Learning Path: A Sequenced Curriculum for RAG Mastery

A sequenced curriculum for mastering RAG — from fundamentals to research-frontier architectures.

Three Learning Tracks

Choose your track based on your available time and interview target.

Week-by-Week Study Plan (4-Week Depth-First)

Week 1: Foundations

Goal: Understand the building blocks all RAG systems rest on.

Code Exercise: Build a minimal retrieval pipeline: embed a 100-document corpus, store in FAISS, retrieve top-5 for 10 test queries, measure recall. Write down chunk size and embedding model choice and why.

Week 2: Core Architectures

Goal: Master Naive, Advanced, Modular, and Corrective RAG. Understand the retrieval problem deeply.

Code Exercise: Implement RRF hybrid retrieval (BM25 + dense). Benchmark hybrid vs. pure dense on 10 test queries. Document the recall gain.

Week 3: Advanced & Specialized Architectures

Goal: Understand when to use Agentic, Self-RAG, Graph, Structured, Multimodal, and Long-Context RAG.

Code Exercise: Write a retrieval strategy selector: given a query, decide whether to use dense, hybrid, graph-based, or SQL-based retrieval. Justify each choice.

Week 4: System Design & Security

Goal: Integrate knowledge into production-grade systems. Understand cost, observability, and security.

Code Exercise: Design a RAG system end-to-end: define the 5 properties, draw the pipeline, compute cost for 10K QPS, instrument for observability, mitigate prompt injection. Write as if you're explaining to a senior engineer.

Concept Dependency Graph

This DAG shows prerequisites for each concept. Follow the arrows: if you want to understand X, make sure you understand everything pointing to it first.

Embeddings  ──┐
             ├──► Vector Databases ──┐
Chunking ────┤                       ├──► Naive RAG ──┐
             └──────────────────────►│               │
                                     ├──► Advanced RAG ──┐
Retrieval Strategies ───────────────►│                   │
                                     └──► Modular RAG ───┤
Reranking ──────────────────────────►│                   │
                                     └──► Corrective RAG │
Evaluation Metrics ──────────────────┤                   │
                                     ├──► Agentic RAG ───┤
System Design Principles ────────────┤                   ├──► System Design Round
                                     ├──► Self-RAG ──────┤
Prompt Injection Risks ──────────────┤                   │
                                     ├──► Graph RAG ─────┤
                                     │                   │
                                     ├──► Structured RAG ┤
                                     │                   │
                                     └──► Long-Context RAG┘

Note: No circular dependencies exist. This is a strict partial ordering — you can start at the top and work downward.

Depth Calibration by Interview Type

Different interview types test different depth. Use this to calibrate your study.

Self-Assessment Questions

Answer these without notes. If you get stuck, find the answer in the files and re-read that section. These questions span multiple files — forcing cross-topic synthesis.

1. Embeddings + Chunking: Your system retrieves documents but misses relevant ones. How do you diagnose whether it's an embedding problem or a chunking problem?

2. Vector DBs + Retrieval: You're evaluating Qdrant vs. FAISS. What's the trade-off? Which would you choose for a 100M vector corpus and why?

3. Retrieval + Reranking: Your dense retrieval has 40% recall@5, but after reranking, it's 42%. Is reranking worth the latency cost?

4. Evaluation + Agentic: You measure NDCG@5 = 0.7 and faithfulness = 0.85. What does this tell you about your Agentic RAG system?

5. Chunking + Reranking: Parent-child chunking vs. fixed-size chunking with reranking — which scales better and why?

6. Embeddings + Graph RAG: Your system uses both text embeddings and a knowledge graph. How do you choose which retrieval source to use for a given query?

7. System Design + Cost: At 10K QPS with OpenAI embedding and GPT-4, what's your primary cost center? How do you optimize?

8. Adaptive RAG + Evaluation: You build an Adaptive RAG that routes queries. How do you measure whether the router improves over Naive RAG?

9. Reranking + Prompt Injection: A reranker is supposed to filter out irrelevant documents. Can a reranker be tricked by prompt injection attacks?

10. Retrieval Strategies + Evaluation: You're choosing between BM25 and dense retrieval for a medical question-answering system. How do you decide? What metrics inform the choice?

11. Long-Context + System Design: With Claude's 200K context window, should you use retrieval or long-context RAG? What's the trade-off?

12. Self-RAG + Observability: Your Self-RAG system fine-tunes a model to improve. How do you monitor whether the fine-tuned model is actually better in production?

13. Prompt Injection + Agentic RAG: An agent retrieves a document that contains prompt injection. What happens? How do you prevent it?

14. Structured RAG + Evaluation: You route some queries to SQL and some to text retrieval. How do you evaluate the system as a whole (one metric across both paths)?

15. Graph RAG + Chunking: In Graph RAG, how do you chunk documents when building the knowledge graph? Does chunking strategy affect graph quality?

Resources Beyond This Repo

These resources fill gaps this repo intentionally leaves. Use them for context.

Recommended Study Order

If you're time-constrained, follow this priority order:

1. Must-know (4 hours): 01_concepts/embeddings.md + 01_concepts/chunking_strategies.md
2. Essential (6 hours): 01_concepts/vector_databases.md + 01_concepts/retrieval_strategies.md
3. High-value (4 hours): 02_interview_bank/01-naive-rag.md Q1–Q5 + 02-advanced-rag.md Q1–Q5
4. Differentiator (4 hours): 00_overview/system_design_principles.md
5. Advanced (6 hours): 01_concepts/evaluation_metrics.md + 01_concepts/reranking.md
6. Everything else: Research-frontier (Agentic, Self-RAG, Graph) and specialized (Multimodal, Long-Context) — nice-to-have for senior rounds