Phase 9 · 3–4 weeks
Retrieval-Augmented Generation
Build a production-grade RAG that beats naive cosine-similarity on your own eval.
What this phase teaches
Embedding models (BGE, E5, GTE, Cohere, Voyage, Nomic), chunking strategies, vector DBs (Pinecone, Qdrant, Chroma, Milvus, pgvector, LanceDB), hybrid BM25 + dense with RRF, rerankers (Cohere Rerank, BGE-reranker, ColBERTv2), HyDE, multi-hop / agentic RAG, GraphRAG, RAG evaluation (RAGAS).
Anchor resources
- Chip Huyen — AI Engineering (O’Reilly, 2025)
- DeepLearning.AI Advanced Retrieval for AI
- LlamaIndex RAG bootcamp
- Papers: 2005.11401 (Lewis RAG), 2404.16130 (Microsoft GraphRAG), 2309.15217 (RAGAS)