RAG Learning Academy
April 3, 2026 · View on GitHub
A structured, multi-agent learning environment for mastering Retrieval-Augmented Generation (RAG) — powered by Claude Code.
Inspired by Claude Code Game Studios, reimagined as an interactive learning guide for RAG systems.
What Is This?
RAG Learning Academy transforms Claude Code into a personal RAG tutor with 20 specialized AI agents, 22 interactive slash commands, and a 9-module curriculum that takes you from "what is RAG?" to production deployment.
Every concept is paired with a hands-on exercise. Every exercise is paired with evaluation metrics. You learn by building.
Quick Start
git clone https://github.com/TakaGoto/rag-learning-academy.git
cd rag-learning-academy
claude # Open Claude Code
/start # Begin your learning journey
That's it. No API keys, no pip install, no setup. Dependencies get installed when you need them (the first hands-on lesson will guide you).
What's Inside
20 Specialist Agents
Agents don't auto-route — they answer directly when asked, and Claude will suggest the right specialist when a question goes deep into their domain.
| Tier | Count | Role | Model |
|---|---|---|---|
| Directors | 3 | Curriculum, Architecture, Research | opus |
| Domain Leads | 5 | Embedding, Retrieval, Indexing, Evaluation, Integration | opus |
| Specialists | 12 | Chunking, Vector DB, Reranking, Prompt Engineering, Hybrid Search, Document Parsing, Metadata, Query Analysis, Deployment, Evaluation Metrics, Graph RAG, Multimodal | sonnet |
22 Slash Commands
| Command | What It Does |
|---|---|
/start | Assess your level, pick a track, get a working pipeline |
/lesson | Start or continue a lesson (checkpoint quizzes between modules) |
/quiz | Test your understanding |
/build | Build a RAG component step by step |
/evaluate | Run metrics on your pipeline |
/debug-rag | Diagnose common RAG failures |
/compare | Compare two approaches with live side-by-side output diffs |
/benchmark | Benchmark pipeline performance |
/architecture | Design a RAG system for a use case |
/paper-review | Walk through a research paper |
/code-review | Get feedback on your RAG code |
/glossary | Look up RAG terminology |
/challenge | Take on a hands-on challenge |
/explain | Deep-dive into any concept (supports ELI5 mode) |
/roadmap | View progress, badges, streaks, and export GitHub badges |
/triage | Not sure where to go? Get routed to the right skill |
/audit-content | Check materials for outdated references |
/recap | Quick summary of what you covered last session |
/sandbox | Spin up a minimal RAG pipeline instantly |
/break-it | Learn by debugging intentionally broken pipelines |
/fix | Skip the teaching, diagnose and fix your pipeline |
/journal | Write notes about what clicked or confused you |
9-Module Curriculum
Each lesson is tagged core or optional. The core path (~8.5 hours) gets you to a working, evaluated RAG system. Optional lessons add depth when you're ready.
| # | Module | Lessons | Core | Key Topics |
|---|---|---|---|---|
| 1 | Foundations | 4 | 4 | What is RAG, architecture, RAG vs fine-tuning |
| 2 | Document Processing | 5 | 2 | Parsing, chunking strategies, metadata |
| 3 | Embeddings | 5 | 2 | Models, vector spaces, similarity, fine-tuning |
| 4 | Vector Databases | 5 | 2 | Chroma, Pinecone, pgvector, indexing |
| 5 | Retrieval Strategies | 5 | 1 | Dense, sparse, hybrid, reranking, MMR |
| 6 | Generation | 5 | 2 | Prompt engineering, grounding, citations |
| 7 | Evaluation | 5 | 2 | RAGAS, retrieval/generation metrics |
| 8 | Advanced Patterns | 5 | 0 | Agentic RAG, Graph RAG, CRAG, multimodal |
| 9 | Production | 5 | 0 | Deployment, caching, monitoring, scaling |
Core path: ~8.5 hours | Full curriculum: ~31 hours
8 Milestones + 4 Proficiency Levels
The curriculum is broken into milestones — concrete checkpoints that mark real capability, not just lessons read.
| # | Milestone | You Can Now... |
|---|---|---|
| 1 | First Light | Build a working RAG system from scratch |
| 2 | Data Wrangler | Turn any document into retrieval-ready chunks |
| 3 | Vector Navigator | Store and search embeddings effectively |
| 4 | Retrieval Engineer | Find the right information for any query |
| 5 | Prompt Architect | Generate grounded, cited answers |
| 6 | Quality Guardian | Measure everything and improve with data |
| 7 | Pattern Master | Go beyond basic RAG when it's warranted |
| 8 | Production Ready | Deploy, monitor, and scale a RAG system |
Complete milestones to earn proficiency levels:
| Level | Milestones | What It Means |
|---|---|---|
| RAG Explorer | 1-2 | Can build a basic pipeline and process documents |
| RAG Practitioner | 3-5 | Can design retrieval systems with proper search and prompting |
| RAG Engineer | 6-7 | Can evaluate, optimize, and apply advanced patterns |
| RAG Architect | 8 + bonus | Can deploy, scale, and maintain production systems |
See milestones.md for full requirements and bonus milestones.
3 Guided Projects
| Level | Project | What You Build |
|---|---|---|
| Starter | Simple Q&A | Basic RAG pipeline over documents |
| Intermediate | Multi-Source Hybrid | Hybrid search + reranking + citations |
| Advanced | Agentic RAG | Self-correcting RAG with routing and tools |
Progress Dashboard
Run make dashboard to generate an HTML progress page showing your milestones, proficiency level, module completion, and quiz scores.
Tech Stack
No API keys required to start. The defaults work out of the box:
| Component | Default (zero config) | Optional Upgrade |
|---|---|---|
| LLM | Claude Code (you're already running it) | Ollama (local, needs 8-16GB RAM) |
| Embeddings | all-MiniLM-L6-v2 (local, no key) | OpenAI text-embedding-3-small |
| Vector DB | ChromaDB (local, no setup) | Pinecone, pgvector, Qdrant |
| Framework | LangChain | LlamaIndex |
| Evaluation | RAGAS | Custom metrics |
| Docs | PyPDF, BeautifulSoup | Unstructured, pdfplumber |
Running models locally? Ollama is free but needs 8-16GB RAM for LLMs. Close heavy apps before running. If your machine slows down, switch to the API path. Local embedding models (all-MiniLM-L6-v2) are lightweight and run fine on most machines.
Project Structure
rag-learning-academy/
├── CLAUDE.md # Master config, voice & tone, agent behavior
├── .claude/
│ ├── settings.json # Hooks, permissions
│ ├── agents/ # 20 specialist agents
│ ├── skills/ # 22 slash commands
│ ├── hooks/ # Freshness checks, validation scripts
│ ├── rules/ # Path-scoped coding standards
│ └── docs/
│ ├── curriculum/ # 9-module learning path (core/optional tagged)
│ ├── reference/ # Glossary, roster, milestones, standards
│ └── templates/ # Architecture, eval, project templates
├── src/ # Your RAG code goes here
├── scripts/ # Dashboard generator, utilities
├── projects/ # Guided build projects
├── tests/ # 616 structural + content tests
├── data/ # Sample documents (6 files incl. 1,247-line chunking doc)
└── progress/ # Your learning progress + dashboard
Learning Tracks
Run /start and you'll be assessed into one of three tracks:
| Track | Score | Modules | Level Earned | Est. Hours |
|---|---|---|---|---|
| Beginner | 0-3 | 1 → 4 | RAG Explorer | 15-25 |
| Intermediate | 4-6 | 3 → 7 | RAG Practitioner | 25-40 |
| Advanced | 7-10 | 6 → 9 | RAG Engineer | 30-45 |
Not sure where to go? Run /triage to get routed based on your situation.
Content Freshness
Academy materials are actively monitored so nothing goes stale:
- Weekly CI — checks PyPI versions, deprecated patterns, MTEB model health, review cycles (details)
- Monthly CI — content age report, creates GitHub issues for stale files
- On-demand — run
/audit-contentfor a deep review with web search verification - Frontmatter — every content file has
last_reviewed,review_cycle, andstaleness_riskmetadata
Philosophy
"Understand → Build → Evaluate → Iterate"
- Agents teach, they don't just code for you
- Every concept comes with a hands-on exercise
- Every exercise has evaluation criteria
- You choose your path — agents advise, you decide
- The tone is conversational and direct — like learning from a smart friend, not reading a textbook
Contributing
Contributions welcome! See CONTRIBUTING.md for guidelines on:
- Reporting bugs and requesting features
- Development setup and running tests (
make ci) - Pull request process and checklist
- Content guidelines and voice & tone
make ci # Run lint + shellcheck + 616 tests before submitting a PR
Credits
Architecture inspired by Claude Code Game Studios by Donchitos.