Course: RAG Techniques — From Basic Pipeline to Graph Architectures

Topics covered

• •Understand the architecture of a classic RAG pipeline: loading → chunking → embeddings → vector store → retriever → generation.
• •Build a working RAG over a PDF using FAISS and OpenAI embeddings.
• •Adapt the pipeline to tabular (CSV) and structured (JSON) data.

Topics covered

• •Understand why document-splitting strategy is RAG quality's main "cheap" lever.
• •Learn to tune chunk size experimentally rather than by eye.
• •Master semantic chunking, propositional chunking, and contextual chunk headers.

Topics covered

• •Understand why "raw" user queries match documents poorly.
• •Master three query transformations: rewriting, step-back, and decomposition into sub-queries.
• •Understand HyDE — retrieval through a hypothetical answer document.

Topics covered

• •Combine lexical (BM25) and vector search in fusion retrieval.
• •Apply reranking: cross-encoder and LLM scoring to reorder candidates.
• •Manage context: a window around the chunk, relevant segment extraction, compression, hierarchical indices.

Topics covered

• •Move from a linear pipeline to loops with self-checking and correction.
• •Master Self-RAG: the "retrieve or not" decision and multi-stage self-assessment of the answer.
• •Master Corrective RAG (CRAG): scoring retrieval quality and falling back to web search.

Topics covered

• •Understand which questions vector search can't solve and why knowledge graphs are needed.
• •Build a knowledge graph from text and combine graph traversal with retrieval.
• •Understand Microsoft GraphRAG: entities, communities, bottom-up summaries.

Topics covered

• •Understand which metrics measure retrieval and which measure generation, and why they must be kept separate.
• •Master deepeval: test cases for correctness, faithfulness, contextual relevancy.
• •Get familiar with GroUSE: 6 grounded-generation metrics and meta-evaluation of your own LLM judge.