Syllabus
RAG & Context Engineering: Building Production-Grade AI Systems — an eight-week applied program in retrieval-augmented generation and context engineering.
Program description
Large language models are powerful but bounded: they know only what they were trained on, and they fail silently when asked to go beyond it. Retrieval-augmented generation (RAG) — connecting models to external knowledge at inference time — is how production AI systems ground their answers in private, fresh, and verifiable data. This program teaches RAG as an engineering discipline: not a framework tutorial, but the design decisions, measurement practices, and production concerns that separate demos from systems.
The program is build-first. Each participant selects a document corpus in Week 1 and constructs a complete system over it across eight weeks: classical and semantic retrieval, hybrid fusion and reranking, systematic evaluation, multihop and agentic workflows, and finally secure deployment. All tooling is local-first and open source.
Learning outcomes
On completion, participants will be able to:
- Decide when and why external retrieval is necessary in LLM systems, and defend that decision
- Design classical, semantic, and hybrid retrieval pipelines, and justify each stage with measurement
- Diagnose accuracy and performance failures using structured evaluation and a root-cause taxonomy
- Build end-to-end RAG systems, including query decomposition and multihop retrieval
- Implement retrieval-aware agentic workflows and judge when they beat static pipelines
- Deploy secure, observable, production-grade RAG systems with layered injection defenses
- Communicate system behavior and design trade-offs to technical and non-technical stakeholders
Schedule
Methodology
This program follows a guided-cohort model: one live 90-minute session per week in which the cohort studies the concepts and builds the labs together, with the instructor facilitating pace, discussion, and debugging. The materials are designed to carry the technical content; the sessions exist so that nobody struggles alone and every question gets worked through out loud. Between sessions, participants complete a weekend challenge (approximately 3–4 hours) that stress-tests the week's build and produces a committed deliverable. Expect a total commitment of 6–8 hours per week. Every artifact is version-controlled from day one; the accumulated repository is itself a program outcome — a portfolio piece demonstrating professional practice.
Assessment
Assessment is continuous and evidence-based: eight weekly checkpoints (committed, working code with measured results), eight challenge deliverables, and a capstone defense. The capstone is presented twice — a technical deep-dive and a stakeholder briefing — and is evaluated on whether design choices are supported by the participant's own evaluation data, not on feature count.
Tools and materials
Python 3.11+, uv, git, Ollama (llama3.1:8b and nomic-embed-text), rank-bm25, Chroma, sentence-transformers, RAGAS, FastAPI, and Docker. All tools are free and run locally; a laptop with 8 GB+ RAM is sufficient. Readings draw on primary sources — the original RAG, ReAct, and Lost-in-the-Middle papers, the Stanford IR textbook, and OWASP security guidance — linked from each week's page.
Prerequisites
Working Python proficiency (functions, classes, virtual environments, pip/uv) and basic command-line comfort. No prior experience with LLM APIs, vector databases, or information retrieval is assumed — Weeks 1–3 build these foundations explicitly.
Instructor
Pablo Grossi — program design and session facilitation, AXLE. Sessions are run in the guided-cohort format: the instructor leads the study, works the labs alongside the cohort, and keeps every participant moving — the accompaniment model AXLE programs are built on.