GyaanStack. Blog

Articles / ragfree-resources

Free resources to learn RAG, in the order I'd use them

The best free courses, docs and papers to learn RAG, in a learning order: basics, chunking, retrieval, reranking, evals and production.

By Raghunath · · 5 min read

Most RAG tutorials teach the same thing: split a PDF, embed the chunks, put them in a vector database, ask a question. That gets you a demo in an hour. It doesn't teach you why the demo gives wrong answers on real documents, and that's the part employers care about.

The resources below cover both. I picked 20 from the GyaanStack RAG and multimodal retrieval subjects (about 245 links together) and ordered them so each one builds on the last. I checked every link before publishing.

Most of this list treats RAG as a search problem, because retrieval decides most of the answer quality. Generation is the easier half.

Stage 1: get the big picture#

Learn RAG From Scratch (freeCodeCamp video course). A full course that builds RAG step by step: indexing, retrieval, generation, then query translation, routing and more advanced indexing. The code is in the rag-from-scratch repo, so you can follow along.

RAG vs Fine-Tuning vs Prompt Engineering (IBM Technology). A short video on when to use which. Watch it early, so you don't reach for RAG when a better prompt or a fine-tune would do.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (paper). The original RAG paper. Read the introduction and method to see what the term first meant; today's pipelines are simpler than the paper's setup.

What AI Engineers Should Know about Search. Written by a search engineer for people coming from the LLM side. It explains why relevance is hard and why "just use embeddings" isn't a search strategy.

Stage 2: chunking and ingestion#

Bad chunks cause many wrong answers that look like model problems.

The 5 Levels Of Text Splitting For Retrieval (video). Walks from fixed-size splitting up to semantic and agent-based splitting, with code. A good way to see the trade-offs side by side.

Chunking Strategies to Improve LLM RAG Pipeline Performance (Weaviate). A written reference for the same ideas, with guidance on chunk size and overlap.

Evaluating Chunking Strategies for Retrieval (Chroma research). Measures chunking methods instead of arguing about them. Read it to learn how to test your own chunking, not to copy its winner.

Getting Started With Embeddings (Hugging Face). A hands-on intro: what an embedding is, how to compute them, how to search with them.

Sentence Transformers documentation. The library most people use to run open embedding and reranking models locally. The docs double as a tutorial on semantic search.

MTEB Leaderboard. Where to compare embedding models. Filter by your task and language, and remember that a leaderboard score is a starting point, not proof the model fits your data.

Hierarchical Navigable Small Worlds (HNSW) (Pinecone). How the most common vector index works. You don't need to implement it, but knowing it's approximate explains some odd misses.

Don't confuse similarity for relevance. A short post on why "close in embedding space" is not the same as "answers the question".

Stage 4: hybrid search and reranking#

This is where a demo becomes a system.

Getting Started with Hybrid Search (Pinecone). Combining keyword search (BM25) with vector search, and why each catches what the other misses. My RAG cheat sheet has a small runnable hybrid example.

Rerankers and Two-Stage Retrieval (Pinecone). Retrieve many candidates cheaply, then rerank the top few with a stronger model. Often one of the biggest quality gains you can add.

Contextual Retrieval (Anthropic). Adding a short, document-aware description to each chunk before embedding and indexing it, combined with BM25 and reranking. The post reports measured drops in retrieval failures, which makes it a useful technique to test on your own data.

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT (paper). The "late interaction" idea behind a family of strong retrievers. Read it once you're comfortable with single-vector search.

Stage 5: evaluate and improve#

Systematically Improving Your RAG. A practical playbook: build synthetic questions, measure retrieval first, look at real queries, then fix the biggest category of failure.

There Are Only 6 RAG Evals. A short framework that lists the relationships you can measure between question, retrieved context and answer. Useful for deciding what to grade.

Ragas documentation. The most common open-source library for RAG metrics such as faithfulness and context recall. Use it once you know which questions you want answered.

Evaluating and optimizing RAG: a talk series with annotated notes. A set of talks by search and RAG practitioners, each with written notes: modern IR evals, late interaction, and more. Good for going deeper after you've built something.

What I'd skip at first#

  • Comparing vector databases. For learning, any one works, and for small collections even brute-force search is fine. Pick one and move on; the choice matters much less than chunking and reranking.
  • GraphRAG and agentic RAG. Both are real techniques, and the GyaanStack RAG subject has a topic on them. Learn them after you can measure plain retrieval, otherwise you can't tell whether the extra complexity helped.
  • Long lists of "advanced RAG techniques". Most are variations of query rewriting, better chunking or reranking. Once you know those three, the lists are easy to read.
  • Fine-tuning embedding models. Worth it later for specialised vocabulary, but only once you have a test set that shows the base model is the bottleneck.

How I'd use this list#

  • Days 1–3: Stage 1. Build the basic pipeline from the video course.
  • Week 2: Stages 2 and 3. Swap chunkers and embedding models and write down what changes.
  • Week 3: Stage 4. Add BM25 and a reranker. Measure recall@k before and after.
  • Week 4: Stage 5. Build a 30-question test set and run it on every change.

If you want a project to apply this to, I wrote a build-along for a PDF question-answering bot that cites its sources.

The full lists live in the GyaanStack RAG (197 links) and multimodal retrieval (48 links) subjects. They're sorted into topics and updated daily.