases, and 2 cases with missing information |
| Day 17 | Retrieval evaluation asks whether enough evidence was found | Identify the evidence needed for each question and check whether retrieval covers it |
| Day 18 | Answer evaluation considers correctness, support, citations, and handling of missing information | Assess each answer using the record below |
| Day 19 | Hybrid retrieval combines keywords and semantics; reranking sorts candidates again | Manually collect candidates with both approaches, merge them, and rank them again; change one factor at a time before comparing |
| Day 20 | Failed cases should guide improvements | Repeat the same 10 questions, compare results, and record the 3 most common problems and possible improvements |
A Simple Evaluation Record
Copy the table and add a row for each test question. Use “not applicable” when a question has no evidence that can support an answer.
| Question | Required evidence | Enough evidence found? | Answer correct? | Citations support the conclusion? | Missing or conflicting information handled appropriately? | Improvement |
|---|---|---|---|---|---|---|
| Example: how many books can I borrow at once? | B; if E is also retrieved, identify the replacement relationship | Yes / no | Yes / no | Yes / no / not applicable | Yes / no / not applicable | Record observations |
First locate the problem in a specific stage:
- The documents contain no answer: add information or explicitly state that information is insufficient.
- The documents contain an answer, but it was not found: inspect chunking, question wording, and retrieval methods.
- The answer was found, but the model responded incorrectly: inspect context completeness, conflicting material, and prompt clarity.
- The answer seems correct, but its citations do not support it: verify each conclusion against the original text rather than merely checking for citation markers.
Checkpoint: complete records for 10 questions and use at least one failed case to explain “what went wrong, what changed, and whether the result improved.” This small test set is for learning and does not represent every real-world use case.
7. A Glossary to Consult
| Term | Initial understanding |
|---|---|
| LLM | A large language model that interprets input and generates text |
| Prompt | Instructions and input content given to a model |
| Context | Information available to the model for its current response |
| Chunk | A passage split out of a document |
| Embedding | A numerical text representation used to compare semantics and other features |
| Index | A data structure built to find information efficiently |
| Top-k | The first k ranked candidate results; they are not necessarily relevant enough |
| Reranking | Sorting the candidates from initial retrieval again |
8. Resources and Next Steps
During Weeks 1–2, focus on Microsoft's RAG fundamentals course, reading about the fundamentals, data preparation, retrieval, and evaluation in that order. Consult the glossary for unfamiliar English terms.
During Weeks 3–4, optionally read LlamaIndex's introduction to RAG to understand loading, indexing, storing, querying, and evaluation. Save installation and code for the programming stage. The reading order and exercises are recommendations from this plan.
After four weeks, if you want to build an automated project, continue with:
- Python variables, lists, dictionaries, loops, and functions.
- Reading text files and understanding API requests and JSON.
- A minimal workflow using simple search: find chunks → pass them to a model → answer.
- Vector retrieval, evaluated using the same 10-question test set.
- Frameworks, reranking, and more complex retrieval strategies as actual problems require them.
9. Progress
- [ ] Week 1: I can explain the basic RAG workflow.
- [ ] Week 2: I have organized chunks with identifiers and sources.
- [ ] Week 3: I have completed 5 manually simulated questions and answers.
- [ ] Week 4: I have completed a 10-question test and a one-page review.
Start your first session with Day 1: list 3 uses for a language model and explain how you would verify the result of each.