RAG Projects
Master Retrieval-Augmented Generation from basics to production
RAG Projects
Retrieval-Augmented Generation combines the power of search with LLM generation for accurate, grounded responses. RAG is one of the most practical AI patterns in production today.
Start Here — The Crash Course
One page, all of RAG
RAG Crash Course — everything on a single page in about 30 minutes. What RAG is, how every stage works, the decisions that matter, how to measure it, and how to fix it when it breaks. Each section links out to the full page on that topic.
Read this first. Follow the links for whatever you need in depth.
The Full Track — Fundamentals
Want the complete grounding? The Fundamentals track explains everything from zero — every stage, every term, every abbreviation — and carries through to advanced production concerns.
| # | Page | Level | Covers |
|---|---|---|---|
| 0 | RAG Crash Course | Everyone | All of RAG on one page, with links out to everything below |
| 1 | What Is RAG? | Beginner | The problem RAG solves; RAG vs fine-tuning vs long context |
| 2 | The RAG Pipeline | Beginner → Intermediate | Every stage of indexing and querying, and how each fails |
| 3 | Document Loading & Parsing | Beginner → Intermediate | PDFs, tables, scans, and the metadata to capture |
| 4 | Chunking Strategies | Beginner → Advanced | Fixed, recursive, semantic, parent-document, late chunking |
| 5 | Embeddings Explained | Beginner → Advanced | Vectors, cosine similarity, model selection |
| 6 | Vector Databases & Indexes | Intermediate → Advanced | HNSW, IVF, quantization, metadata filtering |
| 7 | Retrieval Strategies | Intermediate → Advanced | Dense, sparse, hybrid, BM25, RRF, query transformation |
| 8 | Reranking & Context Assembly | Intermediate → Advanced | Cross-encoders, ColBERT, lost-in-the-middle |
| 9 | Generation & Grounding | Intermediate | Grounded prompting, citations, refusal design |
| 10 | Conversational RAG | Intermediate | Follow-up questions, query rewriting, chat history |
| 11 | Multimodal RAG | Intermediate → Advanced | Images, charts, tables, scanned pages, vision models |
| 12 | RAG Architectures | Intermediate → Advanced | HyDE, Self-RAG, CRAG, Adaptive, Agentic, Graph, Speculative |
| 13 | Evaluation & Metrics | Intermediate → Advanced | Recall@k, nDCG, MRR, faithfulness, LLM-as-judge |
| 14 | Production & Operations | Advanced | Latency, caching, cost, indexing, security, multi-tenancy |
| 15 | Failure Modes & Debugging | Intermediate → Advanced | Symptom-to-cause diagnosis across every stage |
| 16 | Designing a RAG System | Advanced | A full worked design: requirements, sizing, every choice justified |
| 17 | Glossary | Reference | Every term and abbreviation, expanded |
The fundamentals track is self-contained and assumes no prior knowledge. Work through it in order, or jump to the page matching the problem in front of you.
Learning Path
RAG Learning Path
Basic
Intermediate
Advanced
Projects
Beginner
| Project | Description | Time |
|---|---|---|
| Intelligent Document Q&A | Build a complete RAG system for PDF documents | ~2 hours |
Intermediate
| Project | Description | Time |
|---|---|---|
| Hybrid Search | Combine keyword and semantic search — the production baseline | ~4 hours |
| RAG with Reranking | Retrieve wide, rerank narrow: the biggest quality gain in RAG | ~4 hours |
| Conversational RAG | Add memory and context to your RAG system | ~4 hours |
Advanced
| Project | Description | Time |
|---|---|---|
| Agentic RAG | Retrieval as a tool: the agent searches, grades, rewrites, and retries | ~4 days |
| Multi-Modal RAG | Handle images, tables, and complex PDFs | ~4 days |
| Graph RAG | Knowledge graph enhanced retrieval with Neo4j | ~3 days |
| Production RAG Pipeline | End-to-end production system with evaluation, monitoring, caching | ~3 days |
Why Learn RAG?
| Benefit | Description |
|---|---|
| Accuracy | Grounds LLM responses in your data |
| Control | Limits hallucinations with source attribution |
| Scalability | Works with any document corpus size |
| Privacy | Keep your data in your infrastructure |
Case Studies
Real-world implementations showing RAG in production environments.
| Case Study | Industry | Description | Status |
|---|---|---|---|
| Enterprise Customer Support | SaaS | 100K+ ticket handling with intelligent routing and auto-response | Available |
| Medical Literature Search | Healthcare | Medical research papers, clinical trials, and drug interactions | Available |
Case studies demonstrate complete production systems with architecture, code, deployment, and business metrics.
Key Concepts
RAG Key Concepts
Retrieval
Embeddings, Vector Search, Reranking, Hybrid Search
Augmentation
Chunking, Context Window, Prompt Engineering
Generation
LLM Selection, Temperature, Streaming
Frequently Asked Questions
What is RAG and why should I use it?
RAG (Retrieval-Augmented Generation) is a technique that combines information retrieval with LLM generation. Instead of relying solely on the LLM's training data, RAG retrieves relevant documents from your knowledge base and uses them as context for generating responses. This reduces hallucinations, enables up-to-date responses, and allows LLMs to work with your private data.
What's the difference between RAG and fine-tuning?
Fine-tuning modifies the model's weights to learn new information, requiring significant compute and retraining when data changes. RAG keeps the model unchanged and retrieves relevant context at query time, making it easier to update knowledge, more cost-effective, and better for factual accuracy. Most production systems use RAG for knowledge-grounded responses and fine-tuning for style/behavior changes.
Which vector database should I use for RAG?
For learning and prototypes, use ChromaDB (simple, local, free). For production with <1M vectors, consider Pinecone (managed, easy), Weaviate (hybrid search), or Qdrant (performance). For billion-scale search, use FAISS with IVF/HNSW indexes or Milvus. The choice depends on scale, budget, and whether you need managed vs self-hosted.
How do I improve RAG accuracy?
Key techniques include: (1) Better chunking strategies with semantic boundaries, (2) Hybrid search combining keyword + vector retrieval, (3) Reranking with cross-encoders, (4) Query rewriting and expansion, (5) Iterative retrieval with self-correction. Our intermediate and advanced projects cover all these techniques.
What's the typical RAG architecture?
A standard RAG pipeline has: Document ingestion (load, chunk, embed, store) → Query processing (embed query, retrieve documents) → Generation (construct prompt with context, generate response). Production systems add caching, evaluation, monitoring, and fallback mechanisms.
How much does it cost to run a RAG system?
Costs depend on scale. For a small system: Embedding generation ~$0.0001/1K tokens with OpenAI, vector storage ~$25-70/month for 1M vectors on managed services (free with local ChromaDB), LLM generation ~$0.01-0.03 per query with GPT-4o-mini. Semantic caching can reduce costs by 40-60%.
Start with the Intelligent Document Q&A project to learn the fundamentals.