LearnAISign in
Phase 9 · 11 lessons · 42 h · about 3 weeks

🔎 RAG & LLM Application Engineering

Ground models in your data and ship LLM apps that hold up in production

Lessons

9.1Why RAG, and the canonical pipelineOpen-book answers from a closed-book model2 h 30 min9.2Embeddings deep diveHow text becomes searchable geometry3 h 30 min9.3Vector search and indexesFinding nearest neighbours among millions of vectors3 h 20 min9.4Chunking and ingestionThe unglamorous stage that decides retrieval quality⚡ demo · 3 h 20 min9.5Retrieval quality: hybrid search, reranking, query transforms and metricsGetting the right chunks to the top, and proving it⚡ demo · 4 h9.6Generation side: context assembly, citations, structured outputs and tool callingTurning retrieved chunks into answers people can trust4 h9.7Advanced RAG patternsGrading, hierarchies, graphs, agents and structured data3 h 40 min9.8Evaluating and observing LLM appsEvals are the product4 h9.9Production: latency, cost, caching, guardrails, security and privacyWhat changes when real users and real data arrive3 h 20 min9.10Frameworks and SDKsWhat LangChain, LlamaIndex, DSPy and plain SDKs buy you, and what they cost2 h 30 min9.11Project: a cited RAG assistant over your own documents, with evalsPut the whole phase into one system you can defend8 h

About this phase

Phase 8 taught you what large language models can do; this phase teaches you to make them useful and trustworthy on your data. Retrieval-augmented generation (RAG) connects a model to documents it never saw in training, and LLM application engineering is the discipline of turning a demo into a system: measured retrieval, grounded and cited answers, structured outputs, evaluations that catch regressions, and the production realities of latency, cost, caching, security and privacy. You will build the pipeline piece by piece — embeddings, vector indexes, chunking, hybrid search and reranking, prompt assembly and tool calling — then study the advanced patterns (agentic RAG, GraphRAG), the evaluation craft that separates working systems from demos, and how to choose (or skip) frameworks. The capstone is a cited assistant over your own documents with a real evaluation suite.

By the end you can

  • Explain and derive the core math of retrieval: contrastive embedding training, BM25, reciprocal rank fusion, recall@k, MRR and nDCG
  • Build a complete RAG pipeline in plain Python, then upgrade it with hybrid search, reranking, structured outputs and tool calling
  • Choose chunking strategies, embedding models and vector indexes from measurements on your own data rather than folklore
  • Design evaluation suites with golden sets, RAGAS-style metrics and calibrated LLM judges, and trace a system in production
  • Budget latency and cost, exploit prompt caching, and defend against prompt injection, data leakage and permission bypass
  • Decide when advanced patterns (Self-RAG, GraphRAG, agentic RAG) pay off and when a framework helps or hurts